Business plan generation method and device based on large language model, equipment, medium

Abstract

The embodiment of the application provides a business scheme generation method and device based on a large language model, equipment and a medium, and belongs to the technical field of artificial intelligence. The method comprises the following steps: in response to a user business request, performing information extraction on user dialogue information based on a large language model, performing demand sorting on user demand entities based on risk prediction data, constructing a multi-dimensional demand graph based on user demand entities, entity association relationships and demand priorities, the level of a graph node being determined by a demand priority, screening a candidate product matched with a demand category from an insurance knowledge base based on the multi-dimensional demand graph, and generating a scheme for the candidate product based on the large language model. The application can be applied to a financial technology scene, a dynamic demand graph is constructed by using the context understanding capability of an LLM, and upgrading from traditional static label matching to deep mining and sorting of implicit demands of a user is realized, and a change from single product pushing to a transparent customized report of trinity is realized.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology and is applied to financial technology scenarios, particularly to a method, apparatus, device, and medium for generating business solutions based on a large language model. Background Technology

[0002] In the insurance service sector within fintech, intelligent customer service systems and product recommendations can enhance user experience and reduce operating costs. However, the current financial industry faces several bottlenecks: First, intelligent customer service suffers from shallow understanding and rigid responses. Traditional solutions rely on FAQ keyword matching or rule engines, failing to parse implicit user needs, resulting in generated proposals that don't meet user requirements and a fragmented user experience. Second, current intelligent customer service systems primarily push products based on static tags such as age and income, lacking contextual adaptation to user family structure, debt burden, and risk preferences. This leads to the inability to generate timely proposals that meet user needs, and the insurance products recommended in the generated proposals are inaccurate and fail to meet user requirements, resulting in insufficient user conversion rates and severe resource misallocation. Therefore, improving the efficiency and accuracy of business proposal generation has become an urgent problem to be solved. Summary of the Invention

[0003] The main objective of this application is to propose a business solution generation method, apparatus, device, and medium based on a large language model, aiming to solve the technical problems of low efficiency and low accuracy in generating business solutions in the current technology, and to improve the efficiency and accuracy of business solution generation.

[0004] To achieve the above objectives, a first aspect of this application proposes a business solution generation method based on a large language model, the method comprising: In response to a user's business request, obtain user dialogue information; the user dialogue information is natural language information. Information is extracted from the user dialogue information based on a pre-trained large language model to obtain user demand entities and entity relationships; wherein, the user demand entities are used to represent demand categories. Based on risk prediction data, the user demand entities are sorted to obtain demand priorities; A multidimensional demand graph is constructed based on the user demand entity, the entity association relationship, and the demand priority; wherein, the user demand entity serves as the graph node of the multidimensional demand graph, the entity association relationship serves as the edge of the multidimensional demand graph, and the level of the graph node is determined by the demand priority. Based on the multidimensional demand graph, candidate products matching the demand categories are selected from a pre-built insurance knowledge base; Based on the large language model, a solution is generated for the candidate products to obtain a business solution document.

[0005] To achieve the above objectives, a second aspect of this application proposes a business solution generation apparatus based on a large language model, the apparatus comprising: The dialogue acquisition module is used to acquire user dialogue information in response to user business requests; the user dialogue information is natural language information. The information extraction module is used to extract information from the user dialogue information based on a pre-trained large language model to obtain user demand entities and entity associations; wherein, the user demand entities are used to represent demand categories. The demand sorting module is used to sort the user demand entities based on risk prediction data to obtain demand priorities. The demand graph construction module is used to construct a multi-dimensional demand graph based on the user demand entity, the entity relationship, and the demand priority; wherein, the user demand entity is the graph node of the multi-dimensional demand graph, the entity relationship is the edge of the multi-dimensional demand graph, and the level of the graph node is determined by the demand priority. The product screening module is used to select candidate products that match the demand categories from a pre-built insurance knowledge base based on the multi-dimensional demand graph. The proposal generation module is used to generate proposals for the candidate products based on the large language model, thereby obtaining business proposals.

[0006] To achieve the above objectives, a third aspect of the present application provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method described in the first aspect.

[0007] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0008] The business solution generation method, apparatus, device, and medium based on a large language model proposed in this application respond to user business requests by acquiring user dialogue information based on natural language. Based on a large language model, it extracts user demand entities and entity relationships between these entities from the dialogue information to represent demand categories. Furthermore, it prioritizes user demand entities based on risk prediction data to determine demand priorities. This prioritizes the nodes in a multi-dimensional demand graph, which in turn determines their level. Based on this dynamic multi-dimensional demand graph, it filters candidate products matching the demand categories from an insurance knowledge base and generates a three-in-one business solution that meets user needs. This upgrades from traditional static tag matching to in-depth mining and demand prioritization of implicit user needs, transforming from single-product push to a transparent, three-in-one customized solution and improving the efficiency of solution generation. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the application environment for the business solution generation method based on a large language model provided in the embodiments of this application; Figure 2 This is a flowchart of the business solution generation method based on a large language model provided in the embodiments of this application; Figure 3 yes Figure 2 The flowchart for step 203 in the document; Figure 4 This is another flowchart of the business solution generation method based on a large language model provided in the embodiments of this application; Figure 5 This is another flowchart of the business solution generation method based on a large language model provided in the embodiments of this application; Figure 6 yes Figure 2 The flowchart for step 206 in the document; Figure 7 This is another flowchart of the business solution generation method based on a large language model provided in the embodiments of this application; Figure 8 This is a schematic diagram of the structure of the business solution generation device based on a large language model provided in the embodiments of this application; Figure 9 This is a schematic diagram of the hardware structure of the computer device provided in the embodiments of this application. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0011] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0012] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0013] First, let's analyze some of the terms used in this application: Artificial intelligence (AI) is a new branch of computer science that studies, develops, and applies theories, methods, technologies, and systems to simulate, extend, and expand human intelligence. It aims to understand the essence of intelligence and produce intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thought. Furthermore, AI utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to achieve optimal results.

