Intelligent decision-making method and device based on intelligent agent and large model, equipment and medium

Abstract

The invention provides an intelligent decision-making method and device based on an intelligent agent and a large model, equipment and a medium, relates to the technical field of artificial intelligence, and is suitable for the field of finance. The method comprises the following steps: determining a target enterprise from at least two enterprises, and obtaining multi-source enterprise information of the target enterprise; performing feature embedding on the multi-source enterprise information by using the multi-modal representation learning model to obtain enterprise multi-modal features; performing portrait construction according to the enterprise multi-modal features of the at least two enterprises to obtain an enterprise portrait; performing semantic matching on the enterprise portrait and the product knowledge graph through a double-tower semantic matching model to obtain a target recommended product; reasoning the enterprise portrait and the target recommended product through a reasoning large model to obtain a product recommendation reason; and pushing the target recommended product and the product recommendation reason to the target enterprise through the interactive intelligent agent, and pushing the target answer to the target enterprise based on the received feedback problem of the target enterprise. The accuracy of intelligent decision making is improved.

Description

Intelligent decision-making methods, devices, equipment, and media based on intelligent agents and large models Technical Field

[0001] This application relates to the field of artificial intelligence technology, and is applicable to the financial field, particularly to an intelligent decision-making method, device, equipment, and medium based on intelligent agents and large models. Background Technology

[0002] In the field of artificial intelligence, pre-trained large models or intelligent agents can recommend personalized products to businesses. For example, property insurance and liability insurance can be recommended to businesses.

[0003] Currently, insurance companies and other service companies have the following shortcomings when serving small and micro enterprises: (1) Low data utilization: The data of small and micro enterprises are mostly in unstructured form (business pictures, contracts, invoices, audio of conversations, etc.), making it difficult to extract effective information to support intelligent decision-making (including product recommendations, Q&A, etc.). (2) Complex products and low accuracy of matching methods: Current products are designed for standardized scenarios and cannot flexibly adapt to the diversity of small and micro enterprises. Summary of the Invention

[0004] The main objective of this application is to propose intelligent decision-making methods, devices, equipment, and media based on intelligent agents and large models, which can solve the technical problems of low data utilization and low accuracy of matching methods, and improve the accuracy of intelligent decision-making.

[0005] To achieve the above objectives, a first aspect of this application proposes an intelligent decision-making method based on intelligent agents and large models. The method includes: identifying a target enterprise from at least two enterprises and obtaining multi-source enterprise information for the target enterprise; wherein the multi-source enterprise information belongs to at least two of the following modalities: text modality, image modality, structured data modality, and behavioral data modality; embedding features into the multi-source enterprise information using a multi-modal representation learning model to obtain enterprise multi-modal features; constructing a profile based on the enterprise multi-modal features of the at least two enterprises to obtain an enterprise profile of the target enterprise; performing semantic matching between the enterprise profile and a preset product knowledge graph using a dual-tower semantic matching model to obtain a target recommended product; reasoning about the enterprise profile and the target recommended product using a large-scale reasoning model to obtain product recommendation reasons; pushing the target recommended product and the product recommendation reasons to the target enterprise through an interactive intelligent agent, and pushing target answers to the target enterprise based on received feedback questions from the target enterprise.

[0006] Optionally, the step of constructing a profile of the target enterprise based on the multimodal features of at least two enterprises to obtain the enterprise profile of the target enterprise includes: clustering based on the multimodal features of at least two enterprises to obtain the cluster features of the target enterprise; generating multi-labels based on the cluster features and the multimodal features of the enterprises to obtain the enterprise profile; wherein the enterprise profile includes at least two of the following labels: enterprise type label, enterprise risk label, and enterprise potential demand label.

[0007] Optionally, the step of generating the enterprise profile by multi-labeling based on the clustering features and the enterprise multimodal features includes: performing multi-label classification on the clustering features and the enterprise multimodal features using a multi-label classifier to obtain the enterprise profile; the method further includes: obtaining the sample enterprise type, sample enterprise risk records, and product purchase records of the sample enterprises; obtaining the sample clustering features and sample enterprise multimodal features of the sample enterprises; performing multi-label classification on the sample clustering features and sample enterprise multimodal features using an initial classifier to obtain predicted type labels, predicted risk labels, and predicted potential demand labels; performing loss calculation based on the predicted type labels and the sample enterprise type to obtain type classification loss data; performing loss calculation based on the predicted risk labels and the sample enterprise risk records to obtain risk classification loss data; performing loss calculation based on the predicted potential demand labels and the product purchase records to obtain potential demand classification loss data; and adjusting the parameters of the initial classifier based on the type classification loss data, the risk classification loss data, and the potential demand classification loss data to obtain the multi-label classifier.

[0008] Optionally, before embedding features of the multi-source enterprise information using a multimodal representation learning model to obtain enterprise multimodal features, the method further includes: acquiring sample multi-source enterprise information of sample enterprises; extracting sample images belonging to the image modality from the sample multi-source enterprise information, and extracting sample text matching the sample images; performing feature encoding on the sample images using an initial multimodal coding model to obtain image features, and performing feature encoding on the sample text to obtain text features; calculating similarity based on the image features and the text features to obtain image-text alignment loss data; and adjusting the parameters of the initial multimodal coding model based on the image-text alignment loss data to obtain the multimodal representation learning model.

