Product recommendation method and device, computer equipment and storage medium

Abstract

The invention belongs to the field of artificial intelligence, can be applied to the field of finance, and relates to a product recommendation method, which comprises the steps of classifying historical insurance policies according to a preset dimension, and generating a corresponding classification label for each classification category; matching the customer tags of the potential customers with the classification tags to obtain target classification tags; screening out a target customer group according to the target classification label; inputting the obtained target customer information of the target customer group and the target claim settlement case into a verbal skill recommendation model, outputting a target recommended verbal skill, pushing the target recommended verbal skill to potential customers, and receiving voice data fed back by the potential customers; recognizing the voice data to obtain recognized text data; and matching a target product based on customer demand information extracted from the identified text data, and recommending the target product to potential customers. The invention further provides a product recommendation device, computer equipment and a storage medium. In addition, the invention also relates to a block chain technology, and the client tag can be stored in a block chain. According to the method, clients can be accurately positioned, and the product recommendation effect is improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology and can be applied to the financial field, and particularly relates to a product recommendation method, device, computer device, and storage medium. Background Art

[0002] In the financial business scenario, the expansion of business opportunity channels is regarded as one of the key strategies to increase operating income. Taking the insurance business as an example, traditionally, insurance companies mainly build and develop customer systems through their senior salespersons and brokers of each channel, and use this as the main way to obtain business opportunities. However, this method has several inherent limitations, which limit the ability of insurance companies to efficiently and widely explore potential customers and accelerate business growth.

[0003] First of all, the traditional way of obtaining business opportunities highly depends on the personal resources and networks of salespersons and brokers. The target positioning of this method is not accurate enough, and it often relies on the personal experience and relationship networks of salespersons or brokers to gradually accumulate customers, resulting in a slow and inefficient customer exploration process. In addition, due to the limited nature of personal resources, it is difficult to achieve rapid expansion of the customer scale under this model.

[0004] Secondly, it is difficult to effectively spread and replicate successful cases in business practices among different insurance institutions. In the past, the sharing of typical cases mainly occurred through word-of-mouth or in very few cross-institutional and national institutional training sessions. This way of spreading is not only limited in scope but also inefficient, resulting in many successful sales strategies and customer service models not being able to be learned and applied in a timely and extensive manner. This situation not only limits knowledge sharing and innovation within the enterprise but also reduces the competitiveness and development speed of the entire industry.

[0005] Therefore, the existing ways of obtaining business opportunities and exploring customers have significant deficiencies in aspects such as target positioning, efficiency improvement, and knowledge dissemination, which in turn lead to poor product recommendation effects. Summary of the Invention

[0006] The purpose of the present application is to propose a product recommendation method, device, computer device, and storage medium to solve the technical problem that the existing ways of obtaining business opportunities and exploring customers have significant deficiencies in aspects such as target positioning, efficiency improvement, and knowledge dissemination, which in turn lead to poor product recommendation effects.

[0007] To solve the above technical problem, an embodiment of the present application provides a product recommendation method, which adopts the following technical solutions:

[0008] Obtain all historical insurance policies, classify the historical insurance policies according to a preset dimension to obtain a classification result, and generate a corresponding classification label for each classification category in the classification result;

[0009] Obtain the customer tags of potential customers, match the customer tags with the classification tags to obtain target classification tags;

[0010] Filter out the target customer group corresponding to the potential customer according to the target classification tag;

[0011] Obtain the target customer information and target claim cases of the target customer group;

[0012] Input the target customer information and the target claim cases into a trained conversation recommendation model for data processing, and output target recommended conversations;

[0013] Push the target recommended conversation to the potential customer and receive the voice data feedback by the potential customer;

[0014] Perform speech recognition on the voice data to obtain recognized text data;

[0015] Extract the customer demand information from the recognized text data, match the target product based on the customer demand information, and recommend the target product to the potential customer.

[0016] To solve the above technical problems, an embodiment of the present application further provides a product recommendation device, which adopts the following technical solutions:

[0017] A classification module, configured to obtain all historical insurance policies, classify the historical insurance policies according to a preset dimension to obtain a classification result, and generate corresponding classification tags for each classification category in the classification result;

[0018] A tag matching module, configured to obtain the customer tags of potential customers, match the customer tags with the classification tags to obtain target classification tags;

[0019] A screening module, configured to filter out the target customer group corresponding to the potential customer according to the target classification tag;

[0020] An obtaining module, configured to obtain the target customer information and target claim cases of the target customer group;

[0021] A conversation generation module, configured to input the target customer information and the target claim cases into a trained conversation recommendation model for data processing, and output target recommended conversations;

[0022] A pushing module, configured to push the target recommended conversation to the potential customer and receive the voice data feedback by the potential customer;

[0023] A recognition module, configured to perform speech recognition on the voice data to obtain recognized text data;

[0024] A requirement matching module, configured to extract customer requirement information of the recognized text data, match a target product based on the customer requirement information, and recommend the target product to the potential customer.

[0025] To solve the above technical problems, an embodiment of the present application further provides a computer device, which adopts the following technical solution:

[0026] The computer device includes a memory and a processor. Computer-readable instructions are stored in the memory, and when the processor executes the computer-readable instructions, the steps of the product recommendation method described above are implemented.

[0027] To solve the above technical problems, an embodiment of the present application further provides a computer-readable storage medium, which adopts the following technical solution:

[0028] Computer-readable instructions are stored on the computer-readable storage medium, and when the computer-readable instructions are executed by a processor, the steps of the product recommendation method described above are implemented.

[0029] Compared with the prior art, the present application mainly has the following beneficial effects:

[0030] The present application provides a product recommendation method. By classifying all historical insurance policies and generating corresponding classification labels, matching the customer labels of the obtained potential customers with the classification labels, and then screening out the target customer group corresponding to the potential customers, accurate positioning of potential customers can be achieved, so as to provide personalized services for each potential customer, optimize resource allocation, and improve customer satisfaction; by obtaining the target customer information and target claim settlement cases of the target customer group, inputting them into the trained speech recommendation model to generate target recommendation speech, and then pushing the target recommendation speech to the potential customer, personalized recommendation speech can be intelligently provided for the potential customer according to the target customer information and target claim settlement cases of the target customer group, so as to improve customer satisfaction and acceptance, contribute to improving the product recommendation effect and the insurance policy conversion rate, and at the same time improve work efficiency and greatly reduce meaningless communication; receiving the voice data fed back by the potential customer based on the target recommendation speech, further determining the customer requirement information of the potential customer according to the voice data, and matching a target product based on the customer requirement information and recommending it to the potential customer can improve the accuracy of product recommendation, make the recommended product more real-time and reliable, and further improve the product recommendation effect. Description of the Drawings

[0031] To more clearly illustrate the solutions in this application, the following provides a brief introduction to the accompanying drawings required for the description of the embodiments of this application. Obviously, the accompanying drawings in the following description are some embodiments of this application. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0032] Figure 1 is an exemplary system architecture diagram to which this application can be applied;

[0033] Figure 2 is a flowchart of an embodiment of the product recommendation method according to this application;

[0034] Figure 3 is a schematic structural diagram of an embodiment of the product recommendation device according to this application;

[0035] Figure 4 is a schematic structural diagram of an embodiment of the computer device according to this application. Detailed implementation manners

[0036] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs; the terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above accompanying drawings are intended to cover non-exclusive inclusion. The terms "first", "second", etc. in the specification and claims of this application or the above accompanying drawings are used to distinguish different objects and not to describe a specific order.