[0014] Natural Language Processing (NLP): NLP uses computers to process, understand, and utilize human language (such as Chinese and English). NLP is a branch of artificial intelligence and an interdisciplinary field of computer science and linguistics, often referred to as computational linguistics. NLP includes syntactic analysis, semantic analysis, and discourse understanding. It is commonly used in machine translation, handwritten and printed character recognition, speech recognition and text-to-speech conversion, intent recognition, information extraction and filtering, text classification and clustering, sentiment analysis, and opinion mining. It involves data mining, machine learning, knowledge acquisition, knowledge engineering, artificial intelligence research, and linguistic research related to language computation.

[0015] In the insurance service sector within fintech, intelligent customer service systems and product recommendations can enhance user experience and reduce operating costs. However, the current financial industry faces the following bottlenecks: First, intelligent customer service suffers from shallow understanding and rigid responses. Traditional solutions rely on FAQ keyword matching or rule engines, failing to analyze implicit user needs. For example, complex requests like "I want to save for my child's education, but I'm worried about future unemployment affecting my payments" result in a 40%-60% conversion rate to human agents, leading to a fragmented user experience. Second, product recommendations are based on coarse tags and lack contextual understanding. Current intelligent customer service systems mostly push products based on static tags such as age and income, lacking contextual adaptation to user family structure, debt burden, and risk preferences. This results in inaccurate insurance product recommendations that fail to meet user needs, leading to insufficient user conversion rates and severe resource misallocation.

[0016] Based on this, embodiments of this application provide a business solution generation method, apparatus, device, and medium based on a large language model, aiming to improve the efficiency and accuracy of business solution generation.

[0017] The service solution generation method, apparatus, device, and medium based on a large language model provided in this application are specifically described through the following embodiments. First, the service solution generation method based on a large language model in this application embodiment is described.

[0018] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0019] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0020] The business solution generation method based on a large language model provided in this application relates to the field of artificial intelligence technology. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the business solution generation method based on a large language model, but is not limited to the above forms.

[0021] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0022] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user audio, user voice, user behavior, user historical data, and user attribute information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirects to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments obtained.

[0023] The business solution generation method based on a large language model provided in this application embodiment can be applied to, for example, Figure 1In this application environment, a communication connection is established between the user terminal and the server. In one application scenario, the user terminal can be an insurance company. Responding to user business requests, the user terminal obtains user dialogue information based on natural language, extracts information from the dialogue information using a pre-trained large language model, obtains user demand entities and entity relationships, and ranks user demand entities based on risk prediction data to obtain demand priorities. A multi-dimensional demand graph is then constructed based on the user demand entities, entity relationships, and demand priorities. User demand entities serve as nodes in the multi-dimensional demand graph, entity relationships serve as edges, and the level of a graph node is determined by the demand priority. Based on the multi-dimensional demand graph, candidate products matching the demand categories are selected from a pre-built insurance knowledge base. A solution is then generated from the candidate products using the large language model to obtain a business plan. The user terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The following detailed description of specific embodiments further illustrates this application.

[0024] Figure 2 This is an optional flowchart of the business solution generation method based on a large language model provided in the embodiments of this application. Figure 2 The method may include, but is not limited to, steps 201 to 206.

[0025] Step 201: In response to the user's business request, obtain the user's dialogue information; the user's dialogue information is natural language information. Step 202: Extract information from user dialogue information based on a pre-trained large language model to obtain user demand entities and entity relationships; Step 203: Based on the risk prediction data, sort the user demand entities to obtain demand priorities; Step 204: Construct a multi-dimensional demand graph based on user demand entities, entity relationships, and demand priorities; wherein, user demand entities serve as graph nodes of the multi-dimensional demand graph, entity relationships serve as edges of the multi-dimensional demand graph, and the level of the graph nodes is determined by the demand priorities. Step 205: Based on the multi-dimensional demand graph, candidate products that match the demand categories are selected from the pre-built insurance knowledge base; Step 206: Generate a solution for the candidate products based on the large language model to obtain a business solution document.

[0026] Steps 201 to 206, as illustrated in this embodiment, involve responding to user business requests, acquiring user dialogue information based on natural language, extracting information from the user dialogue information based on a pre-trained large language model to obtain user demand entities and entity relationships, ranking user demand entities based on risk prediction data to obtain demand priorities, and determining the level of nodes in a multi-dimensional demand graph based on demand priorities. This allows for the selection of candidate products matching the demand categories from the insurance knowledge base based on the multi-dimensional demand graph, and the generation of a three-in-one business plan that meets user needs. This upgrades from traditional static tag matching to in-depth mining and demand ranking of implicit user needs, realizing a transformation from single product push to a transparent, customized three-in-one plan, thereby improving the efficiency and accuracy of plan generation.

[0027] In step 201 of some embodiments, the user's natural language input is obtained through a user business request to obtain user dialogue information. In some application scenarios, user dialogue information may include: I just had my second child, we still have 20 years left on our mortgage, my husband is a programmer and often works overtime, and I want to buy insurance for the whole family. In another application scenario, user dialogue information may include: I want to save for my child's education, but I'm worried that future unemployment will affect my premium payments. In yet another application scenario, user dialogue information may include: What are the conditions for premium waiver?

[0028] In practical applications, user input can be either text or voice. When the user input is voice, the voice is first converted into text to obtain the user's dialogue information.