[0009] Optionally, after encoding the sample image using an initial multimodal coding model to obtain image features, the method further includes: calling a large language model to perform image reasoning on the sample image to obtain image description text; encoding the image description text to obtain image-enhanced text features; and updating the image features based on the image-enhanced text features.

[0010] Optionally, after performing semantic matching between the enterprise profile and the preset product knowledge graph using a dual-tower semantic matching model to obtain the target recommended product, the method further includes: in response to receiving an insurance application from the target enterprise for the target recommended product, performing risk prediction on the insurance application using a risk control big data model to obtain underwriting risk information, and performing intelligent underwriting on the insurance application based on the underwriting risk information through an underwriting intelligent agent; in response to receiving a claim application from the target enterprise for the recommended product, performing intelligent risk control decision-making on the claim application based on the underwriting risk information through a risk control intelligent agent to obtain claim risk information, and performing intelligent approval through a decision center based on the claim risk information.

[0011] Optionally, the step of reasoning about the enterprise profile and the target recommended product using a large-scale reasoning model to obtain the product recommendation reason includes: obtaining the semantic matching score of the target recommended product by the dual-tower semantic matching model; obtaining the graph path of the target recommended product in the product knowledge graph; obtaining a summary of the enterprise risk points of the target enterprise; and generating text from the enterprise profile, the semantic matching score, the graph path, and the summary of enterprise risk points using the large-scale reasoning model to obtain the product recommendation reason.

[0012] To achieve the above objectives, a second aspect of this application proposes an intelligent decision-making device based on an intelligent agent and a large model. The device includes: an information acquisition module, configured to determine a target enterprise from at least two enterprises and acquire multi-source enterprise information of the target enterprise; wherein the multi-source enterprise information belongs to at least two of the following modalities: text modality, image modality, structured data modality, and behavioral data modality; a feature embedding module, configured to embed features into the multi-source enterprise information using a multi-modal representation learning model to obtain enterprise multi-modal features; a profile construction module, configured to construct a profile based on the enterprise multi-modal features of at least two enterprises to obtain an enterprise profile of the target enterprise; a semantic matching module, configured to perform semantic matching between the enterprise profile and a preset product knowledge graph using a dual-tower semantic matching model to obtain a target recommended product; a model reasoning module, configured to reason about the enterprise profile and the target recommended product using a large reasoning model to obtain product recommendation reasons; and an interaction module, configured to push the target recommended product and the product recommendation reasons to the target enterprise through an interactive intelligent agent, and push target answers to the target enterprise based on received feedback questions from the target enterprise.

[0013] To achieve the above objectives, a third aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the intelligent decision-making method based on intelligent agents and large models described in the first aspect.

[0014] 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 intelligent decision-making method based on intelligent agents and large models described in the first aspect.

[0015] This application proposes an intelligent decision-making method, device, electronic device, and computer-readable storage medium based on intelligent agents and large models. First, it acquires multi-source enterprise information of the target enterprise. Then, it uses a multimodal representation learning model to embed features into the multi-source enterprise information, obtaining multimodal enterprise features. Next, it constructs a profile based on the multimodal features of at least two enterprises, obtaining an enterprise profile of the target enterprise. Then, it uses a dual-tower semantic matching model to perform semantic matching between the enterprise profile and a pre-set product knowledge graph, obtaining a target recommended product. Then, it uses a large-scale reasoning model to reason about the enterprise profile and the target recommended product, obtaining the product recommendation reason. Finally, it pushes the target recommended product and the product recommendation reason to the target enterprise through an interactive intelligent agent, and pushes the target answer to the target enterprise based on the feedback questions received from the target enterprise. This solves the technical problems of low data utilization and low matching accuracy, improving the accuracy of intelligent decision-making.

[0016] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0017] Figure 1 is a flowchart of an intelligent decision-making method based on intelligent agents and large models provided in an embodiment of this application; Figure 2 is a flowchart of an intelligent decision-making method based on intelligent agents and large models provided in another embodiment of this application; Figure 3 is a flowchart of an intelligent decision-making method based on intelligent agents and large models provided in yet another embodiment of this application; Figure 4 is a flowchart of step 103 in Figure 1; Figure 5 is a flowchart of an intelligent decision-making method based on intelligent agents and large models provided in yet another embodiment of this application; Figure 6 is a flowchart of step 105 in Figure 1; Figure 7 is a block diagram of the module structure of an intelligent decision-making device based on intelligent agents and large models provided in an embodiment of this application; Figure 8 is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0018] 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.

[0019] 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.

[0020] 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.

[0021] First, let's clarify some terms used in this application: Artificial intelligence (AI) is a new technical science that studies and develops theories, methods, technologies, and application systems for simulating, extending, and expanding human intelligence. AI is a branch of computer science that attempts to understand the essence of intelligence and produce a new type of intelligent machine 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. AI also 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.

[0022] Natural Language Processing (NLP): NLP uses computers to process, understand, and utilize human language (such as Chinese and English). It 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, information and image processing, 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.

[0023] An AI Agent is an application that can autonomously plan and invoke external tools to execute tasks based on objectives and external information, and reflect on the results. AI Agents are designed around task objectives and business processes in terms of implementation, operational form, and behavioral characteristics.