[0037] Referring to "embodiments" herein means that the specific features, structures, or characteristics described in connection with the embodiments can be included in at least one embodiment of this application. The phrase appears in various places in the specification and does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0038] To enable those skilled in the technical field to better understand the solutions of this application, the following will clearly and completely describe the technical solutions in the embodiments of this application with reference to the accompanying drawings.

[0039] Such as Figure 1As shown in the figure, the system architecture 100 may include a terminal device 101, a network 102, and a server 103. The terminal device 101 may be a laptop computer 1011, a tablet computer 1012, or a mobile phone 1013. The network 102 is a medium for providing a communication link between the terminal device 101 and the server 103. The network 102 may include various connection types, such as wired, wireless communication links, or fiber optic cables, etc.

[0040] Users can use the terminal device 101 to interact with the server 103 through the network 102 to receive or send messages, etc. Various communication client applications may be installed on the terminal device 101, such as a web browser application, a shopping application, a search application, an instant messaging tool, an email client, a social platform software, etc.

[0041] The terminal device 101 may be various electronic devices with a display screen and supporting web browsing. In addition to the laptop computer 1011, the tablet computer 1012, or the mobile phone 1013, the terminal device 101 may also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 (Moving Picture Experts Group Audio Layer IV) player, a laptop portable computer, a desktop computer, etc.

[0042] The server 103 may be a server providing various services, such as a background server supporting the pages displayed on the terminal device 101.

[0043] It should be noted that the product recommendation method provided by the embodiments of the present application is generally executed by the server / terminal device. Correspondingly, the product recommendation device is generally set in the server / terminal device.

[0044] It should be understood that Figure 1 the numbers of the terminal devices, networks, and servers in

[0045] Continue to refer to Figure 2 , which shows a flowchart of an embodiment of the product recommendation method according to the present application, including the following steps:

[0046] Step S201, obtain all historical insurance policies, classify the historical insurance policies according to a preset dimension to obtain a classification result, and generate corresponding classification labels for each classification category in the classification result.

[0047] In this embodiment, all historical insurance policies can be retrieved in batches by directly connecting to the core business system of the insurance company through the API interface, or obtained by connecting to an external data system. The external data systems include insurance agencies, regulatory platforms, third-party procurement platforms, etc. Among them, the policy data of historical insurance policies includes basic information, guarantee information, financial information, special agreement information, service information, and other information. The basic information includes policy identification information (policy number, contract number, application number), time information (application date, effective date, termination date, premium payment period, cooling-off period, etc.), customer entity information (policyholder information, insured information), and sales channel information (agent or broker information, sales channel type); the guarantee information includes insurance liability (insurance type, insurance name, guarantee scope, liability exemption, waiting period), insured amount and compensation rules, and rider information, etc.; the financial information includes premium information (premium amount, payment method, payment cycle), account information (cash value, policy loan amount), and compensation records, etc.; the special agreement information includes endorsement information and special terms, etc.; the service information includes customer service records and value-added service information, etc.; the other information includes contract terms, insurance company name, company address, etc.

[0048] Extract the policy data of all historical insurance policies and preprocess the policy data. The preprocessing includes cleaning invalid and duplicate data, filling or deleting missing values, and standardizing numerical features, etc. Classify all historical insurance policies according to the preset dimensions and the preprocessed policy data to obtain a classification result including multiple classification categories.

[0049] In this embodiment, the preset dimensions include but are not limited to customer industry, customer location, secondary institution, insurance type, sales channel type, etc., and can be specifically set according to the actual situation and are not limited here. Among them, the customer industry and customer location can be obtained from the customer entity information; the insurance type includes but is not limited to life insurance, health insurance, property insurance, accident insurance, liability insurance, etc.; the sales channel type includes but is not limited to insurance agent channels, insurance broker channels, bancassurance channels, direct sales channels of insurance companies, Internet channels, third-party platform channels, corporate group insurance channels, and telemarketing channels, etc.

[0050] In some alternative implementation manners, the step of classifying the historical insurance policies according to the preset dimensions to obtain a classification result includes:

[0051] Determine the optimal number of clusters K through the silhouette coefficient, where K is a positive integer;

[0052] Convert each historical insurance policy into a policy cluster vector according to the preset dimensions, and select K policy cluster vectors from all policy cluster vectors as the cluster centers;

[0053] Calculate the distances between other clustering vectors and each cluster center, and allocate the other clustering vectors to the nearest cluster center according to the distances to form K clusters. The other clustering vectors are policy clustering vectors except for the cluster centers.

[0054] Calculate the centroid of each cluster, use the centroid as the new cluster center, repeat calculating the distances between other clustering vectors and each cluster center, obtain K new clusters, continuously update the cluster centers of each cluster until the clustering stop condition is met, and get the final K classification categories as the classification result.

[0055] Among them, the silhouette coefficient is an index to measure the quality of clustering effect, and its value range is between [-1, 1]. The closer the value is to 1, the better the clustering effect, that is, the points in the same cluster are closer, and the points between different clusters are more dispersed; when the silhouette coefficient is close to -1, it means that the clustering effect is poor, that is, the points may be wrongly assigned to the cluster; when the silhouette coefficient is close to 0, it means that the separation degree and cohesion degree between clusters are similar, and the clustering may not make much sense.

[0056] In this embodiment, before clustering, the historical policies are first vectorized. Specifically, obtain the policy data corresponding to the preset dimension as the policy features, and perform feature encoding on the policy features to obtain the policy clustering vector corresponding to each historical policy. For example, assume that the preset dimension includes categorical policy features such as customer industry, customer location, secondary institution, insurance type, sales channel type, etc., and use one-hot encoding or label encoding to convert the policy data corresponding to the preset dimension into numerical values. If the preset dimension includes numerical policy features such as premium amount and policy term, feature encoding can be performed through standardization (such as Z-Score standardization) or normalization (Min-Max Scaling).