[0029] In step 202 of some embodiments, information is extracted from user dialogue information based on a large language model to obtain user demand entities and entity relationships, thereby facilitating the generation of a dynamic multidimensional demand graph in step 204 to uncover potential user needs. In some application scenarios, user demand entities may include user entities, such as personal characteristics, family structure, and financial status; risk entities, such as personal risk, property risk, and liability risk; product entities, such as insurance products, supplementary insurance, and service rights; financial entities, such as income, liabilities, assets, and cash flow; and protection needs and savings needs. Entity relationships may be, for example, risk-demand relationships, user-risk relationships, demand-product relationships, user-financial relationships, etc.

[0030] In one application scenario, extract user demand entities and entity relationships. Referring to Table 1, in the user dialogue information, the user inputs: "I just had my second child, and there are still 20 years left on the mortgage. My husband is a programmer and often works overtime. I want to buy insurance for the whole family." The extracted user demand entities include family structure, debt, occupational risk, and insurance preference. The family structure is: two-child family, couple + 2 children; the debt is: 20-year mortgage. Occupational risks are: programmers working overtime → potential critical illness risk; protection preference is: coverage for the whole family; the established correlation is: programmers working overtime → priority for critical illness insurance, outstanding mortgage → term life insurance coverage ≥ mortgage balance.

[0031]

[0032] Table 1 In some embodiments, risk prediction data includes the type of risk and the extent of loss. See also... Figure 3 In some embodiments, step 203 may include, but is not limited to: Step 301: Obtain user profile data and extract features from the user profile data to obtain user profile features; Step 302: Based on user demand entities and user profile characteristics, perform risk prediction to obtain the risk type and loss extent; Step 303: Sort the demands based on the type of risk and the degree of loss to obtain the demand priority.

[0033] In steps 301 to 303 of some embodiments, the acquired user profile data may include information such as user gender, age, preferences, and health status. Feature extraction is performed on the user profile data, resulting in user profile features that may include personal characteristics, family structure characteristics, health status characteristics, financial status characteristics, income characteristics, debt characteristics, and asset characteristics. Personal characteristics may include gender, age, preferences, and health status. Based on user demand entities and user profile features, risk prediction is performed to obtain risk prediction data. This risk prediction data includes risk type and loss severity. Risk types may include, for example, critical illness, accidental disability, hospitalization, education fund gaps, and insufficient retirement savings. Risk types can be represented by the probability of occurrence; for example, the probability of critical illness is 0.04. Demands can then be prioritized based on risk type and loss severity.

[0034] In step 203 of an application scenario, risk prediction and demand prioritization are implemented based on actuarial logic. For example, prioritizing protection before financial planning and adults before children, LLM prioritizes demand based on risk type and loss severity. This prioritization is based on the probability of occurrence of the corresponding risk type multiplied by the degree of loss. For example, the impact of mortgage default on family finances is greater than the short-term gap in education funds. In the aforementioned application scenario, the priority is: ① Adult term life insurance (covering mortgage risk) -> ② Family critical illness insurance (addressing health risks from overtime work) -> ③ Children's education funds (long-term savings).

[0035] In a specific application scenario, the user's dialogue information includes: 35 years old, internet engineer, annual income of 500,000 yuan, mortgage of 3 million yuan, one 5-year-old child, frequent overtime work, and a thyroid nodule detected during a physical examination; needs: want to buy insurance for themselves and their child, and also consider saving for their child's education. Risk prediction yields: probability of death ≈ 0.005 × loss severity 90 points = 0.45 points, with a weight of 1.5, resulting in a final score of 0.675 points; probability of critical illness ≈ 0.04 × loss severity 75 points = 3 points, with a weight of 1.3, resulting in a final score of 3.9 points; probability of education savings ≈ 0.6 × loss severity 20 points = 12 points, with a weight of 0.7, resulting in a final score of 8.4 points. Therefore, the needs are ranked as follows: term life insurance, critical illness insurance, and education annuity. Term life insurance (0.675 points) has a low probability but a very high potential loss; critical illness insurance (3.9 points) has a relatively high probability and a significant potential loss; and education annuity (8.4 points) has a high probability but a manageable potential loss. The business plan generated in the final step 206 may include the contents shown in Table 2.

[0036]

[0037] Table 2 In step 204 of some embodiments, the multidimensional demand graph has graph nodes and node weights. User demand entities are used as graph nodes of the multidimensional demand graph, entity relationships are used as edges of the multidimensional demand graph, and the graph is propagated based on demand priority to construct the multidimensional demand graph.

[0038] Compared to traditional product recommendation technologies that rely on explicit tags (such as age, income, etc.), this application embodiment analyzes the user dialogue context through a large language model, constructs a multi-dimensional dynamic demand graph, captures implicit demands, and generates a dynamic multi-dimensional demand graph to uncover users' potential needs.

[0039] Please see Figure 4 Following step 204, this business solution generation method based on a large language model also includes updating the multi-dimensional requirement graph, specifically including: Step 401: Obtain the dialogue context from the multi-turn dialogue; Step 402: Perform intent recognition on the dialogue context to obtain the current user intent; Step 403: Adjust the node weights based on the current user intent to obtain the current weights; Step 404: Update the multidimensional demand graph based on the current weights.

[0040] In steps 401 to 404 of some embodiments, the dialogue context is obtained from the multi-turn dialogue with the user, and intent recognition is performed based on the dialogue context to obtain the latest current user intent. The node weights of the multidimensional demand graph are then corrected based on the current user intent to update the multidimensional demand graph according to the corrected current weights. In one application scenario, for example, if the user adds "the budget does not exceed 80,000 yuan per year," the LLM updates the node weights of the multidimensional demand graph in real time, for example, increasing the weight of "budget constraint" to the highest level to ensure that subsequent recommendations strictly match the latest demands.

[0041] The business solution generation method based on a large language model provided in this application supports the correction of user needs in multi-turn dialogues, which facilitates the subsequent generation of business solution documents that better meet user needs, so as to realize the recommendation of insurance products.

