A risk claim identification method and system based on an agent and multi-modal analysis

Abstract

This invention relates to a risk claim identification method and system based on intelligent agents and multimodal analysis. The method includes: extracting multimodal visual features and reconstructing the text sequence from a graphic carrier containing claim elements to obtain a continuous text sequence to be reviewed; adaptively recalling a subset of relevant regulatory constraints from a regulatory vector knowledge base based on the text sequence to be reviewed through a retrieval enhancement generation mechanism; performing intelligent agent cognitive reasoning based on dynamic prompt words, fusing the text sequence to be reviewed with the recalled regulatory constraint subset, driving a large language model to perform deep semantic deconstruction through a thought chain to identify high-dimensional risk features of the risk claim; and serializing the risk identification results into standardized key-value pair structured decision data for compliance control. This invention can overcome the blind spots of multimodal review, construct a high-confidence dynamic review benchmark, output standardized data to connect with downstream systems, achieve full-link automated compliance control, and improve the accuracy and efficiency of review.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data processing technology, and in particular to a risk claim identification method and system based on intelligent agents and multimodal analysis. Background Technology

[0002] With the continuous improvement of regulatory systems in industries such as cosmetics and food, compliance control of product claims has become a core and mandatory requirement for industry operations, and digital compliance review has become a key component in the digital transformation process of relevant enterprises. During the industry's digital development, product marketing and promotional materials are taking on multimodal and diverse forms, while regulatory laws and standards are constantly being updated and iterated, placing higher demands on the coverage, accuracy, response efficiency, and end-to-end control capabilities of compliance reviews.

[0003] Currently, in digital compliance review scenarios in industries such as cosmetics and food, the industry generally adopts rule engine-based or static dictionary matching methods to conduct compliance verification of claimed content. Some solutions introduce basic optical character recognition technology to extract text content from image materials, and then use conventional natural language analysis models to conduct compliance assessment of the extracted text content, thereby completing the basic compliance review process.

[0004] In existing digital compliance scenarios in industries such as cosmetics and food, three major technical bottlenecks urgently need to be overcome: First, there are technical barriers to the parsing of multimodal unstructured data. The massive amounts of product packaging designs, marketing posters, and other multimedia materials lack effective automated extraction and semantic reconstruction methods, resulting in serious "visual blind spots" in compliance reviews. Second, there is the dynamic evolution of regulations and standards and the lag in understanding them. Regulatory systems such as the "Cosmetic Labeling Management Measures" are iterated frequently, and traditional methods based on rule engines or static dictionary matching cannot handle "implied violations" in complex contexts, and the maintenance costs are extremely high. Third, natural language analysis results are difficult to deeply decouple from business systems. Existing models output mostly unstructured natural language segments, lacking standardized decision data models, and cannot be directly mapped to downstream business flows such as ERP blocking and approval flow control, resulting in low efficiency in the entire closed-loop process. Summary of the Invention

[0005] In view of this, it is necessary to provide a risk claim identification method based on intelligent agents and multimodal analysis to address the aforementioned shortcomings of the existing technology.

[0006] To address the aforementioned problems, in a first aspect, embodiments of the present invention provide a risk claim identification method based on intelligent agents and multimodal analysis, comprising: Multimodal visual feature extraction and text sequence reconstruction are performed on graphic carriers containing claim elements to restore a continuous text sequence to be reviewed in the real context. Based on the text sequence to be reviewed, relevant subsets of regulatory constraints are adaptively retrieved from the regulatory vector knowledge base through a vertical domain regulatory knowledge engine and a retrieval enhancement generation mechanism. Based on the dynamic prompt words, the intelligent agent performs cognitive reasoning and integrates the text sequence to be reviewed with the recalled regulatory constraint subset to drive the large language model to perform deep semantic deconstruction and identify the high-dimensional risk features of risk claims. The risk identification results corresponding to the high-dimensional risk features are serialized into standardized key-value pair structured decision data, and compliance management is performed based on the structured decision data.

[0007] Furthermore, the step of performing multimodal visual feature extraction and text sequence reconstruction on the graphic carrier containing the claim elements to restore a continuous text sequence to be reviewed in the real context includes: Instantiate a fusion model that integrates optical character recognition and layout analysis functions; Using the optical character recognition function of the fusion model, a deep visual neural network is used to extract spatial features from the graphic carrier and identify the coordinates of the text box and its topological relationship. Based on the layout analysis function of the fusion model, the discrete visual text blocks are semantically spliced ​​together by the reading order reconstruction algorithm to restore the continuous text sequence to be reviewed in the real context.

[0008] Furthermore, based on the text sequence to be reviewed, a subset of relevant regulatory constraints is adaptively retrieved from the regulatory vector knowledge base through a vertical domain regulatory knowledge engine and a retrieval enhancement generation mechanism, including: Construct a high-frequency, dynamically updated regulatory vector knowledge base; The industry's guidelines for terminology, relevant regulations, and historical cases of violations are analyzed into atomized knowledge entities; Knowledge entities are mapped into high-dimensional feature vectors through a text embedding model and stored in a regulatory vector knowledge base; When the text sequence to be reviewed is received, a similarity calculation is performed on the text sequence to be reviewed, and a subset of relevant regulatory constraints is adaptively recalled from the vector space.

