Safety product pushing method and device based on large language model

Abstract

The invention discloses a security product pushing method and device based on a large language model, and relates to the technical field of intelligent questions.The security product pushing method comprises the steps that a security product knowledge graph instance is constructed for each security product, a user portrait of each user is constructed, field questions are collected, the field questions are analyzed through the large language model, and user intention categories are obtained; if the user intention type is a product suggestion type, obtaining a product content feature vector based on the security product knowledge graph instance of each security product; and training a purchase intention prediction model based on the user portrait of each user and the product content feature vector of the security product, selecting and recommending the security product for the current user based on the purchase intention prediction model and the user portrait of the current user, deeply mining user demands, and realizing a dynamic personalized recommendation service.

Description

Technical Field

[0001] The present invention relates to the field of intelligent question answering technology, and in particular to a method and device for pushing security products based on a large language model. Background Art

[0002] In recent years, with the continuous improvement of people's living standards, the awareness of security products has gradually increased, and the demand for security products has increased day by day. At the same time, the development of Internet technology has also brought great changes to the security product industry. Online security product sales platforms have emerged, providing users with a more convenient and flexible channel to purchase security products. However, the existing online security product sales system has obvious deficiencies in personalized recommendations and user understanding capabilities.

[0003] Although the traditional offline security product sales model can deeply understand user needs through manual methods, it is inefficient and costly. The collaborative filtering recommendation algorithm currently commonly used by online security product sales platforms can only make simple recommendations based on users' historical behavior data, and lacks a deep understanding of users' real needs. These systems cannot fully explore and utilize the different needs of different users, and the security products recommended to users are often fixed types, lacking flexibility and pertinence; A Chinese patent application with application publication number CN117994001A discloses a method, device, equipment and storage medium for recommending security products, which relates to the field of data processing technology. It includes dividing the users of a security product company to obtain multiple user groups, and then obtaining positive sample data and negative sample data. The users corresponding to the negative sample data are users who are not insured in the security product company. An input sample set is obtained based on the positive sample data and the negative sample data. After the data in the input sample set is converted into feature data that meets the model input standard, the feature data that meets the model input standard is input into a pre-trained recommendation model to obtain a list of recommended security products for each user in the target user group. Finally, the security products in the recommended security product list are sorted to obtain a final security product recommendation list for each user in the target user group. However, this method only analyzes users and fails to consider the content differences in the security products themselves. To this end, the present invention proposes a method and device for pushing security products based on a large language model. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes a security product push method and device based on a large language model, deeply explores user needs, and realizes dynamic personalized recommendation services.