[0042] Please see Figure 5 Before step 202 in some embodiments, the business solution generation method based on a large language model further includes optimizing the large language model, specifically including: Step 501: Obtain multi-source corpus; Step 502: Clean the multi-source corpus to obtain multi-source cleaned data; wherein, the multi-source cleaned data includes historical interaction data; Step 503: Label the historical interaction data to obtain labeled interaction data, and filter out candidate interaction data from the labeled interaction data; Step 504: Calculate the reward for the candidate interaction data based on the preset reward function to obtain the interaction score; Step 505: Optimize the large language model based on interactive scoring.

[0043] In step 501 of some embodiments, the multi-source corpus may include professional knowledge, such as insurance-related clauses and instructions, such as insurance contract terms, regulatory provisions, product manuals, etc.; the multi-source corpus may also include historical interaction data, such as at least 100,000 real customer service dialogues, including labeled user intent, emotions, unresolved issues, etc., where emotions can be anxiety, hesitation, etc., and unresolved issues can include consultation, insurance application, claims, etc.; the multi-source corpus may also include an industry case library, such as typical insurance service scenarios, such as "whether to claim for abnormal physical examination during the waiting period", sales misrepresentation cases, etc.

[0044] In step 502 of some embodiments, cleaning the multi-source corpus may include deduplication, anonymization, or structuring. Deduplication may include removing duplicate FAQs, anonymization may include hiding sensitive information such as user names and ID numbers, and structuring may include labeling by “question-standard answer-compliance points”. The resulting cleaned multi-source data can be used as a high-quality dataset for fine-tuning large language models.

[0045] In steps 503 to 505 of some embodiments, historical interaction data is labeled to obtain labeled interaction data. Candidate interaction data selected from the labeled interaction data are high-quality answers, which are then used as reward signals to optimize the output of the large language model using the PPO (Proximal Policy Optimization) algorithm. Specifically, rewards are calculated on the candidate interaction data based on a preset reward function to obtain an interaction score, and the large language model is optimized based on the interaction score. In practical applications, instruction tuning can be employed. Specifically, based on the LLaMA-2-13B or ChatGLM3-6B base model, task-oriented instruction formats can be used for fine-tuning. For example, for the "product comparison" task, an instruction template can be designed as follows: "User needs: [Two-child family, mortgage balance of 1.5 million, annual income of 300,000, focusing on children's education funds and adults' critical illness protection], please select 3 suitable products from the insurance product library, compare annual premiums, coverage, and certainty of returns, and explain the reasons for the recommendation." In this embodiment, LoRA (Low-Rank Adaptation) technology is used to fine-tune only 0.1% of the model parameters, thereby reducing computational power consumption and injecting insurance expertise while retaining the general capabilities of the base model.

[0046] In this embodiment of the application, high-quality answers are used as reward signals. The output of the large language model is optimized through the PPO algorithm. For example, for claims consultation, the reward function is designed as follows: compliance accounts for 40%, comprehensibility accounts for 30%, and completeness accounts for 30%. Here, compliance is used to indicate whether the correct terms are cited, comprehensibility is used to indicate whether "deductible" and "waiting period" are explained in colloquial language, and completeness is used to indicate whether implicit concerns of users are covered, such as "what if the materials are incomplete".

[0047] In one application scenario, compliance can be indicated by different violation categories, such as serious, moderate, and minor violations. For example, if explicit absolute terms such as "guaranteed returns," "sure-fire profits," or "100% compensation" are used, it is judged as a serious violation, and the reward for the entire answer is directly set to -100, i.e., 100 points are deducted, resulting in a complete rejection. If necessary risk warnings are omitted, such as "This material is for reference only and does not constitute an offer," or if the wrong clause number is cited, it is judged as a moderate violation, and 20 points are deducted for each instance. When vague expressions are used, such as "generally compensation is available" instead of "compensation is available if it meets the contractual agreement," it is judged as a minor violation, and 5 points are deducted for each instance.

[0048] In one application scenario, comprehensibility is not simply about shortening long sentences, but should consider factors such as terminology explanatory power, analogy usage, and sentence complexity. Terminology explanatory power can include: if technical terms such as "deductible," "waiting period," or "cash value" appear in the answer, they must be followed by colloquial explanations. For example, "The waiting period refers to a period from the effective date of the contract during which the insurance company will not be liable for compensation even if an insured event occurs; it is generally 90 or 180 days." Analogy usage can include: encouraging the use of everyday analogies, such as, "The deductible is like the deductible for medical insurance; expenses below this amount must be borne by the individual." Sentence complexity can include: avoiding complex sentences exceeding 30 characters; if a sentence contains more than two clauses, it needs to be broken down.

[0049] In one application scenario, completeness requires the large language model to not only answer user questions but also cover issues the user might care about but hasn't explicitly asked. This includes coverage of explicit questions, implicit concerns, and action guidelines. Explicit question coverage could include whether all three questions asked by the user have been answered. Implicit concern coverage could include: if a user asks, "What materials are needed for claims?" A good answer should not only list the materials but also proactively add: "If the materials are incomplete, you can submit the existing materials first, and we will inform you what you need to supplement." "Generally, the review will be completed within 3-5 business days after all materials are complete." Action guidelines could include: whether the response concludes with clear suggestions for the next steps, such as "You can call the customer service hotline XXX, or log in to the official website to submit a claim application." In this embodiment, high-quality answers are used as reward signals. The output of the large language model is optimized through the PPO algorithm, thereby improving the accuracy of the large language model's answers and reducing the occurrence rate of compliance risk statements. The risk statement may be "guaranteed returns".

[0050] This application employs a fine-tuning strategy of pre-training, instruction fine-tuning, and reinforcement learning to construct a vertical domain model that understands insurance and is compliant, thereby solving the technical problems of lack of general LLM insurance knowledge and high compliance risks.