[0024] Large Language Models (LLMs) are a class of deep learning models that utilize massive amounts of text data for autoregressive or self-supervised learning. They are capable of generating, understanding, translating, and summarizing complex natural language tasks. Common underlying frameworks are often based on the Transformer architecture, with core elements including attention mechanisms, multi-layer stacking, and positional encoding. Typical capabilities include text generation, question answering, dialogue, summarization, translation, code completion, and sentiment analysis.

[0025] A knowledge graph (KG) is a graph-based method for representing and managing knowledge, where "entities" (nodes) are connected by "relationships" (edges) to form a queryable and reasonable knowledge network. Entities are typically real-world things or concepts (such as people, places, companies, products, events, etc.), while relationships describe the semantic connections between entities (such as "belongs to," "creator," "located in," "same category," etc.). The goal of a knowledge graph is to integrate structured, semi-structured, and unstructured data into a unified, reasonable knowledge resource, supporting semantic queries, reasoning, and the discovery of new relationships.

[0026] Currently, there are various attempts in the industry to combine Large Language Models (LLM) with Knowledge Graphs (KG) to improve the accuracy and interpretability of decision-making systems. However, the service industry (such as the insurance industry) has significant pain points in serving small and medium-sized enterprises (SMEs): (1) Complex products with low matching degree. Traditional products are designed for standardized scenarios and cannot flexibly adapt to the diverse risk characteristics and business models of SMEs. (2) High customer acquisition cost and low conversion efficiency. Sales personnel need to manually screen customers and manually recommend products. Marketing relies on experience and manual judgment, and lacks intelligent recommendation and profile insight capabilities. (3) Fragmented and inefficient underwriting and claims processes. Different links rely on independent systems, information transmission is lagging, and there is a lack of intelligent collaborative mechanisms from sales, underwriting to risk control. (4) Low data utilization rate. The data of SMEs is mostly in unstructured form (business pictures, contracts, invoices, audio of conversations, etc.), and existing systems have difficulty extracting effective information to support intelligent decision-making. Existing technologies mostly focus on a single link (such as intelligent customer service or marketing recommendations), lacking an integrated intelligent assistant architecture across the entire chain, making it difficult to achieve linkage and intelligent evolution of links such as sales, underwriting, and risk control.

[0027] Based on this, embodiments of this application propose an intelligent decision-making method, an intelligent decision-making device, an electronic device, and a computer-readable storage medium based on intelligent agents and large models. By constructing a "full-link intelligent platform," a closed-loop intelligent process is realized from customer insight → product recommendation → dynamic underwriting → risk control decision → service tracking. This application adopts a multi-agent collaborative architecture, with a dedicated large model as the scheduling and inference core, integrating 20+ vertical expert agents to achieve distributed intelligent decision-making for tasks.

[0028] The intelligent decision-making method based on intelligent agents and large models provided in this application can be applied to terminals and servers, or it can be software running on the server. The server can be configured as an independent physical server, or as a server cluster or distributed system composed of multiple physical servers, or it can be configured as a cloud server that provides 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 that implements the intelligent decision-making method based on intelligent agents and large models, etc., but is not limited to the above forms.

[0029] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include server computers, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, 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.

[0030] This application provides an intelligent decision-making method based on intelligent agents and large models, an intelligent decision-making device based on intelligent agents and large models, an electronic device, and a computer-readable storage medium. The specific implementation is described through the following embodiments. First, the intelligent decision-making method based on intelligent agents and large models in this application is described.

[0031] It should be noted that in each specific implementation of this application, when it is necessary to process data related to the user's identity or characteristics, such as the user's corporate information (annual report, revenue data), the user's permission or consent will be obtained first. Moreover, the collection, use and processing of this data will comply with relevant laws, regulations and standards.

[0032] Referring to Figure 1, which is an optional flowchart of an intelligent decision-making method based on intelligent agents and large models provided in an embodiment of this application, it may include, but is not limited to, steps 101 to 106.

[0033] Step 101: Identify the target company from at least two companies and obtain multi-source company information for the target company; Step 102: Embed features into the multi-source company information using a multimodal representation learning model to obtain multimodal features of the company; Step 103: Construct a profile based on the multimodal features of the target company to obtain a profile of the target company; Step 104: Perform semantic matching between the company profile and a pre-defined product knowledge graph using a dual-tower semantic matching model to obtain the target recommended product; Step 105: Infer the company profile and the target recommended product using a large-scale inference model to obtain the product recommendation reason; Step 106: Push the target recommended product and the product recommendation reason to the target company through an interactive intelligent agent, and push the target answer to the target company based on the feedback questions received from the target company.

[0034] Steps 101 to 106, as illustrated in this embodiment, first construct a corporate profile of the target company based on multi-source corporate information from multiple enterprises. Then, combine this with a product knowledge graph to find the target recommended product. Next, the inference model outputs product recommendation reasons based on the corporate profile and the target recommended product. This way, product matching can fully utilize multi-source corporate information and also provides additional reasons for recommending the target recommended product. Then, the target recommended product and product recommendation reasons are pushed to the target company through an interactive intelligent agent, and the target answer is pushed to the target company based on the feedback questions received from the target company. In this way, the technical problems of low data utilization and low matching accuracy can be solved, thereby improving the accuracy of intelligent decision-making.