[0005] To achieve the above purpose, a security product push method based on a large language model is proposed, which includes the following steps: Step 1: pre-define the core concepts and concept relationship types in the security product field, and collect security product field knowledge data for various security products; Step 2: Based on the core concepts and concept relationship types, extract knowledge entities and knowledge relationships from the knowledge data in the security product field, and build a security product knowledge graph instance for each security product; Step 3: Collect each user's historical interaction data, user questionnaire data, and user order data, and build a user profile for each user based on the user's historical interaction data and user questionnaire data; Step 4: Collect the domain questions input by the current user, and use the large language model to analyze the domain questions to obtain the user intent category; determine whether the user intent category is a product suggestion type. If it is a product suggestion type, go to step 5; otherwise, reply according to the preset reply strategy; Step 5: Based on the security product knowledge graph instance of each security product, obtain the product content feature vector; based on the user profile of each user and the product content feature vector of the security product, train the purchase intention prediction model; based on the purchase intention prediction model and the user profile of the current user, select and recommend security products for the current user; The pre-definition of the core concepts and concept relationship types in the field of security products includes the following steps: Step 11: Based on common knowledge in the field of security products, determine the concept categories of all core concepts in the field of security products; Step 12: Based on common knowledge in the field of security products, determine the type of conceptual relationships between the core concepts; The construction of a security product knowledge graph instance for each security product includes the following steps: Step 21: Use a named entity recognition model to extract concept entities from the security product domain knowledge data of each security product; Step 22: Extract concept entity attributes from the security product domain knowledge data of each security product by parsing the data table; the concept entity attributes are the attributes possessed by the entity of each core concept, Step 23: Use the relation extraction model to perform relation extraction on the security product domain knowledge data to mine the semantic association relationship between each conceptual entity; the semantic association relationship is included in the conceptual relationship type; Step 24: Organize the core concepts and concept relationship types into a unified knowledge graph ontology framework; Step 25: For each security product, the extracted concept entities are used as concept instance nodes, and the semantic associations between the extracted concept entities are used as edges connecting the concept instance nodes to construct a security product knowledge graph instance; The user historical interaction data of each user is collected in the following manner: Collecting interactive text data input by each user into the question-answering device each time; For each interactive text data, use the intent classification model to identify the user demand type; For each interactive text data, a named entity recognition model is used to identify the target security product type that the user is asking about; For each interactive text data, use the sentiment classification model to identify the user's sentiment category; Count the number of user demand types, target safety product types, and emotion categories in all interactive text data; The number of each user is marked as i, the number of each demand type of the i-th user is marked as Ai, the number of each target safety product type of the i-th user is marked as Bi, and the number of each emotion type of the i-th user is marked as Ci; User historical interaction data includes demand type set Ai, target product set Bi and emotion type set Ci; The user questionnaire data of each user is collected in the following manner: Before the user uses the question-answering device, a questionnaire form filled out by each user is collected in advance to collect each user's basic personal information as user questionnaire data; The user order data of each user is collected in the following manner: Collect the security product types of security products that each user has purchased, and the security product knowledge graph instances corresponding to each security product type purchased; The collection of all security product knowledge graph instances of each user constitutes the user order data; The method of obtaining the product content feature vector based on the security product knowledge graph instance of each security product is as follows: For each security product instance: Build a knowledge graph embedding model; Use the knowledge graph embedding model to learn low-dimensional continuous vector representations for each concept entity and the relationship between concept entities in the security product knowledge graph instance; Aggregate the low-dimensional continuous vector of each concept entity and the low-dimensional continuous vector of the relationship to generate a fixed-dimensional product content feature vector; The method of training the purchase intention prediction model for each security product based on the user portrait of each user and the product content feature vector of the security product is as follows: For each security product type: Construct N training sample vectors for each user, wherein the training sample vectors are formed by concatenating the user's demand type set, target product set, emotion type set, user questionnaire data, and the product content feature vector of the nth security product; wherein N is the number of security products included in the security product type, n=1,2,...N; Set a purchase intention label for each user's nth training sample vector. When the user purchases the nth safety product, the purchase intention label is 1, otherwise, the purchase intention label is 0. All training sample vectors are used as inputs of a purchase intention prediction model, the purchase intention prediction model uses the predicted value of purchase intention corresponding to each training sample vector as output, and the predicted value of purchase intention is a non-negative number in the range of [0,1]; the purchase intention prediction model uses the purchase intention label corresponding to each training sample vector as a prediction target, uses the difference between the predicted value of purchase intention and the purchase intention label as a prediction error, and uses minimizing the mean sum of squares of prediction errors of each prediction value of purchase intention as a training target; the purchase intention prediction model is trained until the mean sum of squares of prediction errors reaches convergence, and the training is stopped; the purchase intention prediction model is any one of the classification models; The method of selecting and recommending security products for the current user based on the purchase intention prediction model and the user profile of the current user is: The security product types included in the domain problem are identified by using a named entity recognition model as the target security product types; Read the product content feature vector of each security product in the target security product type to form a target product content feature vector set; Based on the user profile of the current user and the set of target product content feature vectors, a set of purchase intention vectors is constructed for each security product; Input each group of purchase intention vectors into the purchase intention prediction model to obtain the purchase intention prediction value output by the purchase intention prediction model; Among all safety products, the safety product with the largest predicted value of purchase intention is selected as the recommended safety product.