[0051] In step 503 of some embodiments, historical interaction data is labeled to obtain labeled interaction data, which may include, but is not limited to: Intent recognition is performed based on historical interaction data to obtain historical intents; Emotion recognition is performed based on historical interaction data to obtain historical emotions; Problems are identified based on historical interaction data, resulting in unresolved issues. Annotated interaction data is obtained by labeling historical interaction data based on historical intent, historical sentiment, and unresolved issues.

[0052] In some embodiments, intent recognition is performed based on historical interaction data. The historical intents obtained may include, but are not limited to, inquiries, insurance applications, claims, complaints, and policy cancellations. Emotion recognition is also performed based on historical interaction data. The historical emotions obtained may include, but are not limited to, calmness, anxiety, anger, and hesitation. Problem identification is also performed based on historical interaction data. Unresolved issues, such as those ultimately transferred to human agents or conversations where the user was dissatisfied, are marked as failed cases. Furthermore, specific unresolved issues can be labeled to facilitate subsequent follow-up. Thus, historical interaction data is labeled based on historical intent, historical emotion, and unresolved issues, resulting in labeled interaction data that includes intent labels, emotion labels, and unresolved issue labels.

[0053] In step 503 of some embodiments, intent recognition, emotion recognition, and question recognition can be performed on historical interaction data based on a large language model, and the data can be recorded and labeled to obtain labeled interaction data.

[0054] In some embodiments, the insurance knowledge base includes an insurance product database, a professional knowledge database, and an industry case database. Prior to step 205 in some embodiments, the business solution generation method based on a large language model further includes constructing the insurance knowledge base, specifically: constructing the insurance product database, professional knowledge database, and industry case database based on professional knowledge data, historical interaction data, and industry case data from multi-source cleaned data. The professional knowledge database may include, for example, insurance domain-related clauses and explanations, such as insurance contract terms, regulatory provisions, product manuals, etc.; the insurance product database includes explanations of relevant insurance products; and the industry case database may include, for example, cases of compensation for abnormal medical examinations during the waiting period, and cases of sales misrepresentation, etc.

[0055] In step 205 of some embodiments, highly matching candidate products are selected from the insurance product library in the insurance knowledge base based on the multidimensional demand graph. For example, the insurance product library contains 200 insurance products and is classified into protection type, wealth management type and comprehensive type. Based on the current multidimensional demand graph, three highly matching candidate products are selected from the insurance product library.

[0056] In some embodiments, the business proposal obtained through step 206 is a three-in-one proposal containing comparative analysis, selection rationale, and risk warnings, enabling personalized proposal generation from a product list to a customized report. At least two candidate products are included, and each candidate product has product attribute information in at least two attribute dimensions. Please refer to [link / reference]. Figure 6 In some embodiments, step 206 may include, but is not limited to: Step 601: For each attribute dimension, perform comparative analysis on the product attribute information of the candidate products based on the large language model to obtain multi-dimensional comparative information between each candidate product. Step 602: Based on the large language model, evaluate the candidate products to obtain the description information of the recommendation reasons that match the demand category; Step 603: Conduct a risk assessment on the candidate products based on the large language model to obtain risk assessment information; Step 604: Generate a business proposal based on multi-dimensional comparison information, recommendation reason description information, and risk assessment information.

[0057] The business proposal obtained in step 206 of this application embodiment is a three-in-one proposal with comparative analysis, selection reasons, and risk warnings, realizing personalized proposal generation from product list to customized report, and achieving transparent recommendation.

[0058] In step 601 of some embodiments, the product attribute information may have attribute dimensions including product identifier, product price, product coverage, exclusion clauses, and benefit information. Product attribute information in the product identifier dimension may include the product name; product attribute information in the product price dimension may include the annual premium; product attribute information in the product coverage dimension may include coverage information; product attribute information in the exclusion clauses dimension may include the content of the exclusion clauses; and product attribute information in the benefit information dimension may include benefit demonstration attributes. By comparing and analyzing candidate products through a large language model, the multi-dimensional comparison information matching all candidate products can be a multi-dimensional comparison table. For the "protection needs of families with two children," in one application scenario, the multi-dimensional comparison information may include, but is not limited to: product name, annual premium, coverage, differences in exclusion clauses, and benefit demonstration. Among them, coverage may be, for example, term life insurance with a sum assured of 2 million, and critical illness insurance including additional compensation for special childhood illnesses; differences in exclusion clauses may be, for example, product A does not cover "sudden death," while product B does; and in the benefit demonstration, if it is an investment product, it distinguishes between guaranteed returns and floating dividends.

[0059] In steps 602 to 603 of some embodiments, candidate products are evaluated and recommended based on a large language model to obtain recommendation reason description information, which may include specific reasons for selection; risk assessment is performed on candidate products based on the large language model to obtain risk assessment information, which may include scenario-based explanations and risk warnings. Specifically, in an application scenario, the large language model writes recommendation logic from the user's perspective, i.e., reasons for selection, for example: Plan A requires an annual payment of 50,000 and will receive 1 million after 20 years, suitable for your 'steady growth' education fund needs, as it includes a guaranteed minimum interest rate of 2.5%. Risk assessment information, for example: Plan B has a dividend function, and the long-term returns may be higher, but it requires bearing the risk of market fluctuations. Plan A's 'premium waiver' clause (waiver of remaining premiums after diagnosis of critical illness) is more suitable for your conservative mindset of 'worrying about unemployment affecting premium payments'.

[0060] In step 604 of some embodiments, a three-in-one business proposal with comparative analysis, selection reasons, and risk warnings is generated based on multi-dimensional comparison information, recommendation reason description information, and risk assessment information to achieve transparent recommendation.