[0057] In some alternative implementation manners of this embodiment, vector conversion algorithms such as Word2Vec algorithm and pre-trained BERT can also be used to convert the policy features into policy clustering vectors, and each historical policy corresponds to a policy clustering vector.

[0058] In some specific examples, algorithms such as Euclidean distance, Manhattan distance, and cosine similarity are used to calculate the distances between other clustering vectors and each cluster center.

[0059] In this embodiment, the clustering stop condition can be that no (or the minimum number of) objects are reassigned to different clusters, no (or the minimum number of) cluster centers change anymore, the sum of squared errors is locally minimized, etc., so that the clustering vectors within the cluster are as closely connected as possible, and the distances between different clusters are as large as possible.

[0060] By classifying historical insurance policies into different classification categories, customer needs can be accurately identified, the market can be targeted, which helps to explore potential customer needs, formulate targeted marketing strategies and service plans for each customer group, provide personalized services, and optimize the customer experience.

[0061] In this embodiment, corresponding classification labels are generated for each classification category according to a preset dimension. These classification labels can be descriptive texts or numerical labels. For example, assuming the preset dimension is: the customer industry is manufacturing, the customer's domicile is in South China, and the secondary institution where the customer is located is Institution A, then the generated classification label is a combined label: Manufacturing - South China - Institution A.

[0062] Step S202: Obtain the customer labels of potential customers, match the customer labels with the classification labels, and obtain the target classification labels.

[0063] Among them, potential customers refer to individuals or organizations that may currently be interested in a certain insurance product or service and have the purchasing power.

[0064] In this embodiment, the customer labels of potential customers include customer basic information, customer industry, customer domicile, the institution where the customer's domicile is located, etc. The customer labels are matched with each classification label one by one, and the classification label that matches the customer label is used as the target classification label.

[0065] In some optional implementation manners of this embodiment, the step of matching the customer labels with the classification labels to obtain the target classification labels includes:

[0066] Compare the customer labels with the classification labels through an exact matching algorithm to generate a first candidate classification label set;

[0067] Calculate the similarity score between the customer labels and the classification labels using a fuzzy matching algorithm to generate a second candidate classification label set;

[0068] Calculate the semantic similarity between the customer labels and the classification labels through a semantic matching algorithm to generate a third candidate classification label set;

[0069] Perform weight fusion on the first candidate classification label set, the second candidate classification label set, and the third candidate classification label set to generate the finally matched target classification labels.

[0070] Among them, the exact matching algorithm can adopt at least one of the KMP (Knuth - Morris - Pratt) algorithm, BF (Brute Force) algorithm, hash algorithm, and regular expression matching algorithm. Through the exact matching algorithm, classification labels that are exactly the same as the customer labels can be matched, and these classification labels are used as the first candidate classification label set.

[0071] The fuzzy matching algorithm is the edit distance algorithm, the N-gram similarity algorithm, or the Jaccard similarity algorithm. The similarity scores between the customer label and each classified label are calculated through the fuzzy matching algorithm, and the classified labels with similarity scores greater than or equal to the first threshold are selected as the second candidate classified label set.

[0072] Among them, the semantic matching algorithm can be based on the word vector algorithm. The customer label and the classified label are converted into vector representations through the word embedding model, the semantic similarity between the customer label vector and each classified label vector is calculated, and the classified labels with semantic similarity greater than or equal to the second threshold are selected as the third candidate classified label set. The word embedding model includes Word2Vec, GloVe, and pre-trained BERT, etc.

[0073] Different weights are assigned to different matching algorithms in advance. Through weight fusion, the comprehensive score of each classified label is calculated, and the classified labels with comprehensive scores greater than or equal to the third threshold are used as the final matching result, that is, the target classified label.

[0074] The customer label and the classified label are matched through different matching algorithms, which improves the matching quality in complex scenarios while ensuring the matching efficiency, and enhances the reliability and accuracy of the matching result.

[0075] It should be emphasized that to further ensure the privacy and security of the customer label information, the above-mentioned customer label can also be stored in a node of a blockchain.

[0076] The blockchain referred to in this application is a new application mode of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, and encryption algorithm. Blockchain, in essence, is a decentralized database, a series of data blocks generated by using cryptographic methods. Each data block contains information about a batch of network transactions, which is used to verify the validity of the information (anti-counterfeiting) and generate the next block. The blockchain can include the blockchain underlying platform, the platform product service layer, and the application service layer, etc.

[0077] Step S203, screening out the target customer group corresponding to the potential customer according to the target classified label.

[0078] In this embodiment, the corresponding target classification categories are obtained according to the target classified label. All the customer entity information corresponding to the historical policies under these target classification categories, and the customer group composed of the customers obtained by removing duplicates from the customers corresponding to the customer entity information is the target customer group corresponding to the potential customer.

[0079] In a specific example, the target customers mainly include policyholders and insureds. The policyholder information and insured information corresponding to all historical policies in the target classification category are extracted to obtain all policyholders and insureds under this target classification category, and duplicate entries among all policyholders and insureds are removed, thus obtaining the target customer group.

[0080] Step S204: Obtain the target customer information and target claim cases of the target customer group.

[0081] In this embodiment, the target customer information includes the customer basic information and customer risk information of each target customer in the target customer group. Among them, the customer basic information includes but is not limited to customer age, gender, occupation (customer industry), address (customer domicile), family situation, etc.; the customer risk information includes customer industry risks, local special risks (such as earthquake zones, etc.).

[0082] The target claim cases include classic claim cases related to the target customer group, classic claim cases of products that potential customers are interested in, etc., which are usually used to demonstrate the insurance company's claim settlement process, compensation standards, as well as the service quality and efficiency of the insurance company.

[0083] Step S205: Input the target customer information and target claim cases into the trained conversation script recommendation model for data processing, and output the target recommended conversation scripts.

[0084] In this embodiment, the target customer information and target claim cases are input into the trained conversation script recommendation model, and the AIGC (Artificial Intelligence Generated Content) ability of the conversation script recommendation model is utilized to generate target recommended conversation scripts suitable for potential customers in combination with the target customer information and target claim cases.

[0085] A training set is constructed using multi-source data in the insurance field. The multi-source data includes historical conversation corpora, policy texts, and recommended conversation script labels annotated by experts. The pre-constructed neural network model is trained using the constructed training set to obtain the conversation script recommendation model.