[0006] A safety product push device based on a large language model is proposed, including a graph construction module, a user portrait construction module and a safety product recommendation module; wherein each module is electrically connected to each other; The graph construction module predefines the core concepts and concept relationship types in the security product field, collects the security product field knowledge data of various security products, extracts knowledge entities and knowledge relationships from the security product field knowledge data based on the core concepts and concept relationship types, constructs a security product knowledge graph instance for each security product, and sends the security product knowledge graph instance to the security product recommendation module; The user portrait building module collects each user's historical interaction data, user questionnaire data, and user order data, builds a user portrait for each user based on the user's historical interaction data and user questionnaire data, and sends the user portrait to the security product recommendation module; The security product recommendation module collects the domain questions input by the current user, and uses a large language model to analyze the domain questions to obtain the user intention category; determines whether the user intention category is a product recommendation type. If it is a product recommendation type, obtains the product content feature vector based on the security product knowledge graph instance of each security product; trains a purchase intention prediction model based on the user portrait of each user and the product content feature vector of the security product, and selects recommended security products for the current user based on the purchase intention prediction model and the user portrait of the current user.

[0007] An electronic device is proposed, comprising: a processor and a memory, wherein the memory stores a computer program that can be called by the processor; The processor executes the above-mentioned security product push method based on the large language model by calling the computer program stored in the memory.

[0008] A computer-readable storage medium is provided, on which a rewritable computer program is stored; When the computer program is executed on a computer device, the computer device is enabled to execute the above-mentioned method for pushing security products based on a large language model.

[0009] Compared with the prior art, the present invention has the following beneficial effects: The present invention predefines the core concepts and concept relationship types in the field of security products, collects security product field knowledge data of various security products, extracts knowledge entities and knowledge relationships from the security product field knowledge data based on the core concepts and concept relationship types, constructs a security product knowledge graph instance for each security product, collects each user's historical user interaction data, user questionnaire data and user order data, and constructs a user portrait of each user based on the user's historical interaction data and user questionnaire data, collects the field questions input by the current user, and uses a large language model to analyze the field questions to obtain the user intent category; determines whether the user intent category is a product Recommendation type: If it is a product recommendation type, obtain the product content feature vector based on the security product knowledge graph instance of each security product; train the purchase intention prediction model based on the user portrait of each user and the product content feature vector of the security product, and select and recommend security products for the current user based on the purchase intention prediction model and the user portrait of the current user; for the scenario of online security product sales platform, capture structured domain knowledge by building a security product knowledge graph, use a large model to analyze the user's natural language expression to build a dynamic portrait, and finally make accurate recommendations based on the similarity between the knowledge vector and the user vector, deeply explore user needs, and realize dynamic personalized recommendation services. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a flow chart of a method for pushing security products based on a large language model in Embodiment 1 of the present invention; Figure 2 This is a module connection relationship diagram of a security product push device based on a large language model in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device in Embodiment 3 of the present invention; Figure 4 This is a schematic diagram of the computer-readable storage medium structure in Example 4 of the present invention. DETAILED DESCRIPTION