[0061] The business solution generation method based on a large language model provided in this application can be implemented through a three-layer architecture, including an interaction layer, an LLM core layer, and a knowledge base layer. The interaction layer can include APP, mini-program, H5 entry points, etc.; the LLM core layer is a fine-tuned domain-specific large model; and the knowledge base layer can include an insurance product database, a terms database, a user profile database, etc. The interaction layer collects natural language input from users, including text and speech, and converts speech to text. The LLM core layer includes requirement analysis, graph construction, and solution generation. Requirement analysis includes information extraction and requirement ranking; graph construction refers to building a multi-dimensional requirement graph; and solution generation includes selecting candidate products and generating solutions based on these candidates. The knowledge base layer provides real-time data support for the LLM. The interaction layer, LLM core layer, and knowledge base layer are linked through APIs to ensure real-time response and accuracy.

[0062] Please see Figure 7 After step 206 in some embodiments, the business solution generation method based on a large language model may further include, but is not limited to, updating the business solution document, specifically including: Step 701: Classify the current user intent based on the large language model to obtain the business requirement category; Step 702: Query the preset database based on the business requirement category to obtain evidence chain data; Step 703: Obtain current requirement information from the dialogue context based on the large language model; Step 704: Render the proposal based on the evidence chain data and current requirements information.

[0063] In steps 701 to 702 of some embodiments, the current user intent is classified using a large language model to obtain business demand categories. For example, if the current user intent involves the name of an insurance product and the dividend realization rate, the business demand categories obtained by intent classification include data query and regulatory compliance. Therefore, in step 702, the large language model performs a data query from a preset database based on the business demand category, calling upon historical dividend reports and regulatory filing documents from the knowledge base to generate an evidence chain, thereby obtaining the evidence chain data.

[0064] In steps 703 to 704 of some embodiments, the latest current demand information is obtained from the dialogue context. This ensures that each time a user modifies their demand (e.g., "annual premium reduced to 40,000"), a full recalculation is not possible; instead, an incremental update based on the previous session state is required. Otherwise, the user experience will be disrupted. In this embodiment, the incremental calculation only recalculates the affected parts, such as the total premium, first-year commission, and the first few columns of the cash value table related to the latest current demand information. Other unchanged parts (such as the description of coverage and exclusions) are directly read from the cache. This embodiment uses a template engine to update the business plan. In practical applications, the template is split into multiple independent blocks. When user demands change, only the blocks corresponding to the changed parts are re-rendered, achieving a partial page refresh without reloading the entire document.

[0065] This application embodiment obtains the latest current demand information in the dialogue context through interactive follow-up questions in order to update and optimize the business plan. Specifically, it supports users to ask detailed questions about the plan to obtain current demand information, such as "How is the dividend realization rate guaranteed?" and "What are the conditions for premium waiver?". The large language model calls "product historical dividend reports" and "regulatory filing documents" in the knowledge base to generate evidence chains, such as: the dividend realization rate of this product has been ≥90% for the past 3 years, and the data can be found on the official website. It also allows users to adjust parameters, such as "reduce the annual premium to 40,000", and regenerate the plan to update the business plan, thereby realizing an immersive experience of "modifying while chatting".

[0066] The business solution generation method based on a large language model provided in this application responds to user business requests, acquires user dialogue information based on natural language, extracts information from the user dialogue information based on a pre-trained large language model to obtain user demand entities and entity relationships, and sorts user demand entities based on risk prediction data to obtain demand priorities. Based on the demand priorities, the method determines the level of nodes in a multi-dimensional demand graph, thereby selecting candidate products matching the demand categories from an insurance knowledge base based on the multi-dimensional demand graph, and generating a three-in-one business solution document that meets user needs. This upgrades from traditional static tag matching to in-depth mining and demand sorting of implicit user needs, realizing a transformation from single product push to a transparent, three-in-one customized solution document, thus improving the efficiency and accuracy of solution document generation. This application embodiment can be applied to insurance intelligent customer service and personalized product recommendation systems. The innovations of this application embodiment include constructing a dual-drive architecture of domain LLM and dynamic multi-dimensional demand graph, and constructing a three-layer integrated architecture of domain LLM core, dynamic multi-dimensional demand graph, and personalized business solution document generation. This application's embodiments utilize pre-training, instruction fine-tuning, and RLHF strategies to create vertical domain models that understand insurance and comply with regulations, solving the problems of insufficient knowledge and high compliance risks in general large-scale models. Secondly, it leverages LLM's contextual understanding capabilities to deeply analyze users' implicit needs, thereby constructing a dynamic multi-dimensional demand graph. This breaks through the limitations of traditional tag recommendation, upgrading from traditional static tag matching to in-depth mining and automatic ranking of users' implicit needs (such as family structure, debt pressure, and occupational risks). Finally, it realizes the transformation from single product push to transparent customized business solutions with comparative analysis, scenario-based explanation, and interactive optimization.

[0067] In the property insurance business, this application will: for users, shorten consultation time, improve the accuracy of understanding complex needs, and increase the transparency of the plan, thereby improving the efficiency of insurance decision-making; for insurance companies, it can significantly reduce the rate of customer service being transferred to human agents, the conversion rate of personalized product recommendations is effectively improved compared with traditional tag recommendations, users' recognition of the "clarity of plan explanation" can also be improved, and the complaint rate caused by sales misrepresentation can be reduced, which will help to achieve compliant operation and build brand trust.

[0068] This application's embodiments utilize the contextual understanding, long text generation, and logical reasoning capabilities of a large language model to construct a full-process intelligent service system encompassing in-depth demand mining, precise product matching, and transparent solution explanation. This system not only bridges the semantic gap of traditional customer service but also enables personalized recommendations tailored to each individual, driving the transformation of insurance services from "passive response" to "proactive understanding," thereby improving user satisfaction and enhancing product competitiveness.