[0009] Furthermore, the process of assembling and executing intelligent agent cognitive reasoning based on dynamic prompt words, fusing the text sequence to be reviewed with the recalled regulatory constraint subset, drives a large language model to perform deep semantic deconstruction, and identifies high-dimensional risk features of risk claims, including: Configure a multi-dimensional constraint-type instruction template, which encapsulates character settings, task instructions, anti-illusion boundary constraints, and input slots; The text sequence to be reviewed is dynamically fused with the recalled subset of regulatory constraints and injected into the instruction template to form fused context prompts; The driving language model performs deep semantic deconstruction by simulating the thought process of industry compliance reviewers to identify high-dimensional risk features of risk claims; among which, the high-dimensional risk features include absolute terms, claims beyond the scope, and medical implications.

[0010] Furthermore, the step of serializing the risk identification results corresponding to the high-dimensional risk features into standardized key-value pair structured decision data, and performing compliance management based on the structured decision data, includes: By embedding output paradigm constraint instructions into the prompt word project, the large language model is constrained to serialize the risk identification results into a set of key-value pairs in standard JSON format, thereby obtaining key-value pair structured decision data. Parse key-value pair structured decision data and implement compliance management based on the parsing results.

[0011] Furthermore, the parsing of key-value pair structured decision data, and the execution of compliance management based on the parsing results, includes: Map the entity coordinates of risk words in the key-value pairs back to the physical coordinates of the original image and text carrier, and then perform visual highlighting; Based on the risk identification results in the key-value pairs, the circuit breaker or transfer mechanism of the business system is triggered to complete the compliance control closed loop.

[0012] Furthermore, the graphic carriers containing the claim elements include: product packaging design drawings, marketing posters, PDF documents, and e-commerce detail page images in the cosmetics or food industry.

[0013] Secondly, embodiments of the present invention provide a risk claim identification system based on intelligent agents and multimodal analysis, comprising: The input layer is used to receive graphic carriers containing claim elements; The parsing layer is used to extract multimodal visual features and reconstruct text sequences from graphic carriers containing claim elements, thereby restoring a continuous text sequence to be reviewed in the real context. The core processing layer is used to adaptively recall a subset of relevant regulatory constraints from the regulatory vector knowledge base based on the text sequence to be reviewed, through a vertical domain regulatory knowledge engine and a retrieval enhancement generation mechanism; and is used to assemble and execute intelligent agent cognitive reasoning based on dynamic prompt words, fuse the text sequence to be reviewed and the recalled subset of regulatory constraints, drive the large language model to perform deep semantic deconstruction, and identify the high-dimensional risk features of risk claims. The output and control layer is used to serialize the risk identification results corresponding to the high-dimensional risk features into standardized key-value pair structured decision data, and to perform compliance control based on the structured decision data.

[0014] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the risk claim identification method based on intelligent agent and multimodal analysis as described in the first aspect embodiment of the present invention.

[0015] Fourthly, the present invention also provides a computer-readable storage medium for storing a computer-readable program or instructions, which, when executed by a processor, can implement the steps in the risk claim identification method based on intelligent agent and multimodal analysis as described in the first aspect embodiment of the present invention.

[0016] The risk claim identification method and system based on intelligent agents and multimodal analysis provided by this invention have the following advantages compared with the prior art: (1) Break through the blind spot of unstructured review and realize multimodal holographic control: By instantiating the optical character recognition and layout analysis fusion model, combined with deep visual feature extraction and reading order reconstruction technology, the traditional "unreadable" multimodal graphic carriers are transformed into semantically coherent text sequences to be reviewed, eliminating blind spots in the review of promotional materials and achieving comprehensive coverage and accurate analysis of various complex promotional materials.

[0017] (2) Overcoming model illusion and constructing a high-confidence dynamic review benchmark: An innovative retrieval enhancement generation mechanism is introduced to construct a high-frequency dynamically updated regulatory vector knowledge base, parsing industry guidelines, regulations, and historical violation cases into atomic knowledge entities and storing them quantitatively. By adaptively recalling a subset of regulatory constraints that matches the text to be reviewed through semantic similarity calculation, the AI ​​agent's decision-making is anchored to the latest regulations, effectively avoiding the knowledge lag and excessive divergence of large language models, and improving the legal accuracy and reliability of violation judgment.