[0086] In some alternative implementation manners, the conversation script recommendation model includes a knowledge layer, an embedding layer, multiple decoder layers, and an output layer. The step of inputting the target customer information and target claim cases into the trained conversation script recommendation model for data processing and outputting the target recommended conversation scripts includes:

[0087] Input the target customer information and target claim cases into the knowledge layer, obtain the corresponding policy knowledge graph according to the target claim cases, and inject the knowledge in the policy knowledge graph into the target claim cases to form a sentence tree. Input the sentence tree and target customer information into the embedding layer;

[0088] The sentence tree and target customer information are vector-encoded through the embedding layer to obtain the encoded embedding vector;

[0089] Feature extraction is performed on the encoded embedding vector through multiple decoder layers to obtain enhanced semantic sequence features;

[0090] The enhanced semantic sequence features are autoregressively generated through the output layer to obtain the target recommendation words.

[0091] Among them, the knowledge graph is a knowledge base in which the data is integrated through a graph-structured data model or topology. It usually represents a semantic relationship graph between entities, and is stored in the form of a triple of <entity, relationship, attribute>. An entity represents a specific thing that exists in the real world, a relationship expresses a certain semantic association between entities, and an attribute is a specific information of an entity, which is used to describe the essential characteristics or category of an entity. The nodes (entities) of the insurance policy knowledge graph include insurance products, coverage, insurance conditions, premium rules, etc.; the edges (relationships) include inclusion, requirements, associations, exclusions, etc.; the attributes include product attributes, customer attributes, and dynamic attributes, among which product attributes include coverage period, waiting period, deductible, etc., customer attributes include age, occupation, historical claims records, etc., and dynamic attributes include product popularity, customer click rate, recommendation success rate, etc. For example, the triples of the insurance policy knowledge graph include <critical illness insurance A, including, malignant tumor protection>, <critical illness insurance A, requirement, age 18-60 years old>, <critical illness insurance A, associated, annual payment 5,000 yuan>, <auto insurance, including, third party liability insurance>, <auto insurance, requirement, vehicle use life ≤ 10 years>, etc.

[0092] The knowledge layer is responsible for injecting the knowledge in the policy knowledge graph into the target claim case to form a sentence tree. In this embodiment, forming a sentence tree is divided into two steps: knowledge query (K-Query) and knowledge injection (K-Inject).

[0093] The steps of injecting the knowledge in the policy knowledge graph into the target claim case to form a sentence tree include:

[0094] Extract the case text data of the target claim case, call the knowledge query function of the knowledge layer, identify all entities in the case text data, and query the triples corresponding to each entity in the policy knowledge graph;

[0095] The similarity between the triples and the case text data is calculated through the attention mechanism to obtain the relevant triples;

[0096] The knowledge injection function of the knowledge layer is called to embed the relevant triples into the corresponding positions of the case text data to obtain the sentence tree.

[0097] K-Query is responsible for querying the relationships and attributes corresponding to each entity in each text sentence of the case text data from the policy knowledge graph, that is, triples. The specific query process is as follows:

[0098] E = K_Query(s, K);

[0099] Among them, the function K_Query represents querying the policy knowledge graph K with the text sentence s to obtain the triple set E = {0w o , r i0 , w i0 ), ……, 9w o , r ik , w ik )}.

[0100] Then, the attention mechanism is used to evaluate the relevance of each triple to the input case text data, and select those triples that are highly relevant to the context of the case text data. By calculating the similarity between entities and relationships in the context embedding space, triples with high relevance are selected. Among them, the calculation of similarity can be implemented using cosine similarity, Euclidean distance, or other similarity measurement methods.

[0101] K-Inject is responsible for embedding the triple set E into the corresponding positions in the text sentence s to form a sentence tree. Each triple constitutes a branch, and the matrix sentence tree output by the knowledge layer is:

[0102] t = {w1, w2, w3, ……, w i (r i0 , w i0 ), ……, w i (r ik , w ik ), ……, w n}}.

[0103] In this embodiment, by calculating the similarity between triples and case text data through the attention mechanism, unimportant triples can be filtered out, reducing knowledge noise and improving the accuracy and relevance in the process of injecting the policy knowledge graph, thereby improving the quality and accuracy of dialogue generation.

[0104] In this embodiment, the embedding layer includes token embedding (word embedding) and Rotary Position Embedding (RoPE, rotational position embedding). Among them, word embedding maps each token in the sentence into a word vector representation with a dimension of H; Rotary Position Embedding encodes position information into the word vector through a rotation matrix. Specifically, through the method of complex rotation, tokens are rotated in a high-dimensional space to encode relative position information.

[0105] Specifically, the sentence tree and target customer information are input into the embedding layer for word embedding and rotary position embedding operations to obtain corresponding word encoding vectors and position encoding vectors; the word encoding vectors and position encoding vectors are added together to obtain encoded embedding vectors.

[0106] In some alternative implementation manners of this embodiment, the multi-layer decoder layer is formed by stacking and connecting multiple encoder layers. Each decoder layer includes a masked multi-head self-attention layer and a feed-forward neural network layer. The input encoded embedding vectors are block-processed through the masked multi-head self-attention layer to obtain multiple block vectors. Grouped Query Attention (GQA) calculations are performed within each block vector to obtain each block attention feature, and then inter-block GQA calculations are performed on each block attention feature to obtain global attention features; two fully connected networks of the feed-forward neural network layer perform two-layer fully connected transformations on the global attention features and are processed through the non-linear activation function SwiGLU to obtain enhanced semantic sequence features.

[0107] In some specific examples, residual connections and layer normalization processing are respectively performed on the outputs of the masked multi-head self-attention layer and the feed-forward neural network layer to improve the model performance.

[0108] In this embodiment, the output layer maps the conversation strategy prediction sequence to the vocabulary size to generate the final word probability distribution. According to the word probability distribution, the most appropriate word is selected through greedy search, sampling, or Beam Search to construct the recommended conversation strategy until the end symbol is generated or the maximum length is reached, and the final recommended conversation strategy is output as the target recommended conversation strategy.

[0109] Through the conversation strategy recommendation model, the target recommended conversation strategy is generated based on the target customer information and the target claim case, realizing the precise analysis of potential customers and the intelligent generation of personalized service conversation strategies, improving the efficiency and accuracy of conversation strategy generation, greatly reducing meaningless communication, enhancing customer satisfaction and acceptance, and contributing to improving the recommendation effect.