[0069] Please see Figure 8This application also provides a business solution generation apparatus based on a large language model, which can implement the above-mentioned business solution generation method based on a large language model. The apparatus includes: The dialogue acquisition module is used to obtain user dialogue information in response to user business requests; the user dialogue information is natural language information. The information extraction module is used to extract information from user dialogue information based on a pre-trained large language model, and obtain the user's demand entities and entity relationships. The demand sorting module is used to sort user demand entities based on risk prediction data to obtain demand priorities. The demand graph construction module is used to construct a multi-dimensional demand graph based on user demand entities, entity relationships, and demand priorities. User demand entities serve as graph nodes in the multi-dimensional demand graph, entity relationships serve as edges in the multi-dimensional demand graph, and the level of the graph nodes is determined by the demand priorities. The product screening module is used to filter candidate products that match the demand categories from a pre-built insurance knowledge base based on a multi-dimensional demand graph. The proposal generation module is used to generate proposals for candidate products based on a large language model, resulting in a business proposal.

[0070] In some embodiments, the demand sorting module can specifically be used to implement: Acquire user profile data and extract features from it to obtain user profile features; Risk prediction is performed based on user demand entities and user profile characteristics to obtain risk types and loss levels; Demands are prioritized based on risk type and degree of loss.

[0071] Specifically, the demand sorting module can be used to implement steps 301 to 303 above, which will not be described in detail here.

[0072] In some embodiments, the demand graph construction module can be specifically used to implement: Obtain dialogue context from multi-turn conversations; Perform intent recognition within the dialogue context to obtain the current user intent; The node weights are adjusted based on the current user intent to obtain the current weights; The multidimensional demand graph is updated based on the current weights.

[0073] Specifically, the demand graph construction module can be used to implement steps 401 to 404 above, which will not be elaborated here.

[0074] In some embodiments, the apparatus further includes a model optimization module, which can be specifically used to implement: Obtain multi-source corpora; The multi-source corpus is cleaned to obtain multi-source cleaned data; the multi-source cleaned data includes historical interaction data. Candidate interaction data are filtered from historical interaction data; The candidate interaction data is rewarded based on a preset reward function to obtain an interaction score; Optimize large language models based on interactive scoring.

[0075] Specifically, the model optimization module can be used to implement steps 501 to 505 above, which will not be described in detail here.

[0076] In some embodiments, the proposal generation module can specifically be used to implement: For each attribute dimension, the product attribute information of the candidate products is compared and analyzed based on the large language model to obtain multi-dimensional comparison information between each candidate product. Based on a large language model, candidate products are evaluated and recommended to obtain recommendation reason descriptions that match the demand category; Risk assessment information is obtained by performing risk assessment on candidate products based on a large language model. A business proposal is generated based on multi-dimensional comparative information, descriptions of recommendation reasons, and risk assessment information.

[0077] Specifically, the proposal generation module can be used to implement steps 601 to 604 above, which will not be described in detail here.

[0078] In some embodiments, the proposal generation module can also be used to implement: Data is queried from a pre-defined database based on business requirement categories to obtain evidence chain data; Based on a large language model, current requirement information is obtained from the dialogue context; The proposal is rendered based on evidence chain data and current requirements information.

[0079] Specifically, the proposal generation module can be used to implement steps 701 to 704 above, which will not be described in detail here.

[0080] The specific implementation of the business solution generation device based on the large language model is basically the same as the specific implementation of the business solution generation method based on the large language model described above, and will not be repeated here.

[0081] This application also provides a computer device, which can be any smart terminal including tablet computers, in-vehicle computers, etc. The computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement: In response to user business requests, obtain user dialogue information; user dialogue information is natural language information; Based on a pre-trained large language model, information is extracted from user dialogue information to obtain user demand entities and entity relationships. Based on risk prediction data, user demand entities are sorted to obtain demand priority; A multidimensional demand graph is constructed based on user demand entities, entity relationships, and demand priorities. User demand entities serve as nodes in the multidimensional demand graph, entity relationships serve as edges, and the level of a node is determined by the demand priority. Based on a multi-dimensional demand graph, candidate products matching the demand categories are selected from a pre-built insurance knowledge base; Based on a large language model, solutions are generated for candidate products to obtain business proposals.

[0082] Please see Figure 9 , Figure 9 The hardware structure of a computer device according to another embodiment is illustrated. The computer device includes: The processor 901 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 902 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 902 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 902 and is called and executed by the processor 901 using the service solution generation method based on a large language model according to the embodiments of this application. The input / output interface 903 is used to implement information input and output; The communication interface 904 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 905 transmits information between various components of the device (e.g., processor 901, memory 902, input / output interface 903, and communication interface 904); The processor 901, memory 902, input / output interface 903, and communication interface 904 are connected to each other within the device via bus 905.

[0083] This application embodiment also provides a storage medium, which is a computer-readable storage medium, storing a computer program that is implemented when executed by a processor. In response to user business requests, obtain user dialogue information; user dialogue information is natural language information; Based on a pre-trained large language model, information is extracted from user dialogue information to obtain user demand entities and entity relationships. Based on risk prediction data, user demand entities are sorted to obtain demand priority; A multidimensional demand graph is constructed based on user demand entities, entity relationships, and demand priorities. User demand entities serve as nodes in the multidimensional demand graph, entity relationships serve as edges, and the level of a node is determined by the demand priority. Based on a multi-dimensional demand graph, candidate products matching the demand categories are selected from a pre-built insurance knowledge base; Based on a large language model, solutions are generated for candidate products to obtain business proposals.