[0018] (3) Construct a standardized decision-making data model to enable seamless integration of the entire business flow: By embedding output paradigm constraint instructions in the prompt word project, the large language model is forced to output key-value pair structured data in standard JSON format, covering core information such as risk word positioning, risk level, violation type and compliance strategy; through parsing and mapping by the routing control layer, the risk area is visually highlighted, and the automated circuit breaker or transfer mechanism of the business system is triggered, breaking down the data barriers between the compliance system and downstream systems such as ERP and OA within the enterprise, and completing the millisecond-level full-link compliance control closed loop. Attached Figure Description

[0019] Figure 1 Flowchart of the risk claim identification method based on intelligent agent and multimodal analysis provided by the present invention; Figure 2 A schematic diagram illustrating the dynamic construction of the vertical domain regulatory vector knowledge engine provided by this invention; Figure 3 This is a schematic diagram of the intelligent agent prompt word assembly and key-value pair mapping control provided by the present invention; Figure 4 This invention provides an overall architecture diagram of a risk claim identification system based on intelligent agents and multimodal analysis. Figure 5 This is a structural block diagram of the electronic device provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The naming or numbering of steps in the embodiments of the present invention does not mean that the steps in the method flow must be executed in the time / logical order indicated by the naming or numbering. The execution order of the named or numbered process steps can be changed according to the technical purpose to be achieved, as long as the same or similar technical effect can be achieved.

[0022] Figure 1 A flowchart illustrating the risk claim identification method based on intelligent agents and multimodal analysis provided by this invention. (Refer to...) Figure 1 The method includes the following steps: Step S1 involves extracting multimodal visual features and reconstructing the text sequence of the graphic carrier containing the claim elements to restore the continuous text sequence to be reviewed in the real context.

[0023] In this embodiment, multimodal visual feature extraction and text sequence reconstruction are performed on complex graphic carriers containing claim elements to restore a continuous text sequence to be reviewed in the real context.

[0024] First, a fusion model integrating optical character recognition (OCR) and layout analysis is instantiated. Optical character recognition (OCR) is used to recognize the text content in the image, and layout analysis is used to parse the layout structure of the graphic carrier.

[0025] Next, using the optical character recognition function of the fusion model, a deep visual neural network is used to extract spatial features from the graphic carrier containing the claim elements. The deep visual neural network can accurately capture the visual features of the graphic carrier, thereby identifying the coordinates of the text boxes in the graphic carrier and the topological relationship between each text box, clarifying the specific position of each text block in the graphic carrier and the spatial relationship between them.

[0026] Subsequently, based on the layout analysis function of the fusion model, the reading order reconstruction algorithm is used to semantically splice the extracted discrete visual text blocks, adjust the order of the text blocks according to the normal human reading logic, integrate the originally scattered and disordered text blocks into coherent text, restore the continuous text sequence to be reviewed in the real context, thereby eliminating the interference of complex layout on the semantic coherence of the text.

[0027] Understandably, complex textual and graphic formats containing claims related to cosmetics come in a wide variety of forms, including high-resolution design drafts, PDF documents, and e-commerce product detail page images. This type of multimodal unstructured data presents parsing barriers and lacks effective automated text extraction and semantic reconstruction methods, resulting in significant blind spots in compliance reviews. Furthermore, the diverse layouts of these complex textual and graphic formats, with their fragmented text blocks, can disrupt semantic coherence and affect the accuracy of subsequent compliance reviews.

[0028] Therefore, this embodiment, by instantiating a fusion model of optical character recognition and layout analysis, and combining a deep visual neural network with a reading order reconstruction algorithm, achieves multimodal visual feature extraction and text sequence reconstruction of complex graphic and text carriers. It effectively identifies text box coordinates and topological relationships, eliminates interference from complex layouts, and restores a semantically coherent and continuous sequence of text to be reviewed. This solves the barriers and visual blind spots in multimodal data parsing in existing technologies, providing accurate and complete text input for subsequent steps such as regulatory constraint subset recall and risk feature identification, ensuring the accuracy and effectiveness of the compliance review process.

[0029] Step S2: Based on the text sequence to be reviewed, the relevant subset of regulatory constraints is adaptively retrieved from the regulatory vector knowledge base through a vertical domain regulatory knowledge engine and a retrieval enhancement generation mechanism.

[0030] Understandably, in existing technologies, regulatory rules and standards for industries such as cosmetics and food are constantly and frequently iterated. Traditional compliance review methods based on rule engines or static dictionary matching suffer from cognitive lag and struggle to handle suggestive violations in complex contexts. Furthermore, the cost of updating and maintaining rules and dictionaries is high. Therefore, this invention introduces a Retrieval Augmented Generation (RAG) mechanism. By constructing a dynamically updated regulatory vector knowledge base, it recalls the most relevant regulatory content to the text to be reviewed in each review task, using it as context for the large language model's inference, thereby improving the accuracy and interpretability of the review results.

[0031] In some embodiments of this application, step S2 specifically includes the following steps S21 to S24: S21, Construct a high-frequency, dynamically updated regulatory vector knowledge base.

[0032] The regulatory vector knowledge base is a database that stores textual content such as regulatory clauses, industry guidelines, and historical violation cases converted into high-dimensional vectors using a text embedding model. It supports semantic similarity retrieval and is the core knowledge source for the retrieval enhancement generation (RAG) mechanism. In this embodiment, the regulatory vector knowledge base is built for compliance review scenarios in industries such as cosmetics and food, supporting real-time updates of the knowledge content based on regulatory iterations and the release of new violation cases. The database is stored in vector form, supporting rapid semantic retrieval.