[0011] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0012] Example 1

[0013] like Figure 1 As shown, the method for pushing security products based on a large language model includes the following steps: Step 1: pre-define the core concepts and concept relationship types in the security product field, and collect security product field knowledge data for various security products; Step 2: Based on the core concepts and concept relationship types, extract knowledge entities and knowledge relationships from the knowledge data in the security product field, and build a security product knowledge graph instance for each security product; Step 3: Collect each user's historical interaction data, user questionnaire data, and user order data to build a user profile for each user; Step 4: Collect the domain questions input by the current user, and use the large language model to analyze the domain questions to obtain the user intent category; determine whether the user intent category is a product suggestion type. If it is a product suggestion type, go to step 5; otherwise, reply according to the preset reply strategy; Step 5: Based on the security product knowledge graph instance of each security product, obtain the product content feature vector; based on the user profile of each user and the product content feature vector of the security product, train the purchase intention prediction model; based on the purchase intention prediction model and the user profile of the current user, select and recommend security products for the current user; The pre-defining of the core concepts and concept relationship types in the field of security products includes the following steps: Step 11: Based on common knowledge in the field of security products, determine the concept categories of all core concepts in the field of security products; Specifically, the concept categories of the core concepts may include, but are not limited to, safety product number, safety product type, personnel role, safety product event, insurance policy element, etc.; Step 12: Based on common knowledge in the field of security products, determine the type of conceptual relationships between the core concepts; Specifically, the conceptual relationship type may be, for example: there is a relationship of insured person or beneficiary between the security product type and the personnel role class; Furthermore, the method of collecting safety product domain knowledge data is: Retrieving a professional text database in the field of security products for each security product from the backend database of the Internet, security product companies or security product organizations as security product field knowledge data; It is understandable that the safety product field knowledge data may include the full text of laws and regulations related to safety products, descriptions and rulings of various safety product dispute cases, detailed terms of each safety product business launched by the safety product company, insurance policies for each safety product business, customer notices, etc.; Furthermore, the process of extracting knowledge entities and knowledge relationships from the knowledge data in the field of security products based on core concepts and concept relationship types, and constructing a security product knowledge graph instance for each security product includes the following steps: Step 21: Use a named entity recognition model to extract concept entities from the safety product domain knowledge data of each safety product; the concept entity refers to a concrete instance of each core concept, for example, the safety product types include specific major disease medical safety products, auto insurance, pension safety products, family medical safety products, etc., and synonyms and near synonyms can also be normalized into the same standard entity; Step 22: Extract concept entity attributes from the safety product domain knowledge data of each safety product by parsing the data table; the concept entity attributes are the attributes possessed by the entity of each core concept, for example, the attributes corresponding to a major disease medical safety product may include the safety product amount, premium amount, effective date, expiration date, etc.; Step 23: Use the relation extraction model to extract relations from the security product domain knowledge data to mine the semantic association relations between each conceptual entity; the semantic association relations are included in the conceptual relationship type; specifically, mining the semantic association relations also includes identifying common relationship patterns, such as "include", "satisfy", "restrict", etc., and marking the exact semantics of the association relations between each two associated conceptual entities to build a relationship classification system; Step 24: Organize the core concepts and concept relationship types into a unified knowledge graph ontology framework to define concept categories, attribute categories, relationship categories and their constraints; Step 25: For each security product, the extracted concept entities are used as concept instance nodes, and the semantic associations between the extracted concept entities are used as edges connecting the concept instance nodes to construct a security product knowledge graph instance; Furthermore, the user historical interaction data of each user is collected in the following manner: Collecting interactive text data input by each user into the question-answering device each time; the interactive text data includes inquiries, replies, comments, etc.; For each interactive text data, the intent classification model is used to identify the user demand type; the user demand type includes but is not limited to inquiries about security product terms, security product validity period, premiums, security product status, insurance conditions, and other inquiry needs involving all aspects of security product business; For each interactive text data, a named entity recognition model is used to identify the target safety product type that the user is asking about. It is understandable that when users generally ask about safety products, they ask questions with a purpose. For example, when preparing to purchase medical safety products for themselves, they will ask questions about safety products of the major disease medical safety product type. For each interactive text data, an emotion classification model is used to identify the emotion category of the user; the emotion category includes but is not limited to confusion, anxiety, suspicion, dissatisfaction, expectation, hesitation, etc.; it can be understood that the emotion category of the user input text can reflect the personality characteristics of each user; Count the number of user demand types, target safety product types, and emotion categories in all interactive text