[0084] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0085] The business solution generation method, apparatus, device, and medium based on a large language model provided in this application respond to user business requests, acquire user dialogue information based on natural language, extract information from the user dialogue information based on a pre-trained large language model to obtain user demand entities and entity relationships, and sort user demand entities based on risk prediction data to obtain demand priorities. Based on the demand priorities, the level of nodes in a multi-dimensional demand graph is determined, thereby selecting candidate products that match the demand categories from the insurance knowledge base based on the multi-dimensional demand graph, and generating a three-in-one business solution that meets user needs. This upgrades from traditional static tag matching to in-depth mining and demand sorting of implicit user needs, realizing the transformation from single product push to a transparent three-in-one customized solution, thereby improving the efficiency and accuracy of solution generation. This application's embodiments construct a dual-drive architecture of domain-specific LLM and dynamic multidimensional demand graph. By deeply analyzing users' implicit needs through LLM, it breaks through the limitations of traditional tag recommendation. Furthermore, through instruction fine-tuning and RLHF insurance-vertical LLM training methods, it balances professionalism and compliance, addressing the risk of general models "answering questions irrelevantly and misleading sales." It also constructs a personalized solution generation system of "comparison table + scenario-based explanation + interaction optimization," realizing a service upgrade from "product push" to "demand co-creation."

[0086] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0087] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0088] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0089] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0090] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0091] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0092] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0093] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0094] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0095] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0096] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0097] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A business solution generation method based on a large language model, characterized in that, The method includes: In response to a user's business request, obtain user dialogue information; the user dialogue information is natural language information. Information is extracted from the user dialogue information based on a pre-trained large language model to obtain user demand entities and entity relationships; wherein, the user demand entities are used to represent demand categories. Based on risk prediction data, the user demand entities are sorted to obtain demand priorities; A multidimensional demand graph is constructed based on the user demand entity, the entity association relationship, and the demand priority; wherein, the user demand entity serves as the graph node of the multidimensional demand graph, the entity association relationship serves as the edge of the multidimensional demand graph, and the level of the graph node is determined by the demand priority. Based on the multidimensional demand graph, candidate products matching the demand categories are selected from a pre-built insurance knowledge base; Based on the large language model, a solution is generated for the candidate products to obtain an insurance solution and a business solution containing the candidate products.

2. The method according to claim 1, characterized in that, The candidate products include at least two, and each candidate product has product attribute information in at least two attribute dimensions. The step of generating a business plan based on the large language model for the candidate products includes: For each attribute dimension, the product attribute information of the candidate products is compared and analyzed based on the large language model to obtain multi-dimensional comparison information between each candidate product. Based on the large language model, the candidate products are evaluated and recommended to obtain recommendation reason description information that matches the demand category; Based on the large language model, a risk assessment is performed on the candidate products to obtain risk assessment information; The business plan is generated based on the multidimensional comparison information, the description of the recommendation reasons, and the risk assessment information.

3. The method according to claim 1, characterized in that, The multidimensional demand graph has graph nodes and node weights. After constructing the multidimensional demand graph based on the user demand entities, the entity relationships, and the demand priorities, the method includes updating the multidimensional demand graph, specifically including: Obtain dialogue context from multi-turn conversations; Intent recognition is performed on the dialogue context to obtain the current user intent; The node weights are adjusted based on the current user intent to obtain the current weights; The multidimensional demand graph is updated based on the current weights.

4. The method according to claim 3, characterized in that, The method also includes updating the business plan, specifically including: Based on the large language model, the current user intent is classified to obtain the business requirement category; Based on the business requirement category, a data query is performed on the preset database to obtain evidence chain data; Based on the large language model, the current requirement information is obtained from the dialogue context; The proposal is rendered based on the evidence chain data and the current requirements information.

5. The method according to claim 1, characterized in that, The risk prediction data includes risk type and loss level. The process of prioritizing user needs based on the user need entity to obtain need priorities includes: Acquire user profile data and extract features from the user profile data to obtain user profile features; Based on the user demand entity and the user profile characteristics, risk prediction is performed to obtain the risk type and the degree of loss. The demands are prioritized based on the risk type and the degree of loss.

6. The method according to any one of claims 1 to 5, characterized in that, The method also includes optimizing the large language model, specifically including: Obtain multi-source corpora; The multi-source corpus is cleaned to obtain multi-source cleaned data; wherein, the multi-source cleaned data includes historical interaction data; The historical interaction data is labeled to obtain labeled interaction data, and candidate interaction data is filtered out from the labeled interaction data. The candidate interaction data is used to calculate the reward based on a preset reward function to obtain an interaction score; The large language model is optimized based on the interactive scoring.

7. The method according to claim 6, characterized in that, The step of annotating the historical interaction data to obtain annotated interaction data includes: Based on the historical interaction data, intent recognition is performed to obtain the historical intent; Emotion recognition is performed based on the historical interaction data to obtain historical emotions; Based on the historical interaction data, problems are identified to determine unresolved issues. The historical interaction data is labeled based on the historical intent, the historical sentiment, and the unresolved issues to obtain the labeled interaction data.

8. A business solution generation device based on a large language model, characterized in that, The device includes: The dialogue acquisition module is used to acquire user dialogue information in response to user business requests; the user dialogue information is natural language information. The information extraction module is used to extract information from the user dialogue information based on a pre-trained large language model to obtain user demand entities and entity associations; wherein, the user demand entities are used to represent demand categories. The demand sorting module is used to sort the user demand entities based on risk prediction data to obtain demand priorities. The demand graph construction module is used to construct a multi-dimensional demand graph based on the user demand entity, the entity relationship, and the demand priority; wherein, the user demand entity is the graph node of the multi-dimensional demand graph, the entity relationship is the edge of the multi-dimensional demand graph, and the level of the graph node is determined by the demand priority. The product screening module is used to select candidate products that match the demand category from a pre-built insurance knowledge base based on the multi-dimensional demand graph. The proposal generation module is used to generate proposals for the candidate products based on the large language model, thereby obtaining business proposals.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images