[0033] Figure 2 This is a schematic diagram illustrating the dynamic construction of the vertical domain regulatory vector knowledge engine provided by the present invention. Figure 2 This document fully presents the entire process of building a regulatory vector knowledge base, from data source input to final storage. It is divided into four interconnected core stages. The first stage is the data sources, which form the foundation of the entire knowledge engine. This includes three core categories: guidelines for cosmetic claims, relevant industry regulatory documents, and a historical penalty case database. The second stage is data processing, where the core operations are data cleaning and atomic segmentation. This stage standardizes and cleans the raw data source, breaking it down into semantically independent regulatory clause text chunks, thus achieving atomic parsing of knowledge entities. The third stage is vector transformation, where the core transformation is performed by the text embedding model. This stage converts the regulatory clause text chunks generated in the previous stage into multi-dimensional vector numbers that can represent semantic features, completing the mapping from natural language text to high-dimensional numerical vectors. The fourth stage is storage, culminating in a high-dimensional vector database that supports high-frequency dynamic updates. This database standardizes and stores the transformed multi-dimensional vector numbers, forming the final regulatory vector knowledge base usable for semantic retrieval.

[0034] S22 parses industry claim terminology guidelines, relevant regulations, and historical violation cases into atomized knowledge entities.

[0035] In this embodiment, the input original content, such as guidelines for cosmetic claims, industry regulatory documents, and historical cases of violations and penalties, undergoes data cleaning and standardization. The complete content is then broken down into the smallest, indivisible knowledge units with independent semantics, completing the parsing of atomized knowledge entities. This step is a prerequisite knowledge processing step for the retrieval enhancement generation (RAG) mechanism. Here, atomized knowledge entities refer to the breakdown of regulatory texts or cases into the smallest knowledge units, such as clauses, definitions, and prohibitive descriptions.

[0036] S23 uses a text embedding model to map knowledge entities into high-dimensional feature vectors and stores them in a regulatory vector knowledge base.

[0037] This step is the core semantic transformation process of the retrieval enhancement generation RAG mechanism, and it forms the technical foundation for realizing RAG semantic retrieval. It calls the text embedding model to convert each atomic knowledge entity generated in the previous step into a high-dimensional feature vector that can represent its semantic features. Then, all high-dimensional feature vectors are stored in the vector space corresponding to the regulatory vector knowledge base, completing the vectorization transformation and storage of knowledge entities, thus providing the foundation for subsequent semantic retrieval.

[0038] S24, when the text sequence to be reviewed is received, a similarity calculation is performed on the text sequence to be reviewed, and a subset of relevant regulatory constraints is adaptively recalled from the vector space.

[0039] Specifically, after receiving the text sequence to be reviewed generated in step S1, the system converts the text sequence into a corresponding high-dimensional feature vector using the same text embedding model. In this embodiment, the text embedding model is a neural network model that maps text into fixed-length vectors, which can be used for semantic similarity calculation and information retrieval tasks.

[0040] Then, in the vector space of the regulatory vector knowledge base, similarity calculation between vectors is performed to filter out high-dimensional feature vectors that are semantically related to the text sequence to be reviewed, recall the atomic knowledge entities associated with the high-dimensional feature vectors, and form a subset of regulatory constraints. This subset will serve as the dynamic contextual prior knowledge for subsequent large language model reasoning, complete the core retrieval action of the RAG mechanism, inject dynamic and authoritative regulatory knowledge into the large model reasoning, and avoid the problem of static knowledge illusion.

[0041] This embodiment effectively addresses the issues of static knowledge illusion in large models and the lag in understanding iterative regulations by constructing a frequently updated regulatory vector knowledge base and combining it with a retrieval enhancement mechanism, thereby reducing the maintenance cost of compliance rules. It adaptively recalls a subset of regulatory constraints that matches the text to be reviewed, providing authoritative and accurate dynamic contextual prior knowledge for subsequent risk identification and reasoning, thus improving the accuracy and timeliness of compliance reviews.

[0042] Step S3: Based on the dynamic prompt words, assemble and execute the intelligent agent's cognitive reasoning, integrate the text sequence to be reviewed with the recalled regulatory constraint subset, drive the large language model to perform deep semantic deconstruction, and identify the high-dimensional risk features of risk claims.

[0043] In compliance review scenarios, traditional methods based on rule engines or static dictionaries can only identify surface-level keywords and cannot handle suggestive violations in complex contexts, such as those conveyed through metaphors or vague expressions. While large language models possess powerful semantic understanding capabilities, direct use may lead to knowledge illusions or deviations from the latest regulations. Therefore, this invention introduces an intelligent agent architecture, combining dynamic regulatory knowledge recalled by a retrieval-enhanced generation mechanism. Through multi-dimensional constraint instruction templates, the model is guided to perform step-by-step reasoning using a thought chain, thereby achieving accurate identification of deep semantic risks, improving the accuracy and interpretability of the review, and preventing the model from fabricating regulatory basis out of thin air.

[0044] In some embodiments of this application, step S3 specifically includes the following steps S31 to S33: S31, Configure a multi-dimensional constraint-type instruction template, which encapsulates role settings, task instructions, anti-illusion boundary constraints, and input slots.