[0110] Step S206: Push the target recommended conversation strategy to the potential customer and receive the voice data fed back by the potential customer.

[0111] In this embodiment, after pushing the target recommended conversation strategy to the potential customer and reaching the customer, the feedback data of the potential customer is continuously monitored. The feedback data includes voice data, conversation reply text data, etc. fed back by the customer.

[0112] In some alternative implementation manners, the step of pushing the target recommended conversation strategy to the potential customer includes:

[0113] Determine the pushing manner of the target recommended conversation strategy;

[0114] When the push method is text-based, directly push the target recommended speech to potential customers;

[0115] When the push method is audio-based, synthesize the speech of the target recommended speech and push it to potential customers.

[0116] When pushing the target recommended speech to potential customers, first confirm the push method. The push methods include text-based, audio-based, etc. When it is determined that the push method is text-based, the target recommended speech can be sent to potential customers through methods such as email, text message, message push, etc.; when it is determined that the push method is audio-based, convert the target recommended speech into speech of the speech, and push the speech of the speech to potential customers.

[0117] In some alternative implementation manners, converting the target recommended speech into speech of the speech can be implemented by using a trained speech synthesis model. Input the target recommended speech into the trained speech synthesis model. The speech synthesis model includes an emotion analysis module, an acoustic structure, and a vocoder; perform emotion classification and emotion embedding on the target recommended speech through the emotion analysis module to obtain a speech emotion embedding vector; extract acoustic features from the speech emotion embedding vector through the acoustic structure to obtain speech acoustic features; input the speech acoustic features into the vocoder for speech waveform synthesis to obtain the speech of the speech.

[0118] Among them, the emotion analysis module includes an emotion classification model and an emotion embedding layer. Perform emotion classification on the target recommended speech through the emotion classification model to obtain an emotion category, and then perform emotion embedding on the emotion category through the emotion embedding layer to generate a speech emotion embedding vector. The emotion classification model can adopt common classification models, such as decision trees, random forests, Transformer architecture models, etc., or a custom neural network model. The acoustic structure includes a phoneme conversion layer, an acoustic encoder, and an acoustic decoder. Map the speech emotion embedding vector to phonemes through the phoneme conversion layer to obtain a phoneme embedding sequence. Extract features from the phoneme embedding sequence through the acoustic encoder to obtain a phoneme feature sequence; decode the phoneme feature sequence through the acoustic decoder to obtain speech acoustic features. The vocoder adopts a HifiGAN vocoder. Through the GAN generator network in the vocoder, convert the input speech acoustic features into high-fidelity speech waveforms, that is, obtain the synthesized speech of the speech.

[0119] Construct a sample data set to train the speech synthesis model. The sample data set contains a large number of paired data. Each paired data includes a text data sample and its corresponding acoustic feature sample, and the acoustic feature sample is labeled with a corresponding emotion label. Use the constructed sample data set to train the pre-constructed neural network model to obtain the finally trained speech synthesis model.

[0120] By selecting different push methods to push the target recommended sales talk to the corresponding potential customers, it is possible to better meet the needs of different customers, provide personalized services to customers, improve communication efficiency. At the same time, it can better display the features and advantages of the product, increase the conversion rate, and improve the service quality.

[0121] Step S207: Perform speech recognition on the speech data to obtain recognized text data.

[0122] Input the speech data into the trained speech recognition model. The speech recognition model includes an acoustic feature extraction layer and a speech recognition layer. Extract features from the speech data through the acoustic feature extraction layer to obtain an acoustic feature vector. Input the acoustic feature vector into the speech recognition layer for inference and prediction to obtain recognized text data.

[0123] Among them, the acoustic feature extraction layer can adopt an LSTM network. Specifically, a bidirectional Bi-LSTM can be used. For the input enhanced audio data, the recurrent neural network is respectively used in sequence and reverse order to obtain two independent hidden layer representations, and then certain calculations (concatenation or addition) are performed on these two hidden layer representations to obtain a final hidden layer representation, that is, the acoustic feature vector, and the acoustic feature vector is output for subsequent calculations. This hidden layer representation of the acoustic feature vector contains both speech information from the previous moment and the next moment.

[0124] The speech recognition layer is implemented by using a Transformer decoding network. Input the acoustic feature vector into the Transformer decoding network for calculation to obtain the text output probability distribution corresponding to each frame. Decode the text output probability distribution. Commonly used decoding algorithms include greedy decoding, beam search decoding, etc. to obtain the corresponding text sequence. Finally, the text sequences are concatenated in chronological order to obtain the complete recognized text data.

[0125] In this embodiment, the speech recognition model is trained by using a large amount of historical corpus data. The training process of the model continuously adjusts the weight parameters of the network through the backpropagation algorithm to minimize the difference between the prediction result of the model and the true label, that is, the loss value no longer changes. Among them, the CTC loss function is used to calculate the loss value between the prediction result and the true label during the training process of the speech recognition model.

[0126] Performing speech recognition on the speech data through the speech recognition model can improve the accuracy, speed, and stability of speech recognition.

[0127] Step S208: Extract the customer demand information from the recognized text data, match the target product based on the customer demand information, and recommend the target product to the potential customer.

[0128] In this embodiment, the customer demand information of the recognized text data is extracted to further determine the customer's appeal, and relevant products are matched according to the customer demand information.

[0129] Specifically, the customer demand information is extracted from the recognized text data through keyword matching or regular expressions. The customer demand information includes premium requirements, insured amount requirements, protection content, etc., and the customer demand information is matched and associated with the products in the insurance product library.

[0130] In some alternative implementation manners of this embodiment, the step of matching the target product based on the customer demand information includes:

[0131] Feature extraction and quantization processing are performed on the customer demand information to obtain a demand feature vector;

[0132] Feature extraction and quantization processing are performed on each product in the insurance product library to obtain a product feature vector;

[0133] Calculate the similarity between the demand feature vector and the product feature vector;

[0134] According to the similarity, at least one target product that matches the customer demand information is screened out from the insurance product library.

[0135] Among them, common quantization processing methods are as follows:

[0136] 1) One-hot encoding: For categorical features (such as product type, payment method, etc.), convert them into multiple binary features;

[0137] 2) Normalization: For numerical features (such as premium, insured amount, etc.), scale them to a specific range, such as [0, 1], for easy comparison between different features;

[0138] 3) Binarization: For some features, such as whether to include a certain additional service, it can be converted into a binary feature of 0 or 1, such as 1 for inclusion and 0 for non-inclusion;

[0139] 4) Text vectorization: For text features (such as the description of protection content), methods such as TF-IDF or Word2Vec can be used to convert them into numerical vectors.