data; The number of each user is marked as i, the number of each demand type of the i-th user is marked as Ai, the number of each target safety product type of the i-th user is marked as Bi, and the number of each emotion type of the i-th user is marked as Ci; User historical interaction data includes demand type set Ai, target product set Bi and emotion type set Ci; Furthermore, the user questionnaire data of each user is collected in the following manner: Before the user uses the question-answering device, a questionnaire form filled out by each user is collected in advance to collect each user's basic personal information as user questionnaire data; Specifically, the basic personal information refers to the basic information of the user, the insured person or the insured property that needs to be known when recommending safety products to the user, including but not limited to name, age, gender, marital status, occupation, income level, asset status, debt status, number of family members, children status, support needs, health status of the user and his family, coverage of existing safety products, risk tolerance and risk avoidance, etc.; Furthermore, the user order data of each user is collected in the following manner: Collect the security product types of security products that each user has purchased, and the security product knowledge graph instances corresponding to each security product type purchased; The collection of all security product knowledge graph instances of each user constitutes the user order data; Furthermore, the method of collecting the domain questions input by the current user and analyzing the domain questions using the large language model to obtain the user intent category is as follows: The question-answering device backend server receives a question text in the field of security products input by a user currently using the question-answering device as a field question; Pre-collect a set of user intention types in the field of security products; specifically, the user intention types in the set of user intention types include but are not limited to product suggestion type, insured query, information query type, insurance consultation type, claim application type, product detail type, etc.; Collect question sample libraries in advance, and manually annotate the user intent type for each question sample to form an annotated corpus containing question sample-user intent type pairs; Using each question sample in the annotated corpus as input and the user intent type as the supervised training target, a large language model is trained to identify the user intent type corresponding to the question text; the large language model is any one of BERT, GPT, and T5; Furthermore, a preferred embodiment of replying according to a preset reply strategy may be as follows: When the user intention category is "insurance query", the question-answering device queries from the background and outputs to the user the specific information of the safety products that the current user has purchased; the specific information may be all the information in the safety product knowledge graph instance corresponding to the purchased safety product; Furthermore, the method of obtaining the product content feature vector based on the security product knowledge graph instance of each security product is as follows: For each security product instance: Construct a knowledge graph embedding model; the knowledge graph embedding model includes TransE, RotatE, etc.; Use the knowledge graph embedding model to learn low-dimensional continuous vector representations for each concept entity and the relationship between concept entities in the security product knowledge graph instance; Aggregate the low-dimensional continuous vector of each concept entity and the low-dimensional continuous vector of the relationship to generate a fixed-dimensional product content feature vector; Furthermore, the method of training the purchase intention prediction model for each security product based on the user portrait of each user and the product content feature vector of the security product is as follows: For each security product type: Construct N training sample vectors for each user, wherein the training sample vectors are formed by concatenating the user's demand type set, target product set, emotion type set, user questionnaire data, and the product content feature vector of the nth security product; wherein N is the number of security products included in the security product type, n=1,2,...N; Set a purchase intention label for each user's nth training sample vector. When the user purchases the nth safety product, the purchase intention label is 1, otherwise, the purchase intention label is 0. All training sample vectors are used as inputs of a purchase intention prediction model, the purchase intention prediction model uses the predicted value of purchase intention corresponding to each training sample vector as output, and the predicted value of purchase intention is a non-negative number in the range of [0,1]; the purchase intention prediction model uses the purchase intention label corresponding to each training sample vector as a prediction target, uses the difference between the predicted value of purchase intention and the purchase intention label as a prediction error, and uses minimizing the mean sum of squares of prediction errors of each prediction value of purchase intention as a training target; the purchase intention prediction model is trained until the mean sum of squares of prediction errors reaches convergence, and the training is stopped; the purchase intention prediction model is any one of the classification models, including but not limited to a DNN model, a SVM model, etc.; Furthermore, the method of selecting and recommending security products for the current user based on the purchase intention prediction model and the user portrait of the current user is: The security product types included in the domain problem are identified by using a named entity recognition model as the target security product types; Read the product content feature vector of each security product in the target security product type to form a target product content feature vector set; According to the user portrait of the current user and the target product content feature vector set, a set of purchase intention vectors is constructed for each security product; it can be understood that the construction method of the purchase intention vector is consistent with the construction method of the training sample vector; Input each group of purchase intention vectors into the purchase intention prediction model to obtain the purchase intention prediction value output by the purchase intention prediction model; Among all safety products, the safety product with the largest predicted value of purchase intention is selected as the recommended safety product.