[0045] Specifically, a structured prompt template is pre-designed, comprising four core components: First, role setting, positioning the large language model as a professional cosmetics compliance review expert, requiring it to possess industry regulatory knowledge and review experience; second, task instructions, explicitly requiring the model to identify and mark any non-compliant claims based on the text to be reviewed and the recalled regulatory constraint subset; third, anti-illusion boundary constraints, instructing the model to make judgments only based on the provided recalled regulations, and prohibiting it from fabricating or referencing external regulations, ensuring the authority of the output basis; and fourth, input slots, reserving two dynamically filled positions for subsequent injection of the text sequence to be reviewed and the regulatory constraint subset, respectively.

[0046] S32, dynamically fuse the text sequence to be reviewed with the recalled regulatory constraint subset, and inject it into the instruction template to form a fusion context prompt.

[0047] The reconstructed continuous sequence of text to be reviewed obtained in step S1, and the subset of legal constraints related to the semantics of the text to be reviewed recalled in step S2, are respectively filled into the input slots corresponding to the instruction template to generate a complete prompt word, namely, a context-aware prompt word. This prompt word enables the large language model to simultaneously obtain the content to be reviewed and the relevant legal basis, providing sufficient contextual information for subsequent reasoning.

[0048] S33, the driving large language model performs deep semantic deconstruction by simulating the thought process of industry compliance reviewers to identify high-dimensional risk characteristics of risk claims; wherein, the high-dimensional risk characteristics include absolute terms, claims beyond the scope, and medical implications.

[0049] The fused contextual cue words are input into the large language model, and the chain of thought reasoning mechanism is enabled. After receiving the fused contextual cue words, the large language model simulates the thinking logic of industry compliance reviewers, analyzes the semantic content of the text to be reviewed step by step through the chain of thought, and makes compliance judgments in combination with the recalled subset of regulatory constraints, accurately identifying high-dimensional risk features, including absolute terms, claims beyond the scope, and medical implications.

[0050] For example, the analysis can be broken down into the following steps: First, understand the semantic meaning of the text to be reviewed and identify potentially problematic words and expressions. Next, compare each clause in the recalled regulatory subset to determine if the text contains absolute terms, claims beyond its scope, or medical implications. Then, synthesize the reasoning results and output high-dimensional risk characteristics, including specific prohibited words, violation types such as absolute terms, claims beyond their scope, medical implications, and brief explanations. This process simulates the thinking logic of industry compliance reviewers, ensuring accurate capture of deep semantic risks.

[0051] This embodiment standardizes the reasoning behavior of large language models by assembling multidimensional constraint-type instruction templates and dynamic prompt words. Combined with agent cognitive reasoning and thought chain analysis, it achieves accurate identification of high-dimensional risk features in complex contexts. It effectively avoids the knowledge illusion of large models, overcomes the limitations of traditional rule matching, and can accurately identify risks such as absolute terms, overreaching claims, and medical implications, thus improving the accuracy and reliability of compliance reviews.

[0052] Step S4: Serialize the risk identification results corresponding to the high-dimensional risk features into standardized key-value pair structured decision data, and perform compliance management based on the structured decision data.

[0053] In existing compliance review technologies, the assessment results generated by large language models are usually output in the form of natural language paragraphs. Although this unstructured data is easy for humans to read, it is difficult for downstream business systems such as Enterprise Resource Planning (ERP) and Office Automation (OA) systems to directly identify and process it. As a result, the review results cannot be quickly transformed into actual control actions, and the efficiency of the entire automated closed loop is low.

[0054] To address this issue and break down the barriers between intelligent decision-making and business execution, this invention introduces output paradigm constraint instructions in the prompt word engineering, forcing the large language model to serialize the risk identification results into standardized key-value pair structured data. This ensures that the output results retain semantic information while also being machine readable. The output is then parsed and mapped to specific business actions through the routing control layer, thereby achieving full-process automation from risk identification to business control.

[0055] In some embodiments of this application, step S4 is implemented through structured semantic mapping and routing control mechanisms, specifically including the following sub-steps S41~S42: S41. In the prompt word project, an output paradigm constraint instruction is embedded to constrain the large language model to serialize the risk identification results into a set of key-value pairs (KV) in standard JSON (JavaScript Object Notation) format, thereby obtaining key-value pair structured decision data.

[0056] The system adds an output paradigm constraint instruction to the multidimensional constraint instruction template used in step S3. This instruction explicitly requires the large language model to serialize the identified high-dimensional risk features into a set of key-value pairs in JSON format according to a predefined feature space after completing the thought chain reasoning. The predefined feature space includes four core fields: target entity, risk level, violation type, and compliance strategy. The first field is the target entity, Target_Entity, which serves as the key. Its value is the risky keyword entity that is precisely anchored, such as a specific illegal word or phrase. The second field is the Risk Level, which is used as the Value to quantify the risk level, including three levels: critical risk, suspected violation, and compliance. The third field is the Violation Type (Value), which maps to the specific legal category of the violation, such as exaggerated efficacy claims or unauthorized use of medical terminology. The fourth field is Compliance_Strategy, which, as the value, is a smartly generated compliance alternative or refactoring suggestion.