[0140] In this embodiment, the similarity algorithms include but are not limited to the cosine similarity algorithm, Euclidean distance algorithm, and Pearson correlation coefficient algorithm.

[0141] Calculate the similarity, sort it from high to low according to the similarity, and select the preset number of products ranked at the front as the target products.

[0142] By matching target products according to customer demand information, it is possible to accurately match the needs of customers, meet the personalized needs of customers, improve customer satisfaction and ensure efficiency. At the same time, it is conducive to optimizing resource allocation.

[0143] In this embodiment, after the target product is matched, the reason for recommending the target product is generated synchronously, and the target product and the reason for recommendation are pushed to potential customers synchronously.

[0144] In this application, all historical insurance policies are classified and corresponding classification labels are generated. The customer labels of the obtained potential customers are matched with the classification labels, and then the target customer group corresponding to the potential customers is screened out, which can realize the accurate positioning of potential customers, so as to provide personalized services for each potential customer, optimize resource allocation, and improve customer satisfaction; by obtaining the target customer information and target claim settlement cases of the target customer group and inputting them into the trained speech recommendation model to generate target recommendation speech, and then pushing the target recommendation speech to potential customers, it is possible to intelligently provide personalized recommendation speech for potential customers according to the target customer information and target claim settlement cases of the target customer group, so as to improve customer satisfaction and acceptance, help improve product recommendation effect and improve the insurance policy conversion rate, and at the same time improve work efficiency and greatly reduce meaningless communication; receiving the voice data feedback by the potential customer based on the target recommendation speech, further determining the customer demand information of the potential customer according to the voice data, and matching the target product based on the customer demand information and recommending it to the potential customer can improve the accuracy of product recommendation, make the recommended product more real-time and reliable, and further improve the product recommendation effect.

[0145] In some optional implementation manners, after the step of recommending the target product to the potential customer, the following steps are further included:

[0146] Obtain the actual insurance purchase selection result of the potential customer, and adjust the recommendation factor coefficient of the target product according to the actual insurance purchase selection result.

[0147] In this embodiment, the feedback result of the potential customer on the target product is continuously monitored. The feedback result includes the emotional tendency, satisfaction degree of the target product, and whether to choose to purchase insurance, etc. When the actual insurance purchase selection result is non-coverage, the recommendation factor coefficient of the target product for this potential customer is reduced; when the actual insurance purchase selection result is coverage, the recommendation factor coefficient of the target product for this potential customer is enhanced.

[0148] By adjusting the recommendation factor coefficient of the target product according to the actual insurance purchase selection result, the recommendation algorithm can be continuously optimized, ensuring the continuous improvement of service quality and the continuous improvement of customer satisfaction.

[0149] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is a theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results.

[0150] Artificial intelligence basic technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, and mechatronics. Artificial intelligence software technologies mainly include several major directions such as computer vision technology, robotics, biometric technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0151] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through computer-readable instructions. The computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, read-only memory (ROM), or a random access memory (RAM), etc.

[0152] It should be understood that although the steps in the flowchart of the accompanying drawings are shown in sequence according to the arrows, these steps do not necessarily have to be executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps does not have a strict order limit, and they can be executed in other orders. Moreover, at least some of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages. These sub-steps or stages do not necessarily have to be executed at the same time, but can be executed at different times, and their execution order does not necessarily have to be sequential, but can be executed alternately or alternately with at least some of the sub-steps or stages of other steps.

[0153] Further reference Figure 3 to Figure 2 As an implementation of the method shown above, an embodiment of a product recommendation device is provided in the present application. This device embodiment corresponds to the method embodiment shown in Figure 2 and can be specifically applied to various electronic devices.

[0154] Such as Figure 3As shown in the figure, the product recommendation device 300 described in this embodiment includes: a classification module 301, a label matching module 302, a screening module 303, an acquisition module 304, a script generation module 305, a push module 306, an identification module 307, and a demand matching module 308.

[0155] Among them:

[0156] The classification module 301 is used to obtain all historical insurance policies, classify the historical insurance policies according to a preset dimension to obtain a classification result, and generate corresponding classification labels for each classification category in the classification result;

[0157] The label matching module 302 is used to obtain the customer labels of potential customers, match the customer labels with the classification labels to obtain target classification labels;

[0158] The screening module 303 is used to screen out the target customer group corresponding to the potential customer according to the target classification label;

[0159] The acquisition module 304 is used to obtain the target customer information and target claim cases of the target customer group;

[0160] The script generation module 305 is used to input the target customer information and the target claim cases into a trained script recommendation model for data processing, and output target recommended scripts;

[0161] The push module 306 is used to push the target recommended script to the potential customer and receive the voice data fed back by the potential customer;

[0162] The identification module 307 is used to perform speech recognition on the voice data to obtain recognition text data;

[0163] The demand matching module 308 is used to extract the customer demand information from the recognition text data, match the target product based on the customer demand information, and recommend the target product to the potential customer.

[0164] It should be emphasized that to further ensure the privacy and security of customer label information, the above customer labels can also be stored in a node of a blockchain.

[0165] Based on the above product recommendation device, by screening out the target customer groups corresponding to potential customers, the accurate positioning of potential customers can be achieved, so as to provide personalized services for each potential customer, optimize resource allocation, and improve customer satisfaction; by generating target recommendation scripts according to the target customer information and target claim settlement cases of the target customer groups and pushing the target recommendation scripts to potential customers, personalized recommendation scripts can be intelligently provided for potential customers to improve customer satisfaction and acceptance, which helps to improve the product recommendation effect and the policy conclusion rate, and at the same time improves work efficiency and greatly reduces meaningless communication; receiving the voice data feedback by potential customers based on the target recommendation scripts, further determining the customer demand information of potential customers according to the voice data, and matching target products to recommend to potential customers based on the customer demand information can improve the accuracy of product recommendation, make the recommended products more real-time and reliable, and further improve the product recommendation effect.

[0166] In some optional implementation manners of this embodiment, the classification module 301 is further configured to:

[0167] Determine the optimal number of clusters K through the silhouette coefficient, where K is a positive integer;

[0168] Convert each of the historical insurance policies into a policy clustering vector according to a preset dimension, and select K of the policy clustering vectors from all the policy clustering vectors as the clustering centers;

[0169] Calculate the distances between other clustering vectors and each of the clustering centers, and allocate the other clustering vectors to the nearest clustering center according to the distances to form K clustering clusters, where the other clustering vectors are the policy clustering vectors except the clustering centers;

[0170] Calculate the centroid of each of the clustering clusters, use the centroid as the new clustering center, repeat calculating the distances between the other clustering vectors and each of the clustering centers, obtain K new clustering clusters, and continuously update the clustering centers of each of the clustering clusters until the clustering stop condition is met, and obtain the final K classification categories as the classification result.

