A product promotion content auditing system, a risk filtering system comprising the same and a method

Abstract

The present application relates to a kind of product publicity content's auditing system, the risk filtering system and method comprising it.Auditing system includes the serial connection in turn: text acquisition module, for extracting text to be audited;Text segmentation module, for cutting text to be audited into text to be audited paragraph;Identification module, for screening out the first structured object set with constituting violation risk in text to be audited paragraph;Original text backtracking module, for each structured object in the first structured object set is combined as complete context with text to be audited, obtain the second structured object set;Review module, for checking the second structured object set, obtain the second structured object set after cleaning;Report output module, for the second structured object set after cleaning is converted into audit report.The auditing system of the present application improves the consistency, accuracy and adjustability of audit result.

Description

Technical Field

[0001] This invention relates to the field of advertising compliance review technology; specifically, this invention relates to a product advertising content review system, a risk filtering system and method including the same. Background Technology

[0002] With the booming development of e-commerce and social media marketing, product promotion channels are becoming increasingly diversified, and the forms and quantity of promotional content are growing exponentially. At the same time, the compliance requirements of laws and regulations such as the Advertising Law, the Anti-Unfair Competition Law, and increasingly specific product areas (e.g., the Cosmetics Supervision and Management Regulations and the Shanghai Cosmetics Industry Advertising Compliance Guidelines in the cosmetics field; the Medical Device Supervision and Management Regulations in the medical field; and the Health Food Registration and Filing Management Measures in the health food field) are becoming increasingly stringent. Companies face the following dual pressures: internal compliance pressure, needing to ensure that large-scale, frequent promotional content does not violate advertising regulations; and external whistleblowing pressure, as professional whistleblowers use automated web scraping tools to monitor brand promotional pages in bulk. Once they discover violations that could lead to administrative enforcement, they initiate reports, potentially resulting in administrative penalties, product removal, reputational damage, and the need to pay substantial settlements to professional whistleblowers.

[0003] Currently, companies mainly use the following methods to conduct compliance reviews of product promotional content:

[0004] Firstly, the manual review method involves corporate legal personnel or external compliance experts reviewing each piece of promotional material. This method heavily relies on the professional experience of the reviewers, resulting in low efficiency and high costs. Furthermore, when faced with a large number of SKUs and frequently updated promotional content, it is prone to oversights and inconsistent standards, failing to meet the needs of large-scale review.

[0005] Secondly, the automated review method based on keyword matching, which uses a pre-set list of prohibited words such as "first," "most," and "national level" for string matching, has the following fatal flaws: it can only identify obviously prohibited words and cannot understand semantic synonyms and implicit expressions; it is completely ineffective for scenarios that require semantic inference. For example, although "water-oil balance" does not directly use the words "oil control," it clearly points to "oil control effect" from the semantic perspective of advertising law, and the keyword matching method cannot identify this kind of association; it cannot handle numerical comparisons of quantitative claims, such as "30% increase in translucency," which needs to be compared with the filed experimental data "20%" to determine whether it is a violation.

[0006] Third, while the comprehensive risk warning method based on a generalized large language model can identify more potential problems, its output has structural flaws: there is a fundamental misalignment between the semantic understanding logic of the large language model and the logic of corporate compliance review. Taking the common cosmetic claim of "brightening skin" as an example, whether it constitutes a "whitening effect claim" is legally open to interpretation; it is a "soft risk" that can be defended through supplementary evidence or reasonable explanation. Directly indicating a violation would lead to an "over-rectification" dilemma for the company. Furthermore, the generalized large language model has the following shortcomings in corporate compliance review scenarios: First, it lacks knowledge in the initial state, requiring additional investment to acquire domain knowledge; second, even if knowledge is acquired through investment, reliability at the execution level cannot be guaranteed; third, even if execution reliability is barely usable under certain conditions, rule changes brought about by product iterations require renewed investment for each adjustment, making it commercially unsustainable. Even if these shortcomings are overcome, the general language model cannot distinguish between the certainty and severity of risks. Companies need to invest a lot of manpower to re-examine each risk output by the model, which does not actually reduce the compliance burden. At the same time, it lacks the ability to compare and correlate with the specific approved scope of product claims, such as the efficacy of cosmetics registration or the indications for medical device registration, and cannot determine whether a certain promotional statement exceeds the scope of promotion that the product is legally allowed to make.