[0057] After receiving contextual prompts, the large language model strictly follows the above constraints in the output stage to generate a JSON object. This JSON object is key-value pair structured decision data, which fully encapsulates risk words, risk levels, violation types, and compliance suggestions, providing standardized input for downstream management.

[0058] S42 parses key-value pair structured decision data and performs compliance management based on the parsing results.

[0059] The system parses the key-value pair structured decision data generated by the routing control layer. First, it extracts the coordinates of risk-related keywords from the target entity (Target_Entity), maps these coordinates back to the physical coordinates of the original text and image carrier, and performs visual highlighting to emphasize the location of the risk content for quick manual identification. Then, based on the risk level (Risk_Level) and violation type (Violation_Type), the routing control layer triggers corresponding control actions from the business system: for high-risk content, a circuit breaker mechanism is implemented to block it from entering subsequent business processes; for content awaiting review or low-risk content, a flow mechanism is implemented to push it to the corresponding processing node for further review or rectification, ultimately completing the intelligent compliance control closed loop.

[0060] Figure 3 This is a schematic diagram illustrating the intelligent agent prompt word assembly and key-value pair mapping control provided by the present invention. Figure 3 The collaborative implementation process of step S3, intelligent agent cognitive reasoning, and step S4, structured control is presented. Figure 3 The left side represents the input pool of prompt words, which is the core source of contextual prompt words. It comprises three elements: first, the text fragment to be reviewed generated in step S1, such as "the number one anti-inflammatory and whitening essence on the entire internet"; second, a subset of regulatory constraints adaptively recalled in step S2, such as "cosmetics are prohibited from using medical terms like 'anti-inflammatory' and absolute terms like 'number one'"; and third, output paradigm constraints, explicitly requiring the model to output key-value pairs in JSON format. The middle section is the large-scale model's thought process reasoning stage. The model simulates the thinking logic of industry compliance reviewers, first analyzing the claim terms in the text to be reviewed, identifying "anti-inflammatory" and "number one on the entire internet" as two suspected violations, and then comparing them with the prohibitive clauses of the recalled regulations, determining them as violations of medical suggestion and violations of absolute terms, respectively, completing deep semantic deconstruction and risk identification.

[0061] Figure 3The right side shows the key-value pair structured output. The model strictly follows the output paradigm constraints to generate decision data in standard JSON format, covering four core fields: Target_Entity anchors risky terms such as "anti-inflammatory" and "number one on the entire network"; Risk_Level is marked as "high risk"; Violation_Type corresponds to "medical implications" and "absolute terms"; and Compliance_Strategy provides rectification solutions such as "suggest deletion or replacement with 'soothing' and 'best-selling'". This structured data not only fully preserves semantic reasoning conclusions but also has a machine-readable standardized format, providing accurate basis for the routing control layer. It supports visual highlighting of risky terms and automated control actions such as circuit breaking and transfer in business systems, achieving a closed loop from risk identification to business handling.

[0062] This embodiment transforms unstructured risk identification results into machine-parseable standard JSON data through output paradigm constraints and key-value pair structured mapping, enabling seamless integration with downstream business systems. The routing control layer directly parses the JSON data and maps it to physical coordinates and business actions, achieving visual highlighting of risk terms and automated circuit breaking or transfer of business systems, forming a closed-loop control across the entire chain. This mechanism improves the usability and response efficiency of compliance review results, reduces the cost of manual intervention, ensures that risks can be detected and handled in a timely manner, and builds an efficient, accurate, and integrable intelligent compliance system for enterprises.

[0063] Figure 4 The overall architecture diagram of the risk claim identification system based on intelligent agent and multimodal analysis provided by this invention is shown below. Figure 4 Risk claim identification systems based on intelligent agents and multimodal analysis include: The input layer is used to receive graphic carriers containing claim elements; The parsing layer is used to extract multimodal visual features and reconstruct text sequences from graphic carriers containing claim elements, thereby restoring a continuous text sequence to be reviewed in the real context. The core processing layer is used to adaptively recall a subset of relevant regulatory constraints from the regulatory vector knowledge base based on the text sequence to be reviewed, through a vertical domain regulatory knowledge engine and a retrieval enhancement generation mechanism; and is used to assemble and execute intelligent agent cognitive reasoning based on dynamic prompt words, fuse the text sequence to be reviewed and the recalled subset of regulatory constraints, drive the large language model to perform deep semantic deconstruction, and identify the high-dimensional risk features of risk claims. The output and control layer is used to serialize the risk identification results corresponding to the high-dimensional risk features into standardized key-value pair structured decision data, and to perform compliance control based on the structured decision data.

[0064] The risk claim identification system based on intelligent agent and multimodal parsing provided by the present invention is used to execute the risk claim identification method based on intelligent agent and multimodal parsing provided in the foregoing embodiments. The risk claim identification method based on intelligent agent and multimodal parsing has been described in detail in the above embodiments, and will not be repeated here.

[0065] Reference Figure 4 The input layer serves as the system's data entry point, receiving multimodal source files containing claim elements, such as complex graphic and textual carriers like images, PDF design drafts, and screenshots of e-commerce detail pages, providing a data foundation for subsequent processing.