[0035] For example, in the insurance scenario in the financial sector, based on multi-source corporate information of company A (basic information, transaction records, credit history, invoice images, call texts, social behavior, etc.), it can be determined to push employer's liability insurance B to company A, along with the reasons for pushing the product recommendation. Then, based on the feedback questions received from company A (such as how much is the premium for employer's liability insurance?), the target answer can be pushed to company A.

[0036] In step 101 of some embodiments, a target enterprise is identified from at least two enterprises, and multi-source enterprise information of the target enterprise is obtained. Each of the at least two enterprises can be the target enterprise. Specifically, the enterprise can be a small or micro-sized enterprise. Small and micro-sized enterprises are usually a classification of enterprise size, mainly defined by dimensions such as number of employees, output / revenue, and asset size. For example, micro enterprises typically have 0-9 employees and low revenue / output; small enterprises typically have 10-49 employees and medium to low revenue / output; medium-sized enterprises typically have 50-300 employees and medium to high revenue / output; large enterprises are generally not included in the category of small and medium-sized enterprises.

[0037] Multi-source enterprise information refers to enterprise-related data and information from different sources, integrated or aggregated, used for description, analysis, and decision-making. Multi-source enterprise information belongs to at least two of the following modalities: text modality, image modality, structured data modality, and behavioral data modality. For example, multi-source enterprise information includes: basic information, credit reports, and call transcripts (text modality); invoice images and ID photos (image modality); transaction records (structured data modality); and social behaviors (behavioral data modality). Multi-source enterprise information can be obtained through direct database connections, data warehouse extraction, Enterprise Service Bus (ESB), or data integration tools.

[0038] In step 102 of some embodiments, a multimodal representation learning model is used to embed features of multi-source enterprise information to obtain enterprise multimodal features. A multimodal representation learning model is a model capable of simultaneously learning and generating unified or coordinated representation vectors from multiple modalities (such as text modality, image modality, structured data modality, behavioral data modality, etc.). The core objective of a multimodal representation learning model is to enable the semantic information between different modalities to be aligned and integrated, thereby improving the performance of tasks such as cross-modal retrieval, understanding, and generation. Multimodal representation learning models can be models like CLIP and ALIGN, which establish strong alignment between text and images through large-scale contrastive learning. Enterprise multimodal features include target image features for the image modality, target text features for the text modality, structured data features for the structured data modality, and behavioral data features for the behavioral data modality.

[0039] In one example, a multimodal contrastive learning encoder architecture (MM-CLR) is employed to map all multi-source enterprise information containing the following modalities to a unified shared embedding space Z∈ d The data types include: text modalities (such as contracts, call logs, and invoice OCR), image modalities (invoice images and ID photos), structured data modalities (business indicators, bank statements, and tax invoice records), and behavioral data modalities (access trajectories and call characteristics). The mapping method is as follows: Then modal alignment is performed: .in, Embedding features representing text modalities Embedding features representing image modalities Embedding features representing structured data modalities Data representing text modality, Data representing image modalities, Data representing structured data modalities, , , These represent encoders for text modality, encoders for image modality, and encoders for structured data modality, respectively.

[0040] In one embodiment, referring to FIG2, before step 102, the intelligent decision-making method based on intelligent agents and large models may further include: step 201, obtaining sample multi-source enterprise information of sample enterprises; step 202, extracting sample images belonging to the image modality from the sample multi-source enterprise information, and extracting sample text matching the sample images; step 203, performing feature encoding on the sample images through an initial multimodal coding model to obtain image features, and performing feature encoding on the sample text to obtain text features; step 204, calculating similarity based on image features and text features to obtain image-text alignment loss data; step 205, adjusting the parameters of the initial multimodal coding model based on the image-text alignment loss data to obtain a multimodal representation learning model.

[0041] This embodiment primarily considers the difficulty in training traditional models due to the lack of continuous structured data among small and medium-sized enterprises (SMEs). Therefore, this embodiment improves representation robustness through "cross-modal weak supervision constraints." Specifically, the image-text alignment loss data is shown below: , where sim() represents the similarity function (such as the cosine similarity formula). Embedding features representing text modalities This represents the embedding features of the image modality. In this way, even if the image quality is poor (weak data) such as ticket effects, it can still be stably represented by other modalities.

[0042] The initial multimodal coding model can be CLIP, ALIGN, or other models.

[0043] In one embodiment, referring to FIG3, after step 203, the intelligent decision-making method based on the agent and the big model may further include: step 301, calling the big language model to perform image reasoning on the sample image to obtain image description text; step 302, performing text encoding on the image description text to obtain image-enhanced text features; step 303, updating the image features according to the image-enhanced text features.

[0044] This embodiment implements Synthetic Weak-Data Augmentation, specifically by using a large language model to synthesize image description text (such as ticket image description text), thereby encoding image-enhanced text features, which in turn update image features. This helps improve the accuracy of image-text alignment loss data, thereby improving the model performance of the multimodal representation learning model.

[0045] In step 303, the image-enhanced text may be concatenated with the image features to obtain updated image features.