[0007] Fourth, review methods based on keyword matching or Retrieval-Augmented Generation (RAG) also struggle to address the aforementioned issues. It's important to note that RAG essentially still uses a large language model as its core engine for reasoning and generation—its workflow involves converting multimodal promotional content into vectors, retrieving relevant fragments, and then providing the retrieval results as context for the large language model's final judgment. Therefore, the dual shortcomings of the large language model discussed earlier regarding "execution reliability" and "rule iteration cost" are fully inherited in the RAG architecture. Specifically, first, RAG's output also possesses inherent randomness. Due to the probabilistic generation mechanism of the underlying large language model, even if the retrieval stage returns the same contextual content, the model may still produce inconsistent conclusions in different calls during comprehensive judgment, failing to guarantee consistent results for multiple reviews of the same promotional material. This is unacceptable in compliance review scenarios requiring determinism and repeatability. Second, RAG also lacks the ability to clearly define and strictly enforce rules. When a company's review needs change—for example, shifting the focus from detecting absolute terms to identifying false advertising, or adjusting the comparison criteria for efficacy claims—the RAG system cannot adapt to these changes with simple configuration modifications like a rule engine. Modifying search strategies, re-tuning prompts, and re-verifying output stability all come with unpredictable costs and risks. While RAG alleviates the "knowledge blind spots" of general-purpose language models to some extent, it doesn't solve the problems of "unreliable execution" or "uneconomical iteration." Therefore, in terms of both output determinism and rule flexibility, RAG-based review methods are essentially no different from directly calling general-purpose language models in practical effect; neither can meet the basic requirements of corporate compliance review for determinism, consistency, and controllability. Summary of the Invention

[0008] In view of this, the present invention provides a product promotion content review system, a risk filtering system and method including the same, thereby solving or at least mitigating one or more of the above-mentioned problems and other problems existing in the prior art.

[0009] To achieve the aforementioned objectives, a first aspect of the present invention provides a product promotion content review system, the review system comprising, in sequence, the following components:

[0010] The text acquisition module is used to acquire the product's promotional content and extract the text to be reviewed from the promotional content. The promotional content includes one or more of the following: promotional copy, promotional images containing text, and promotional videos containing audio.

[0011] The text segmentation module is used to segment the text to be reviewed into multiple text segments.

[0012] The identification module is used to filter out a first set of structured objects that pose a risk of violation from the text paragraph to be reviewed based on a preset keyword database and / or semantic understanding model. The keys of the first set of structured objects include the original sentence of the paragraph and the words used in the original text.

[0013] The original text backtracking module is used to combine each structured object in the first set of structured objects with the text to be reviewed as a complete context, and to derive a second set of structured objects based on preset backtracking rules. The keys of the second set of structured objects include one or more of the following: the original sentence in the full text, the words used in the original text, relevant legal provisions, review conclusions, and reasons for judgment. The backtracking rules are user-defined and configured.

[0014] The review module is used to check the second structured object set based on preset review rules to obtain a cleaned second structured object set. The review rules are user-defined and configured.

[0015] The report output module is used to convert the cleaned second set of structured objects into an audit report in human natural language.

[0016] Optionally, in the auditing system described above, the auditing system further includes a support detection module, wherein,

[0017] The text acquisition module is also used to acquire the approved and compliant advertising scope of the product;

[0018] The support detection module is used to compare each structured object in the first set of structured objects with the approved compliant publicity scope using a preset semantic understanding model, so as to obtain a first set of structured objects that exceeds the compliant support scope.

[0019] The original text backtracking module is used to combine each structured object in the first set of structured objects that exceeds the scope of compliance support with the text to be reviewed as a complete context, and obtain the second set of structured objects based on the backtracking rules.

[0020] Optionally, in the aforementioned review system, the review system further includes a pre-review module, wherein,

[0021] The pre-approval module is used to check the first set of structured objects based on preset rules to obtain a cleaned set of structured objects.

[0022] The original text backtracking module is used to combine each structured object in the first set of cleaned structured objects with the text to be reviewed as a complete context, and obtain the second set of structured objects based on the backtracking rules.

[0023] In the aforementioned review system, optionally, in the identification module, the semantic understanding model deconstructs the rule text into the smallest logical unit, and constructs a structured template object of applicable conditions-behavioral patterns-result constraints based on the smallest logical unit. The rule text includes one or more of laws and regulations, industry standards, platform specifications, and corporate regulations.

[0024] In the aforementioned review system, optionally, in the text acquisition module, when the promotional content is a promotional image containing text, the text acquisition module is used to recognize and extract the OCR text information from the promotional image using an OCR model; when the promotional content is a promotional video containing audio, the text acquisition module is used to recognize the audio in the promotional video using an ASR model and convert the audio into ASR text information; simultaneously...

[0025] When the promotional content includes one type, the text acquisition module is used to use the promotional copy, the OCR text information, or the ASR text information as the text to be reviewed; when the promotional content includes multiple types, the text acquisition module is used to merge the promotional copy, the OCR text information, and / or the ASR text information according to preset rules to obtain the text to be reviewed.