[0171] By dividing historical insurance policies into different classification categories, the customer needs can be accurately identified, the market can be positioned, which helps to explore the potential needs of customers, formulate targeted marketing strategies and service plans for each customer group, provide personalized services, and optimize the customer experience.

[0172] In some optional implementation manners, the label matching module 302 includes:

[0173] An exact matching sub-module, configured to compare the customer label with the classification label through an exact matching algorithm to generate a first candidate classification label set;

[0174] A fuzzy matching sub-module, which is used to calculate the similarity score between the customer label and the classification label by using a fuzzy matching algorithm, and generate a second candidate classification label set;

[0175] A semantic matching sub-module, which is used to calculate the semantic similarity between the customer label and the classification label through a semantic matching algorithm, and generate a third candidate classification label set;

[0176] A fusion sub-module, which is used to perform weight fusion on the first candidate classification label set, the second candidate classification label set, and the third candidate classification label set to generate a target classification label for the final match.

[0177] The customer label and the classification label are matched through different matching algorithms, which can improve the matching quality in complex scenarios while ensuring the matching efficiency, and enhance the reliability and accuracy of the matching result.

[0178] In some optional implementation manners, the conversation recommendation model includes a knowledge layer, an embedding layer, multiple decoder layers, and an output layer. The conversation generation module 305 includes:

[0179] A knowledge injection sub-module, which is used to input the target customer information and the target claim case into the knowledge layer, obtain the corresponding policy knowledge graph according to the target claim case, inject the knowledge in the policy knowledge graph into the target claim case to form a sentence tree, and input the sentence tree and the target customer information into the embedding layer;

[0180] An embedding sub-module, which is used to perform vector encoding on the sentence tree and the target customer information through the embedding layer to obtain an encoded embedding vector;

[0181] An encoding sub-module, which is used to extract features from the encoded embedding vector through the multiple decoder layers to obtain enhanced semantic sequence features;

[0182] An output sub-module, which is used to perform autoregressive generation on the enhanced semantic sequence features through the output layer to obtain a target recommended conversation.

[0183] Through the conversation recommendation model, the target recommended conversation is generated according to the target customer information and the target claim case, realizing the precise analysis of potential customers and the intelligent generation of personalized service conversations, improving the efficiency and accuracy of conversation generation, greatly reducing meaningless communication, enhancing customer satisfaction and acceptance, and helping to improve the recommendation effect.

[0184] In some optional implementation manners, the push module 306 is further used for:

[0185] Determine the push method of the target recommended conversation;

[0186] When the push method is text reach, directly push the target recommended words to the potential customer;

[0187] When the push method is audio reach, synthesize the target recommended words into speech and push it to the potential customer.

[0188] By selecting different push methods to push the target recommended words to the corresponding potential customers, it can better meet the needs of different customers, provide personalized services for customers, improve communication efficiency. At the same time, it can better display the features and advantages of the product, improve the conversion rate, and improve the service quality.

[0189] In some alternative implementation manners of this embodiment, the requirement matching module 308 is further configured to:

[0190] Extract features and perform quantization processing on the customer requirement information to obtain a requirement feature vector;

[0191] Extract features and perform quantization processing on each product in the insurance product library to obtain a product feature vector;

[0192] Calculate the similarity between the requirement feature vector and the product feature vector;

[0193] According to the similarity, screen out at least one target product that matches the customer requirement information from the insurance product library.

[0194] By matching target products according to customer requirement information, it can accurately match the needs of customers, meet the personalized needs of customers, improve customer satisfaction and guarantee efficiency. At the same time, it is beneficial to optimize resource allocation.

[0195] In some alternative implementation manners, the product recommendation device further includes an optimization module, which is configured to obtain the actual insurance purchase selection result of the potential customer, and adjust the recommendation factor coefficient of the target product according to the actual insurance purchase selection result.

[0196] By adjusting the recommendation factor coefficient of the target product according to the actual insurance purchase selection result, it can continuously optimize the recommendation algorithm, ensuring continuous improvement of service quality and continuous improvement of customer satisfaction.

[0197] To solve the above technical problems, an embodiment of the present application also provides a computer device. Specifically, please refer to Figure 4 , Figure 4 which is the basic structural block diagram of the computer device in this embodiment.

[0198] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are communicatively connected to each other via a system bus. It should be noted that only the computer device 4 having a memory 41, a processor 42, and a network interface 43 is shown in the figure. However, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Among them, those skilled in the art of the present technology can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0199] The computer device can be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The computer device can perform human-computer interaction with the user through means such as a keyboard, a mouse, a remote control, a touchpad, or a voice control device.

[0200] The memory 41 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (such as SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disc, etc. In some embodiments, the memory 41 can be an internal storage unit of the computer device 4, such as the hard disk or memory of the computer device 4. In other embodiments, the memory 41 can also be an external storage device of the computer device 4, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the computer device 4. Of course, the memory 41 can also include both the internal storage unit and the external storage device of the computer device 4. In this embodiment, the memory 41 is generally used to store the operating system and application software of the computer device 4, such as computer-readable instructions of the product recommendation method. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or will be output.

[0201] In some embodiments, the processor 42 may be a Central Processing Unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips. The processor 42 is generally used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to run the computer-readable instructions stored in the memory 41 or process data, such as running the computer-readable instructions of the product recommendation method.

[0202] The network interface 43 may include a wireless network interface or a wired network interface, and the network interface 43 is generally used to establish a communication connection between the computer device 4 and other electronic devices.

[0203] By screening out the target customer group corresponding to the potential customers, the accurate positioning of the potential customers can be achieved, so as to provide personalized services for each potential customer, optimize the resource allocation, and improve the customer satisfaction; by generating the target recommendation words according to the target customer information and target claim cases of the target customer group and pushing the target recommendation words to the potential customers, the personalized recommendation words can be intelligently provided for the potential customers to improve the customer satisfaction and acceptance, which helps to improve the product recommendation effect and the policy conclusion rate, and at the same time improves the work efficiency and greatly reduces the meaningless communication; receiving the voice data feedback by the potential customers based on the target recommendation words, further determining the customer demand information of the potential customers according to the voice data, and matching the target products to the potential customers based on the customer demand information can improve the accuracy of the product recommendation, make the recommended products more real-time and reliable, and further improve the product recommendation effect.