[0046] In one embodiment, after step 102, the intelligent decision-making method based on intelligent agents and large models may further include: calling a large language model to perform image reasoning on the target image contained in multi-source enterprise information to obtain target image description text; text encoding the target image description text to obtain target image enhanced text features; and updating the target image features contained in the enterprise multimodal features according to the target image enhanced text features. This allows for multimodal generation and filling of missing data (such as missing tax invoices or no transaction records), improving data utilization, thereby increasing the accuracy of profile construction and ultimately improving the accuracy of intelligent decision-making.

[0047] In step 103 of some embodiments, a profile is constructed based on the multimodal features of at least two enterprises to obtain a profile of the target enterprise.

[0048] In one embodiment, referring to FIG4, step 103 may include: step 401, clustering based on the multimodal features of at least two enterprises to obtain the cluster features of the target enterprise; step 402, generating multi-labels based on the cluster features and the multimodal features of the enterprises to obtain an enterprise profile.

[0049] The advantage of the above embodiments is that clustering based on the multimodal characteristics of multiple enterprises can obtain the clustering characteristics of the target enterprise among multiple enterprises, and then generate an enterprise profile. This is equivalent to generating a profile by using common features and personalized features, which improves the accuracy of the generation.

[0050] In step 401, clustering features can indicate the enterprise category of the target enterprise, such as: "trading micro and small enterprises", "construction engineering customers", "technology service enterprises".

[0051] In one example, step 401 may include the following steps: (1) Inputting the enterprise multimodal features for each enterprise: (2) Dimensionality reduction (optional): (3) Clustering algorithm: HDBSCAN (robust handling of "weak data" scenarios) is adopted: .

[0052] Understandably, considering clustering features when generating enterprise profiles allows for the utilization of multi-source enterprise information from other companies, further improving data utilization.

[0053] In step 402, the enterprise profile includes at least two of the following tags: enterprise type tag, enterprise risk tag, and enterprise potential demand tag.

[0054] In one embodiment, step 402 may include: performing multi-label classification on cluster features and enterprise multimodal features using a multi-label classifier to obtain an enterprise profile.

[0055] In one embodiment, referring to Figure 5, before step 402, the intelligent decision-making method based on intelligent agents and large models may further include: step 501, obtaining the sample enterprise type, sample enterprise risk records, and product purchase records of the sample enterprises; step 502, obtaining the sample clustering features and sample enterprise multimodal features of the sample enterprises; step 503, performing multi-label classification on the sample clustering features and sample enterprise multimodal features using an initial classifier to obtain predicted type labels, predicted risk labels, and predicted potential demand labels; step 504, performing loss calculation based on the predicted type labels and sample enterprise types to obtain type classification loss data; step 505, performing loss calculation based on the predicted risk labels and sample enterprise risk records to obtain risk classification loss data; step 506, performing loss calculation based on the predicted potential demand labels and product purchase records to obtain potential demand classification loss data; step 507, adjusting the parameters of the initial classifier based on the type classification loss data, risk classification loss data, and potential demand classification loss data to obtain a multi-label classifier.

[0056] In step 501, the sample enterprise type can be determined based on industry codes and transaction structures, such as "cross-border trade micro and small enterprises," "construction engineering contracting enterprises," and "asset-light technology micro and small enterprises." Sample risk types can be generated from the sample enterprise risk records based on the risk model: "high cash flow volatility," "high probability of abnormal bills," and "moderate operational stability." Potential demand types can be generated from product purchase records based on the demand inferrer (MLP+Attention): "preference for insurance products with high payout cycles," "employee health insurance demand," and "insufficient coverage of corporate property insurance."

[0057] Step 502, obtaining the sample clustering features is similar to obtaining the target company's clustering features, and will not be described again here. Obtaining the sample company's multimodal features is similar to obtaining the target company's multimodal features, and will not be described again here.

[0058] In step 503, the multi-label classification process is represented as follows: ,in, This indicates the forecast type label, forecast risk label, and forecast potential demand label. Z represents the initial classifier, Z represents the multimodal data of the sample enterprises, and C represents the clustering features of the samples. This indicates that Z and C are concatenated.

[0059] In steps 504 to 506, the loss can be calculated using the multi-label cross-entropy loss function.

[0060] Finally, in step 507, the parameters of the initial classifier can be adjusted based on the fused loss data of type classification loss data, risk classification loss data, and potential demand classification loss data to obtain a multi-label classifier.

[0061] The advantage of the above embodiments is that a multi-label classifier can be trained based on multiple dimensions, which improves the performance of the multi-label classifier and thus improves the accuracy of intelligent decision-making.

[0062] In step 104 of some embodiments, the enterprise profile and the preset product knowledge graph are semantically matched using a dual-tower semantic matching model to obtain the target recommended product.

[0063] The process of constructing a product knowledge graph may include: (1) Structured parsing of clause texts: extracting the following content using an insurance clause extraction model: coverage liability, exclusion clauses, customer suitability, and premium calculation rules. (2) Embedding of regulatory rules: extracting the following content from the regulations of the China Banking and Insurance Regulatory Commission: prohibition conditions, access requirements, sales constraints, and constructing the relationship between regulatory nodes and clause nodes. (3) Extraction of industry-adaptive information: industry accident rate, capital chain risk, and business model characteristics. (4) Knowledge graph construction: Where: V = {clause node, risk node, regulatory node, industry node}, E = {guarantee relationship, applicability relationship, risk propagation relationship}.