[0026] In the aforementioned review system, optionally, when recognizing and extracting OCR text information from the promotional image using an OCR model:

[0027] The text acquisition module is used to perform text detection on the promotional image, generating multiple candidate text boxes containing text; post-processing the multiple candidate text boxes to obtain multiple valid text regions; based on the multiple valid text regions, cropping the promotional image into multiple text sub-images; performing text recognition on each of the multiple text sub-images to obtain text recognition results; and organizing the text recognition results into the OCR text information according to the reading order; wherein...

[0028] The post-processing includes: non-maximum suppression processing; classifying the candidate text boxes into subscript text boxes and text text boxes, merging the subscript text boxes with the nearest text text box; determining the paragraph relationship between multiple candidate text boxes, merging candidate text boxes belonging to the same paragraph region; determining the nesting relationship between multiple candidate text boxes, deleting candidate text boxes nested inside other candidate text boxes; and expanding the borders of the candidate text boxes.

[0029] To achieve the aforementioned objectives, a second aspect of the present invention provides a risk filtering system for product promotional content, implemented based on the aforementioned review system. The risk filtering system includes multiple concurrent risk review workflows, all of which perform risk filtering on the promotional content through the review system. The risk review workflows include one or more of the following: efficacy consistency review workflow, absolute terminology review workflow, prohibited vocabulary review workflow, and quantitative parameter review workflow.

[0030] The risk filtering system also includes a report fusion workflow, which merges multiple cleaned second structured object sets obtained from the risk audit workflow to obtain a unified second structured object set, and transforms the unified second structured object set into a final audit report in human natural language.

[0031] In the risk filtering system described above, optionally, in the efficacy consistency review workflow, the review system further includes a support detection module. The text acquisition module is used to acquire the product's promotional content and the approved compliant promotional scope, and convert the promotional content into the text to be reviewed. The recognition module is used to filter out a first set of structured objects from the text paragraphs to be reviewed based on a preset semantic understanding model, where the original text uses efficacy descriptions. The support detection module is used to compare each structured object in the first set of structured objects with the approved compliant promotional scope to obtain a first set of structured objects that exceeds the approved compliant promotional scope. The original text backtracking module is used to combine each structured object in the first set of structured objects that exceeds the approved compliant promotional scope with the text to be reviewed as a complete context, and obtain a second set of structured objects based on the backtracking rules.

[0032] In the absolute term review workflow, the identification module is used to filter out a first set of structured objects from the text paragraph to be reviewed based on a preset keyword database and / or by a semantic understanding model according to preset rules, wherein the original text uses absolute terms; the original text backtracking module is used to combine each structured object in the first set of structured objects with the text to be reviewed as a complete context, exclude structured objects that are obviously not related to product promotional content, and obtain a second set of structured objects; the review module is used to check the second set of structured objects based on the review rules, and obtain a second set of structured objects after cleaning up special exempted words that are permitted by laws and regulations, and use this as the cleaned second set of structured objects;

[0033] In the aforementioned workflow for reviewing prohibited words, the identification module is used to filter out a first set of structured objects from the text paragraph to be reviewed based on a preset keyword database and a semantic understanding model, wherein the words used in the original text are prohibited words.

[0034] In the quantitative parameter review workflow, the review system further includes a support detection module. The text acquisition module is used to acquire the product's promotional content and the verified compliant promotional scope, and convert the promotional content into the text to be reviewed. The recognition module is used to filter out a first set of structured objects from the text paragraphs to be reviewed based on a preset semantic understanding model, where the original text uses quantitative descriptions. The support detection module is used to compare each structured object in the first set of structured objects with the boundary values ​​of the quantitative parameters in the verified compliant promotional scope to obtain a first set of structured objects that exceed the boundary values ​​of the quantitative parameters. The original text backtracking module is used to combine each structured object in the first set of structured objects that exceed the boundary values ​​of the quantitative parameters with the text to be reviewed as a complete context, and obtain a second set of structured objects based on the backtracking rules.

[0035] To achieve the aforementioned objectives, a third aspect of the present invention provides a risk filtering method for product promotional content, implemented based on the risk filtering system described above, the risk filtering method comprising:

[0036] Step S1: Construct a risk audit workflow with multiple concurrent tasks based on different risk audit dimensions. The risk audit workflow includes one or more of the following: efficacy consistency audit workflow, absolute term audit workflow, illegal word audit workflow, and quantitative parameter audit workflow.

[0037] Step S2: Obtain the product's promotional content and extract the text to be reviewed from the promotional content. The promotional content includes one or more of the following: promotional copy, promotional images containing text, and promotional videos containing audio.