[0066] The parsing layer includes an OCR (Optical Character Recognition) module and a layout analysis and sequence reconstruction module. The OCR module extracts spatial features of the text and image carrier through a deep visual neural network, and identifies the coordinates and topological relationships of text boxes. The layout analysis module, based on a reading order reconstruction algorithm, semantically concatenates discrete text blocks into coherent text, restores the text sequence to be reviewed in the real context, and eliminates barriers to multimodal data parsing and visual blind spots.

[0067] The core processing layer (Agent layer) comprises three core modules: a regulatory vector knowledge base (including the RAG engine, which stores industry regulations, claim guidelines, and violation cases), a prompt word engineering module, and a large language model inference engine. First, the RAG mechanism adaptively recalls a subset of regulatory constraints that semantically match the text to be reviewed from the frequently updated regulatory vector knowledge base. Then, the prompt word engineering module merges the text to be reviewed with the regulatory basis to generate contextual prompt words. Finally, the large language model simulates the thought process of compliance reviewers, deeply deconstructing semantics to accurately identify high-dimensional risk features such as absolute terms, claims beyond their scope, and medical implications, thus avoiding model knowledge illusions.

[0068] The output and control layer includes a key-value structured data parser and a highlighting interception and compliance suggestion module. The parser serializes the risk identification results into standard JSON format key-value pair decision data; after parsing the data, the highlighting interception module maps the coordinates of risk terms back to the original carrier to perform visual highlighting, and triggers the circuit breaker or transfer mechanism of the business system according to the risk level, completing the compliance control closed loop and achieving seamless integration with downstream ERP, OA and other systems.

[0069] This invention provides a risk claim recognition system based on intelligent agents and multimodal analysis. By integrating an optical character recognition and layout analysis model, it extracts visual features and reconstructs the reading order of text and image carriers containing claim elements, achieving automated semantic reconstruction of complex multimedia materials and eliminating visual blind spots in compliance review. By constructing a dynamically updated regulatory vector knowledge base, it adaptively recalls relevant regulatory constraint subsets based on the text sequence to be reviewed, accurately addressing high-frequency regulatory iterations and suggestive violations, reducing manual maintenance costs. Through intelligent agent cognitive reasoning, it accurately identifies high-dimensional risk features and outputs standardized key-value pairs that directly connect to downstream business systems, achieving a closed-loop, efficient compliance management system across the entire chain.

[0070] Figure 5 A structural block diagram of the electronic device provided by the present invention, such as Figure 5 As shown, the present invention also provides an electronic device 500, which can be a mobile terminal, desktop computer, laptop, handheld computer, server, or other computing device. The electronic device 500 includes a processor 501 and a memory 502, wherein the memory 502 stores a risk claim identification program 503 based on intelligent agent and multimodal analysis.

[0071] In some embodiments, memory 502 may be an internal storage unit of a computer device, such as a hard drive or memory. In other embodiments, memory 502 may be an external storage device of a computer device, such as a plug-in hard drive, smart media card (SMc), secure digital card (SD), flash card, etc. Further, memory 502 may include both internal and external storage units of the computer device. Memory 502 is used to store application software and various types of data installed on the computer device, such as program code for installing the computer device. Memory 502 can also be used to temporarily store data that has been output or will be output. In one embodiment, when the risk claim identification program 503 based on intelligent agent and multimodal analysis is executed by processor 501, the following steps are implemented: Multimodal visual feature extraction and text sequence reconstruction are performed on graphic carriers containing claim elements to restore a continuous text sequence to be reviewed in the real context. Based on the text sequence to be reviewed, relevant subsets of regulatory constraints are adaptively retrieved from the regulatory vector knowledge base through a vertical domain regulatory knowledge engine and a retrieval enhancement generation mechanism. Based on the dynamic prompt words, the intelligent agent performs cognitive reasoning and integrates the text sequence to be reviewed with the recalled regulatory constraint subset to drive the large language model to perform deep semantic deconstruction and identify the high-dimensional risk features of risk claims. The risk identification results corresponding to the high-dimensional risk features are serialized into standardized key-value pair structured decision data, and compliance management is performed based on the structured decision data.

[0072] In some embodiments, processor 501 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 502 or process data, such as executing a risk claim identification program based on intelligent agent and multimodal analysis.

[0073] This embodiment also provides a computer-readable storage medium storing a risk claim identification program based on intelligent agents and multimodal parsing. When the risk claim identification program based on intelligent agents and multimodal parsing is executed by a processor, it implements the steps of the risk claim identification method based on intelligent agents and multimodal parsing.

[0074] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An agent-based and multi-modal resolution-based risk claim identification method, characterized in that, include: Multimodal visual feature extraction and text sequence reconstruction are performed on graphic carriers containing claim elements to restore a continuous text sequence to be reviewed in the real context. Based on the text sequence to be reviewed, relevant subsets of regulatory constraints are adaptively retrieved from the regulatory vector knowledge base through a vertical domain regulatory knowledge engine and a retrieval enhancement generation mechanism. Based on the dynamic prompt words, the intelligent agent performs cognitive reasoning and integrates the text sequence to be reviewed with the recalled regulatory constraint subset to drive the large language model to perform deep semantic deconstruction and identify the high-dimensional risk features of risk claims. The risk identification results corresponding to the high-dimensional risk features are serialized into standardized key-value pair structured decision data, and compliance management is performed based on the structured decision data.