[0014] Example 2

[0015] like Figure 2 As shown, a security product push device based on a large language model includes a graph construction module, a user portrait construction module and a security product recommendation module; wherein each module is electrically connected; The graph construction module predefines the core concepts and concept relationship types in the security product field, collects the security product field knowledge data of various security products, extracts knowledge entities and knowledge relationships from the security product field knowledge data based on the core concepts and concept relationship types, constructs a security product knowledge graph instance for each security product, and sends the security product knowledge graph instance to the security product recommendation module; The user portrait building module collects each user's historical interaction data, user questionnaire data, and user order data, builds a user portrait for each user based on the user's historical interaction data and user questionnaire data, and sends the user portrait to the security product recommendation module; The security product recommendation module collects the domain questions input by the current user, and uses a large language model to analyze the domain questions to obtain the user intention category; determines whether the user intention category is a product recommendation type. If it is a product recommendation type, obtains the product content feature vector based on the security product knowledge graph instance of each security product; trains a purchase intention prediction model based on the user portrait of each user and the product content feature vector of the security product, and selects recommended security products for the current user based on the purchase intention prediction model and the user portrait of the current user.

[0016] Example 3

[0017] Figure 3 1 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present application. Figure 3 As shown, according to another aspect of the present application, an electronic device 100 is also provided. The electronic device 100 may include one or more processors and one or more memories. The memories store computer readable codes, and when the computer readable codes are executed by one or more processors, the method for pushing security products based on a large language model as described above may be executed.

[0018] The method or device according to the embodiment of the present application can also be used by Figure 3 The electronic device architecture shown in FIG. Figure 3 As shown, the electronic device 100 may include a bus 101, one or more CPUs 102, ROM 103, RAM 104, a communication port 105 connected to a network, an input / output component 106, a hard disk 107, etc. The storage device in the electronic device 100, such as ROM 103 or hard disk 107, may store the security product push method based on a large language model provided in the present application.

[0019] Furthermore, the electronic device 100 may further include a user interface 108. Figure 3 The architecture shown is only exemplary and can be omitted according to actual needs when implementing different devices. Figure 3 One or more components of an electronic device are shown.

[0020] Example 4

[0021] Figure 4 Schematic diagram of a computer-readable storage medium structure provided by an embodiment of the present application. Figure 4As shown, a computer-readable storage medium 200 according to an embodiment of the present application is shown. Computer-readable instructions are stored on the computer-readable storage medium 200. When the computer-readable instructions are executed by the processor, the method for pushing a secure product based on a large language model according to an embodiment of the present application described with reference to the above figures can be executed. The computer-readable storage medium 200 includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory (cache), etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0022] In addition, according to the implementation of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the present application provides a non-transitory machine-readable storage medium, the non-transitory machine-readable storage medium stores machine-readable instructions, the machine-readable instructions can be run by a processor to execute instructions corresponding to the method steps provided by the present application, and when the computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present application are executed.

[0023] The methods, apparatuses, and devices of the present application may be implemented in many ways. For example, the methods, apparatuses, and devices of the present application may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified. In addition, in some embodiments, the present application may also be implemented as programs recorded in a recording medium, which include machine-readable instructions for implementing the method according to the present application. Therefore, the present application also covers recording media storing programs for executing the method according to the present application.