[0064] The dual-tower semantic matching model is used to match products to businesses. The input to the dual-tower semantic matching model is the business tower. Input enterprise profile features to obtain enterprise semantic features. Product Tower Input the structured features of the insurance product to obtain its semantic features. Output of the dual-tower semantic matching model: semantic matching score .in, This represents the cosine similarity formula. The product with the highest semantic matching score can be used as the target recommended product.

[0065] In step 105 of some embodiments, the enterprise profile and target recommended products are inferred through the reasoning big model to obtain the product recommendation reasons.

[0066] In one embodiment, referring to FIG6, step 105 may include: step 601, obtaining the semantic matching score of the target recommended product by the dual-tower semantic matching model; step 602, obtaining the graph path of the target recommended product in the product knowledge graph; step 603, obtaining the enterprise risk point summary of the target enterprise; step 604, generating text from the enterprise profile, semantic matching score, graph path and enterprise risk point summary by the reasoning big model to obtain the product recommendation reason.

[0067] In one example, the inputs to the inference model include: enterprise profile, semantic matching score, graph path, and enterprise risk point summary; the outputs of the inference model include: product recommendation reasons (explanatory text), risk disclosure statement, and alternative product comparison report. An example of a product recommendation reason is as follows: Since your company is an "equipment leasing company" and its cash flow has been relatively stable in the past six months, we recommend "corporate property insurance + employer's liability insurance". The recommendation is based on the following criteria: (1) the industry average accident rate is relatively high; (2) the coverage covers your company's core assets; and (3) the premium matches the risk level.

[0068] In step 106 of some embodiments, an interactive agent pushes target recommended products and reasons for product recommendation to the target enterprise, and pushes target answers to the target enterprise based on the feedback questions received from the target enterprise. Specifically, the interactive agent has multi-turn dialogue capabilities, simulates "consultant sales" through role-based prompts, and supports real-time Q&A and solution comparison. Through semantic generation driven by a large model, "dynamic recommendation + explainable marketing" is achieved, significantly improving user trust and conversion rates, and enhancing the accuracy of intelligent decision-making.

[0069] In one embodiment, after step 104 or step 106, the intelligent decision-making method based on the intelligent agent and the big model may further include: in response to receiving an insurance application from the target enterprise for the target recommended product, performing risk prediction on the insurance application through the risk control big model to obtain underwriting risk information, and performing intelligent underwriting on the insurance application based on the underwriting risk information through the underwriting intelligent agent.

[0070] Specifically, the risk control big data model is a customized model fine-tuned for risk assessment scenarios of SMEs. It takes as input information such as enterprise profiles, historical claims records, public opinion signals, and financial anomalies, and outputs risk distribution and suggested premiums. The dynamic underwriting rule engine supports flexible underwriting logic configuration based on rule graphs, automatically triggering the collection of supplementary materials or manual review. The intelligent underwriting agent can autonomously invoke rule judgment, data comparison, and risk level interpretation modules to achieve "intelligent batch underwriting." This embodiment supports human-machine collaborative review, greatly improving approval efficiency and consistency.

[0071] In one example, the fine-tuning process of the risk control model is as follows: (1) Training data comes from: historical underwriting records, claims results, corporate profiles, abnormal financial statements, and public opinion data (negative news, high-risk events). (2) Labels include: risk level (low / medium / high), underwriting conclusion (approved / rejected / requires supplementary materials), and expected premium. (3) Loss function: .in, Indicates the true risk level. This indicates the probability of predicting each risk level. () represents the consistency loss function. This is the weighting coefficient (which can be set according to requirements). This represents the true underwriting conclusion. This indicates the predicted underwriting conclusion. This represents the loss data used to adjust the large-scale risk control model.

[0072] In one example, the underwriting agent uses an underwriting rule graph to implement underwriting in the following steps: (1) Extracting rules from the underwriting manual and terms: constructing rule nodes (Age requirements, abnormal business operations, industry restrictions). (2) Generate edges based on the relationship between rule nodes. To compile the rules into a rule graph (3) Map the enterprise facts to the rule graph during underwriting. (4) Automatically determine the trigger path, such as "financial anomaly → trigger supplementary information".

[0073] In one embodiment, after step 104 or step 106, the intelligent decision-making method based on intelligent agents and large models may further include: in response to receiving a claim application for the recommended product from a target enterprise, making intelligent risk control decisions on the claim application based on underwriting risk information through a risk control intelligent agent to obtain claim risk information, and making intelligent approval based on the claim risk information through a decision center.

[0074] The risk control agent is deployed with a unified risk scoring model: combining the inference results of a large model with structured risk factors (industry volatility, geographical exposure, and claims ratio) to generate a comprehensive risk score. The risk control agent also employs a small model collaboration mechanism: deploying lightweight models for single dimensions (credit, operations, geography, public opinion), which are then uniformly scheduled and weighted by the L2 large model. Furthermore, the risk control agent performs multimodal anomaly detection: utilizing image recognition to detect document forgery and OCR comparison to identify policy anomalies. This embodiment proposes a "combination of large and small models" risk control system, achieving multi-dimensional fusion of risk perception while ensuring computational efficiency.