[0204] The present application also provides another implementation manner, that is, to provide a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor, so that the at least one processor executes the steps of the product recommendation method as described above.

[0205] By screening out the target customer groups corresponding to potential customers, the accurate positioning of potential customers can be achieved, so as to provide personalized services for each potential customer, optimize resource allocation, and improve customer satisfaction; by generating target recommendation scripts based on the target customer information and target claim settlement cases of the target customer groups and pushing the target recommendation scripts to potential customers, personalized recommendation scripts can be intelligently provided for potential customers to improve customer satisfaction and acceptance, which helps to improve the product recommendation effect and the policy conclusion rate, while improving work efficiency and greatly reducing meaningless communication; receiving the voice data feedback by potential customers based on the target recommendation scripts, further determining the customer demand information of potential customers according to the voice data, and matching target products to recommend to potential customers based on the customer demand information can improve the accuracy of product recommendation, make the recommended products more real-time and reliable, and further improve the product recommendation effect.

[0206] From the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions for causing a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in various embodiments of the present application.

[0207] Obviously, the above-described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The preferred embodiments of the present application are given in the drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosed content of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions described in the foregoing specific embodiments, or perform equivalent replacements on some of the technical features. Any equivalent structure directly or indirectly using the content of the specification and drawings of the present application in other related technical fields is equally within the scope of the patent protection of the present application.

Claims

1. A product recommendation method, characterized in that: The steps include: Obtain all historical insurance policies, classify the historical insurance policies according to preset dimensions, obtain classification results, and generate corresponding classification labels for each classification category in the classification results; Acquire a customer tag of a potential customer, and match the customer tag with the classification tag to obtain a target classification tag; Filter out the target customer group corresponding to the potential customer according to the target classification label; Obtain target customer information and target claims cases of the target customer group; Input the target customer information and the target claim case into the trained speech recommendation model for data processing, and output the target recommended speech; Pushing the target recommendation words to the potential customer and receiving voice data fed back by the potential customer; Performing speech recognition on the speech data to obtain recognized text data; The customer demand information of the recognized text data is extracted, a target product is matched based on the customer demand information, and the target product is recommended to the potential customer.

2. The product recommendation method according to claim 1, characterized in that: The step of classifying the historical insurance policies according to the preset dimensions to obtain the classification results comprises: The optimal number of clusters K is determined by the silhouette coefficient, where K is a positive integer; Convert each of the historical insurance policies into an insurance policy cluster vector according to a preset dimension, and select K insurance policy cluster vectors from all the insurance policy cluster vectors as cluster centers; Calculate the distance between other clustering vectors and each of the clustering centers, and assign other clustering vectors to the clustering center with the closest distance according to the distance to form K clusters, wherein the other clustering vectors are the policy clustering vectors other than the clustering center; The centroid of each cluster is calculated, and the centroid is used as the new cluster center. The distance between the other cluster vectors and each cluster center is repeatedly calculated to obtain K new clusters. The cluster centers of each cluster are continuously updated until the clustering stop condition is met, and the final K classification categories are obtained as the classification results.

3. The product recommendation method according to claim 1, characterized in that: The step of matching the customer tag with the classification tag to obtain a target classification tag comprises: By using an exact matching algorithm, the customer tag is compared with the classification tag to generate a first candidate classification tag set; Calculating the similarity score between the customer tag and the classification tag using a fuzzy matching algorithm to generate a second candidate classification tag set; Calculating the semantic similarity between the customer tag and the classification tag through a semantic matching algorithm to generate a third candidate classification tag set; The first candidate classification label set, the second candidate classification label set and the third candidate classification label set are weightedly fused to generate a final matching target classification label.

4. The product recommendation method according to claim 1, characterized in that: The speech recommendation model includes a knowledge layer, an embedding layer, a multi-layer decoder layer and an output layer. The step of inputting the target customer information and the target claim case into the trained speech recommendation model for data processing and outputting the target recommended speech includes: Input the target customer information and the target claim case into the knowledge layer, obtain the corresponding policy knowledge graph according to the target claim case, and inject the knowledge in the policy knowledge graph into the target claim case to form a sentence tree, and input the sentence tree and the target customer information into the embedding layer; Performing vector encoding on the sentence tree and the target customer information through the embedding layer to obtain an encoded embedding vector; Performing feature extraction on the encoded embedding vector through the multi-layer decoder layer to obtain enhanced semantic sequence features; The enhanced semantic sequence features are autoregressively generated through the output layer to obtain target recommendation words.

5. The product recommendation method according to claim 1, characterized in that: The step of pushing the target recommendation words to the potential customer comprises: Determine the push method of the target recommendation words; When the push mode is text contact, the target recommendation words are directly pushed to the potential customer; When the push mode is audio contact, the synthesized speech voice of the target recommended speech is pushed to the potential customer.

6. The product recommendation method according to claim 1, characterized in that: The step of matching target products based on the customer demand information includes: Extracting and quantifying the customer demand information to obtain a demand feature vector; Perform feature extraction and quantification on each product in the insurance product database to obtain a product feature vector; Calculating the similarity between the demand feature vector and the product feature vector; At least one target product matching the customer demand information is screened out from the insurance product library according to the similarity.

7. The product recommendation method according to any one of claims 1 to 6, characterized in that: After the step of recommending the target product to the potential customer, the method further includes: The actual insurance selection result of the potential customer is obtained, and the recommendation factor coefficient of the target product is adjusted according to the actual insurance selection result.

8. A product recommendation device, characterized in that: include: A classification module is used to obtain all historical insurance policies, classify the historical insurance policies according to preset dimensions, obtain classification results, and generate corresponding classification labels for each classification category in the classification results; A tag matching module is used to obtain a customer tag of a potential customer, match the customer tag with the classification tag, and obtain a target classification tag; A screening module, used to screen out target customer groups corresponding to the potential customers according to the target classification labels; An acquisition module, used to acquire target customer information and target claims cases of the target customer group; A speech generation module, used for inputting the target customer information and the target claim case into the trained speech recommendation model for data processing, and outputting the target recommended speech; A push module, used to push the target recommendation words to the potential customer and receive voice data fed back by the potential customer; A recognition module, used for performing speech recognition on the speech data to obtain recognition text data; The demand matching module is used to extract the customer demand information from the identified text data, match the target product based on the customer demand information, and recommend the target product to the potential customer.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of the product recommendation method according to any one of claims 1 to 7 when executing the computer-readable instructions.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the steps of the product recommendation method according to any one of claims 1 to 7 are implemented.

Images