[0024] In addition, the parts of the above-mentioned technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0025] The specific implementation modes as described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation mode of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0026] The above preset parameters or preset thresholds are all set by technicians in this field according to actual conditions or obtained through large-scale data simulation.

[0027] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A security product push method based on a large language model, characterized in that: The following steps are involved: Step 1: pre-define the core concepts and concept relationship types in the security product field, and collect security product field knowledge data for various security products; Step 2: Based on the core concepts and concept relationship types, extract knowledge entities and knowledge relationships from the knowledge data in the security product field, and build a security product knowledge graph instance for each security product; Step 3: Collect each user's historical interaction data, user questionnaire data, and user order data, and build a user profile for each user based on the user's historical interaction data and user questionnaire data; Step 4: Collect the domain questions input by the current user and use the large language model to analyze the domain questions to obtain the user intent category; Determine whether the user intent category is a product suggestion type. If so, go to step 5; otherwise, respond according to the preset response strategy; Step 5: Based on the security product knowledge graph instance of each security product, obtain the product content feature vector; based on the user portrait of each user and the product content feature vector of the security product, train the purchase intention prediction model; based on the purchase intention prediction model and the user portrait of the current user, select and recommend security products for the current user.

2. The method for pushing security products based on a large language model according to claim 1, characterized in that: The pre-definition of the core concepts and concept relationship types in the field of security products includes the following steps: Step 11: Based on common knowledge in the field of security products, determine the concept categories of all core concepts in the field of security products; Step 12: Based on common knowledge in the field of security products, determine the type of conceptual relationships between the core concepts.

3. The method for pushing security products based on a large language model according to claim 2, characterized in that: The construction of a security product knowledge graph instance for each security product includes the following steps: Step 21: Use a named entity recognition model to extract concept entities from the security product domain knowledge data of each security product; Step 22: Extract concept entity attributes from the security product domain knowledge data of each security product by parsing the data table; the concept entity attributes are the attributes possessed by the entity of each core concept, Step 23: Use the relation extraction model to perform relation extraction on the security product domain knowledge data to mine the semantic association relationship between each conceptual entity; the semantic association relationship is included in the conceptual relationship type; Step 24: Organize the core concepts and concept relationship types into a unified knowledge graph ontology framework; Step 25: For each security product, the extracted concept entities are used as concept instance nodes, and the semantic association relationships between the extracted concept entities are used as edges connecting the concept instance nodes to construct a security product knowledge graph instance.

4. The method for pushing security products based on a large language model according to claim 3, characterized in that: The user historical interaction data of each user is collected in the following manner: Collecting interactive text data input by each user into the question-answering device each time; For each interactive text data, use the intent classification model to identify the user demand type; For each interactive text data, a named entity recognition model is used to identify the target security product type that the user is asking about; For each interactive text data, use the sentiment classification model to identify the user's sentiment category; Count the number of user demand types, target safety product types, and emotion categories in all interactive text data; The number of each user is marked as i, the number of each demand type of the i-th user is marked as Ai, the number of each target safety product type of the i-th user is marked as Bi, and the number of each emotion type of the i-th user is marked as Ci; User historical interaction data includes demand type set Ai, target product set Bi and emotion type set Ci.

5. The method for pushing security products based on a large language model according to claim 4, characterized in that: The user questionnaire data of each user is collected in the following manner: Before the user uses the question-and-answer device, a questionnaire form filled out by each user is collected in advance to collect each user's basic personal information as user questionnaire data.

6. The method for pushing security products based on a large language model according to claim 5, characterized in that: The user order data of each user is collected in the following manner: Collect the security product types of security products that each user has purchased, and the security product knowledge graph instances corresponding to each security product type purchased; The collection of all security product knowledge graph instances of each user constitutes the user order data.