2. The risk claim identification method based on intelligent agent and multimodal analysis according to claim 1, characterized in that, The process of extracting multimodal visual features and reconstructing text sequences from graphic carriers containing claim elements to restore a continuous text sequence to be reviewed in the real context includes: Instantiate a fusion model that integrates optical character recognition and layout analysis functions; Using the optical character recognition function of the fusion model, a deep visual neural network is used to extract spatial features from the graphic carrier and identify the coordinates of the text box and its topological relationship. Based on the layout analysis function of the fusion model, the discrete visual text blocks are semantically spliced ​​together by the reading order reconstruction algorithm to restore the continuous text sequence to be reviewed in the real context. 3.The risk claim identification method based on agent and multi-modal resolution according to claim 1, characterized in that, Based on the text sequence to be reviewed, a subset of relevant regulatory constraints is adaptively retrieved from the regulatory vector knowledge base using a vertical domain regulatory knowledge engine and a retrieval enhancement generation mechanism, including: Construct a high-frequency, dynamically updated regulatory vector knowledge base; The industry's guidelines for terminology, relevant regulations, and historical cases of violations are analyzed into atomized knowledge entities; Knowledge entities are mapped into high-dimensional feature vectors through a text embedding model and stored in a regulatory vector knowledge base; When the text sequence to be reviewed is received, a similarity calculation is performed on the text sequence to be reviewed, and a subset of relevant regulatory constraints is adaptively recalled from the vector space. 4.The risk claim identification method based on agent and multi-modal resolution according to claim 1, characterized in that, The process involves assembling and executing an intelligent agent for cognitive reasoning based on dynamic prompts, fusing the text sequence to be reviewed with the recalled subset of regulatory constraints, driving a large language model to perform deep semantic deconstruction, and identifying high-dimensional risk features of risk claims, including: Configure a multi-dimensional constraint-type instruction template, which encapsulates character settings, task instructions, anti-illusion boundary constraints, and input slots; The text sequence to be reviewed is dynamically fused with the recalled subset of regulatory constraints and injected into the instruction template to form a fused context prompt. The driving language model performs deep semantic deconstruction by simulating the thought process of industry compliance reviewers to identify high-dimensional risk features of risk claims; among which, the high-dimensional risk features include absolute terms, claims beyond the scope, and medical implications. 5.The risk claim identification method based on agent and multi-modal resolution according to claim 1, characterized in that, The step of serializing the risk identification results corresponding to the high-dimensional risk features into standardized key-value pair structured decision data, and performing compliance management based on the structured decision data, includes: By embedding output paradigm constraint instructions into the prompt word project, the large language model is constrained to serialize the risk identification results into a set of key-value pairs in standard JSON format, thereby obtaining key-value pair structured decision data. Parse key-value pair structured decision data and implement compliance management based on the parsing results. 6.The risk claim identification method based on agent and multi-modal resolution according to claim 5, characterized in that, The parsing of key-value pair structured decision data, and the execution of compliance management based on the parsing results, include: Map the entity coordinates of risk words in the key-value pairs back to the physical coordinates of the original image and text carrier, and then perform visual highlighting; Based on the risk identification results in the key-value pairs, the circuit breaker or transfer mechanism of the business system is triggered to complete the compliance control closed loop. 7.The risk claim identification method based on agent and multi-modal resolution according to claim 1, characterized in that, The graphic and textual carriers containing claim elements include: product packaging design drawings, marketing posters, PDF documents, and e-commerce detail page images in the cosmetics or food industry.

8. An agent and multi-modal resolution based risk claim identification system for performing the agent and multi-modal resolution based risk claim identification method of any one of claims 1-7. The system includes: The input layer is used to receive graphic carriers containing claim elements; The parsing layer is used to extract multimodal visual features and reconstruct text sequences from graphic carriers containing claim elements, thereby restoring a continuous text sequence to be reviewed in the real context. The core processing layer is used to adaptively recall a subset of relevant regulatory constraints from the regulatory vector knowledge base based on the text sequence to be reviewed, through a vertical domain regulatory knowledge engine and a retrieval enhancement generation mechanism; and is used to assemble and execute intelligent agent cognitive reasoning based on dynamic prompt words, fuse the text sequence to be reviewed and the recalled subset of regulatory constraints, drive the large language model to perform deep semantic deconstruction, and identify the high-dimensional risk features of risk claims. The output and control layer is used to serialize the risk identification results corresponding to the high-dimensional risk features into standardized key-value pair structured decision data, and to perform compliance control based on the structured decision data.

9. An electronic device, characterized in that Including memory and processor, among which, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the risk claim identification method based on intelligent agent and multimodal analysis as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the steps in the risk claim identification method based on intelligent agent and multimodal analysis as described in any one of claims 1 to 7.

Images