[0075] In one example, this application designs a full-link multi-agent collaborative controller, including: (1) Agent registration and scheduling mechanism: Each Agent has an independent role and responsibility (such as interactive Agent, underwriting Agent, risk control Agent), is registered through a unified controller and assigned tasks by a large model. (2) Collaborative memory mechanism based on RAG (retrieval-enhanced generation): Agents share context and store interaction history through a vector database (such as Milvus) to achieve cross-task knowledge reuse. (3) Human-machine collaborative interface: When the task complexity exceeds the threshold, the system automatically switches to manual review mode and records the context backtracking chain. In this way, by constructing a "multi-agent + knowledge enhancement + dynamic scheduling" architecture, intelligent closed-loop management from sales to claims is realized.

[0076] In one example, an Insurance Knowledge RAG module is introduced to dynamically retrieve insurance terms, regulatory announcements, and product manuals when generating responses and recommendations, improving the accuracy and compliance of answers. The Insurance Knowledge RAG module specifically includes: vector index construction: segmenting and embedding insurance terms; semantic matching retrieval: retrieving similar terms in real time based on user questions; and contextual concatenation generation: concatenating the search results into the Prompt to enhance the facts.

[0077] In summary, this application can achieve at least the following beneficial effects: (1) Intelligentization and full-link collaboration: realize a closed-loop intelligent process from marketing and customer acquisition to underwriting and risk control, significantly shortening the time required for operations. (2) Precise customer insight and flexible customization: achieve in-depth characterization of the risks and needs of SMEs based on multimodal learning, supporting personalized product recommendations and dynamic premium calculation. (3) Strong interpretability and compliance: through knowledge enhancement and rule graphs, the output of the large model has interpretability and regulatory transparency. (4) Significant cost-effectiveness: through unmanned and batch processing, significantly reduce manpower and communication costs; the model self-learning mechanism continuously optimizes the accuracy of risk control. (5) Innovative architectural advantages: propose a new insurance intelligent system of "multi-agent + large model + knowledge enhancement + risk detection", providing industry-leading digital solutions for SME insurance business.

[0078] Please refer to Figure 7. This application embodiment also provides an intelligent decision-making device based on intelligent agents and large models, which can realize the above-mentioned intelligent decision-making method based on intelligent agents and large models. Figure 7 is a block diagram of the module structure of the intelligent decision-making device based on intelligent agents and large models provided in this application embodiment. The device includes: an information acquisition module 701, used to determine the target enterprise from at least two enterprises and acquire multi-source enterprise information of the target enterprise; a feature embedding module 702, used to embed features of multi-source enterprise information using a multimodal representation learning model to obtain enterprise multimodal features; a profile construction module 703, used to construct a profile based on the enterprise multimodal features of at least two enterprises to obtain the enterprise profile of the target enterprise; a semantic matching module 704, used to perform semantic matching between the enterprise profile and a preset product knowledge graph through a dual-tower semantic matching model to obtain the target recommended product; a model reasoning module 705, used to reason about the enterprise profile and the target recommended product through a reasoning large model to obtain the product recommendation reason; and an interaction module 706, used to push the target recommended product and the product recommendation reason to the target enterprise through an interactive intelligent agent, and push the target answer to the target enterprise based on the feedback questions received from the target enterprise.

[0079] It should be noted that the specific implementation of this intelligent decision-making device based on intelligent agents and large models is basically the same as the specific implementation of the intelligent decision-making method based on intelligent agents and large models described above, and will not be repeated here.

[0080] This application also provides an electronic device, which includes: a memory, a processor, a program stored in the memory and executable on the processor, and a data bus for communication between the processor and the memory. When the program is executed by the processor, it implements the aforementioned intelligent decision-making method based on intelligent agents and large models. This electronic device can be any intelligent terminal, including tablet computers, in-vehicle computers, etc.

[0081] Please refer to Figure 8, which illustrates the hardware structure of an electronic device according to another embodiment. The electronic device includes: a processor 801, which can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, for executing related programs to implement the technical solutions provided in the embodiments of this application; and a memory 802, which can be implemented using a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM), etc. The memory 802 can store the operating system and other applications. 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 802 and is called and executed by the processor 801 to implement the intelligent decision-making method based on intelligent agents and large models in the embodiments of this application. The input / output interface 803 is used to realize information input and output. The communication interface 804 is used to realize communication interaction between this device and other devices. Communication can be realized through wired means (such as USB, network cable, etc.) or through wireless means (such as mobile network, WIFI, Bluetooth, etc.). The bus 805 transmits information between the various components of the device (such as the processor 801, memory 802, input / output interface 803 and communication interface 804). The processor 801, memory 802, input / output interface 803 and communication interface 804 realize communication connection between each other within the device through the bus 805.

[0082] This application also provides a computer-readable storage medium for computer-readable storage, wherein the storage medium stores one or more programs, which can be executed by one or more processors to implement the above-described intelligent decision-making method based on intelligent agents and large models.

[0083] 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.

[0084] 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.

[0085] 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.

[0086] 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.

[0087] 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.

[0088] 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 a 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.

[0089] 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.

[0090] 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 units 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.