7. The method for pushing security products based on a large language model according to claim 6, characterized in that: The method of obtaining the product content feature vector based on the security product knowledge graph instance of each security product is as follows: For each security product instance: Build a knowledge graph embedding model; Use the knowledge graph embedding model to learn low-dimensional continuous vector representations for each concept entity and the relationship between concept entities in the security product knowledge graph instance; The low-dimensional continuous vector of each concept entity and the low-dimensional continuous vector of the relationship are aggregated to generate a fixed-dimensional product content feature vector.

8. The method for pushing security products based on a large language model according to claim 7, characterized in that: The method of training the purchase intention prediction model for each security product based on the user portrait of each user and the product content feature vector of the security product is as follows: For each security product type: Construct N training sample vectors for each user, wherein the training sample vectors are formed by concatenating the user's demand type set, target product set, emotion type set, user questionnaire data, and the product content feature vector of the nth security product; wherein N is the number of security products included in the security product type, n=1,2,...N; Set a purchase intention label for each user's nth training sample vector. When the user purchases the nth safety product, the purchase intention label is 1, otherwise, the purchase intention label is 0. All training sample vectors are used as inputs of a purchase intention prediction model, the purchase intention prediction model uses the predicted value of purchase intention corresponding to each training sample vector as output, and the predicted value of purchase intention is a non-negative number in the range of [0,1]; the purchase intention prediction model uses the purchase intention label corresponding to each training sample vector as a prediction target, the difference between the predicted value of purchase intention and the purchase intention label as a prediction error, and minimizing the sum of squares of the prediction errors of the predicted values ​​of each purchase intention as a training target; the purchase intention prediction model is trained until the sum of squares of the prediction errors converges and the training is stopped; the purchase intention prediction model is any one of the classification models.

9. The method for pushing security products based on a large language model according to claim 8, characterized in that: The method of selecting and recommending security products for the current user based on the purchase intention prediction model and the user profile of the current user is: The security product types included in the domain problem are identified by using a named entity recognition model as the target security product types; Read the product content feature vector of each security product in the target security product type to form a target product content feature vector set; Based on the user profile of the current user and the set of target product content feature vectors, a set of purchase intention vectors is constructed for each security product; Input each group of purchase intention vectors into the purchase intention prediction model to obtain the purchase intention prediction value output by the purchase intention prediction model; Among all safety products, the safety product with the largest predicted value of purchase intention is selected as the recommended safety product.

10. A device for pushing a safe product based on a large language model, which is used to implement the method for pushing a safe product based on a large language model as claimed in any one of claims 1 to 9, characterized in that: It includes a graph construction module, a user portrait construction module and a security product recommendation module; wherein each module is electrically connected to another; The graph construction module predefines the core concepts and concept relationship types in the security product field, collects the security product field knowledge data of various security products, extracts knowledge entities and knowledge relationships from the security product field knowledge data based on the core concepts and concept relationship types, constructs a security product knowledge graph instance for each security product, and sends the security product knowledge graph instance to the security product recommendation module; The user portrait building module collects each user's historical interaction data, user questionnaire data, and user order data, builds a user portrait for each user based on the user's historical interaction data and user questionnaire data, and sends the user portrait to the security product recommendation module; The security product recommendation module collects the domain questions input by the current user, and uses a large language model to analyze the domain questions to obtain the user intention category; determines whether the user intention category is a product recommendation type. If it is a product recommendation type, obtains the product content feature vector based on the security product knowledge graph instance of each security product; trains a purchase intention prediction model based on the user portrait of each user and the product content feature vector of the security product, and selects recommended security products for the current user based on the purchase intention prediction model and the user portrait of the current user.

11. An electronic device, characterized in that: include: processor and memory, wherein The memory stores a computer program that can be called by the processor; The processor executes the large language model-based security product push method described in any one of claims 1 to 9 in the background by calling the computer program stored in the memory.

12. A computer-readable storage medium, characterized in that: A rewritable computer program is stored thereon; When the computer program runs on a computer device, the computer device executes the method for pushing a security product based on a large language model as described in any one of claims 1 to 9 in the background.

Images