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

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

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

[0094] 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. An intelligent decision-making method based on intelligent agents and large models, characterized in that, The method includes: identifying a target enterprise from at least two enterprises and obtaining multi-source enterprise information of the target enterprise; wherein the multi-source enterprise information belongs to at least two of the following modalities: text modality, image modality, structured data modality, and behavioral data modality; embedding features into the multi-source enterprise information using a multi-modal representation learning model to obtain enterprise multi-modal features; constructing a profile based on the enterprise multi-modal features of at least two enterprises to obtain an enterprise profile of the target enterprise; performing semantic matching between the enterprise profile and a preset product knowledge graph using a dual-tower semantic matching model to obtain a target recommended product; reasoning about the enterprise profile and the target recommended product using a large-scale reasoning model to obtain a product recommendation reason; pushing the target recommended product and the product recommendation reason to the target enterprise through an interactive intelligent agent, and pushing the target answer to the target enterprise based on the feedback questions received from the target enterprise.

2. The method according to claim 1, characterized in that, The step of constructing a profile of the target company based on the multimodal features of at least two companies to obtain the company profile of the target company includes: clustering based on the multimodal features of at least two companies to obtain the cluster features of the target company; generating multi-labels based on the cluster features and the multimodal features of the companies to obtain the company profile; wherein the company profile includes at least two of the following labels: company type label, company risk label, and company potential demand label.

3. The method according to claim 2, characterized in that, The step of generating the enterprise profile by multi-labeling based on the clustering features and the enterprise multimodal features includes: classifying the clustering features and the enterprise multimodal features using a multi-label classifier to obtain the enterprise profile; the method further includes: obtaining the sample enterprise type, sample enterprise risk records, and product purchase records of the sample enterprises; obtaining the sample clustering features and sample enterprise multimodal features of the sample enterprises; classifying the sample clustering features and sample enterprise multimodal features using an initial classifier to obtain predicted type labels, predicted risk labels, and predicted potential demand labels; calculating losses based on the predicted type labels and the sample enterprise type to obtain type classification loss data; calculating losses based on the predicted risk labels and the sample enterprise risk records to obtain risk classification loss data; calculating losses based on the predicted potential demand labels and the product purchase records to obtain potential demand classification loss data; and adjusting the parameters of the initial classifier based on the type classification loss data, the risk classification loss data, and the potential demand classification loss data to obtain the multi-label classifier.

4. The method according to any one of claims 1 to 3, characterized in that, Before embedding features of the multi-source enterprise information using a multimodal representation learning model to obtain enterprise multimodal features, the method further includes: acquiring sample multi-source enterprise information of sample enterprises; extracting sample images belonging to the image modality from the sample multi-source enterprise information, and extracting sample text matching the sample images; performing feature encoding on the sample images using an initial multimodal coding model to obtain image features, and performing feature encoding on the sample text to obtain text features; calculating similarity based on the image features and the text features to obtain image-text alignment loss data; and adjusting the parameters of the initial multimodal coding model based on the image-text alignment loss data to obtain the multimodal representation learning model.

5. The method according to claim 4, characterized in that, After encoding the sample image using an initial multimodal coding model to obtain image features, the method further includes: calling a large language model to perform image reasoning on the sample image to obtain image description text; encoding the image description text to obtain image-enhanced text features; and updating the image features based on the image-enhanced text features.

6. The method according to any one of claims 1 to 3, characterized in that, After semantically matching the enterprise profile and the preset product knowledge graph using a dual-tower semantic matching model to obtain the target recommended product, the method further includes: in response to receiving an insurance application from the target enterprise for the target recommended product, performing risk prediction on the insurance application using a risk control big data model to obtain underwriting risk information, and performing intelligent underwriting on the insurance application based on the underwriting risk information through an underwriting intelligent agent; in response to receiving a claim application from the target enterprise for the recommended product, performing intelligent risk control decision-making on the claim application based on the underwriting risk information through a risk control intelligent agent to obtain claim risk information, and performing intelligent approval through a decision center based on the claim risk information.

7. The method according to any one of claims 1 to 3, characterized in that, The step of reasoning about the enterprise profile and the target recommended product using a large-scale reasoning model to obtain the product recommendation reason includes: obtaining the semantic matching score of the target recommended product by the dual-tower semantic matching model; obtaining the graph path of the target recommended product in the product knowledge graph; obtaining a summary of the enterprise risk points of the target enterprise; and generating text from the enterprise profile, the semantic matching score, the graph path, and the summary of enterprise risk points using the large-scale reasoning model to obtain the product recommendation reason.

8. An intelligent decision-making device based on intelligent agents and large models, characterized in that, The device includes: an information acquisition module for identifying a target enterprise from at least two enterprises and acquiring multi-source enterprise information of the target enterprise; wherein the multi-source enterprise information belongs to at least two of the following modalities: text modality, image modality, structured data modality, and behavioral data modality; a feature embedding module for embedding features into the multi-source enterprise information using a multi-modal representation learning model to obtain enterprise multi-modal features; a profile construction module for constructing a profile based on the enterprise multi-modal features of at least two enterprises to obtain an enterprise profile of the target enterprise; a semantic matching module for performing semantic matching between the enterprise profile and a preset product knowledge graph using a dual-tower semantic matching model to obtain a target recommended product; a model reasoning module for reasoning between the enterprise profile and the target recommended product using a large reasoning model to obtain product recommendation reasons; and an interaction module for pushing the target recommended product and the product recommendation reasons to the target enterprise through an interactive intelligent agent, and pushing target answers to the target enterprise based on the feedback questions received from the target enterprise.

9. An electronic device, characterized in that, The electronic 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