[0038] Step S3: Divide the text to be reviewed into multiple text segments;

[0039] Step S4: In the text paragraph to be reviewed, a first set of structured objects is selected based on a preset keyword database and / or semantic understanding model. The keys of the first set of structured objects include the original sentence of the paragraph and the words used in the original text.

[0040] Step S5: Combine each structured object in the first set of structured objects with the text to be reviewed as a complete context, and derive a second set of structured objects based on preset backtracking rules. The keys of the second set of structured objects include one or more of the following: the original sentence in the full text, the words used in the original text, relevant legal provisions, review conclusions, and reasons for judgment.

[0041] Step S6: Check the second structured object set based on the preset review rules to obtain the cleaned second structured object set;

[0042] Step S7: Merge the multiple cleaned second structured object sets obtained from the risk audit workflow to obtain a unified second structured object set;

[0043] Step S8: Transform the unified second structured object set into an audit report in human natural language.

[0044] In the risk filtering method described above, optionally, in step S2: simultaneously obtain the approved compliant advertising scope of the product; in step S4: compare each structured object in the first set of structured objects with the approved compliant advertising scope using a preset semantic understanding model to obtain a first set of structured objects that exceeds the compliance support scope; in step S5: combine each structured object in the first set of structured objects that exceeds the compliance support scope with the text to be reviewed as a complete context, and obtain a second set of structured objects based on the backtracking rules.

[0045] The review system of this invention eliminates the need to convert text into feature vectors and complete the review based on traditional RAG. Instead, it directly employs a serial workflow of text segmentation, keyword recognition, inputting the original text, and verification. This simplifies the review process while maintaining high review accuracy, improving the consistency, accuracy, and repeatability of the review results. Furthermore, the databases, semantic understanding models, and rules used by each module can be explicitly defined and modified in real time according to review requirements. This allows for flexible adjustments to the review focus and standards, and the final review report can trace back to specific clauses and reasons for judgment, making the review results more interpretable. In other words, the review system of this invention effectively utilizes the semantic understanding capabilities of general-purpose large language models while avoiding the random uncertainty of large language model outputs and the high costs of reinforcement training or fine-tuning. Optionally, the review system also integrates a support detection module to accurately determine whether the semantic content of promotional text falls within the approved and compliant scope of product promotion. This specifically avoids high-frequency risks such as over-scope claims and inaccurate quantitative statements in promotional content, which are easily reported by professional whistleblowers, thus possessing higher application value in real-world business scenarios.

[0046] The risk filtering system of this invention includes multiple risk review workflows that run concurrently based on the aforementioned review system. Each review dimension is decomposed into an independent parallel architecture, allowing for flexible selection or expansion of risk review workflows according to product domain and review requirements. Furthermore, each risk review workflow can have its review rules adjusted independently, enhancing the versatility and flexibility of risk filtering in promotional content across various industries. Moreover, the parallel operation of each template significantly improves work efficiency.

[0047] The present invention further provides a risk filtering method for product promotional content, and therefore this risk filtering method also has the above-mentioned advantages. Attached Figure Description

[0048] The disclosure of this invention will become more apparent from the accompanying drawings. It should be understood that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings:

[0049] Figure 1 This is a schematic block diagram illustrating the structure of an embodiment of the review system of the present invention.

[0050] Figure 2 This is a schematic block diagram of an embodiment of the risk filtering system of the present invention. Detailed Implementation

[0051] Referring to the accompanying drawings and specific embodiments, the structure, composition, features, and advantages of the product promotion content review system, risk filtering system and method including it of the present invention will be described below by way of example. However, all descriptions should not be used to limit the present invention in any way.

[0052] Furthermore, for any single technical feature described or implied in the embodiments mentioned herein, or any single technical feature shown or implied in the various figures, the present invention still allows for any combination or deletion of these technical features (or their equivalents) without any technical obstacle, and thus these further embodiments according to the present invention should also be considered within the scope of this description.

[0053] When reviewing and filtering product promotional content such as product information pages on e-commerce platforms, advertising images and text, and promotional videos published on social media for compliance review and risk filtering, existing technical solutions based on keyword comparison and semantic vector retrieval typically lack the ability to categorize efficacy at the semantic level. They cannot automatically classify diverse promotional terms such as "water-oil balance" and "brightening" into corresponding product-approved claim dimensions such as "oil control" and "whitening," thus failing to determine whether the promotional content exceeds the legally permitted scope of product claims. Furthermore, they lack mechanisms for automatically extracting quantitative parameters and comparing them against boundaries, making it impossible to extract quantitative descriptions from promotional content and automatically compare them with approved standard values. Moreover, existing technical solutions are often sequential, fixed processes that focus only on a single or a small number of violation types, unable to simultaneously cover multiple independent review dimensions or achieve real-time flexible adjustments using a scalable parallel architecture.

[0054] Figure 1 This is a schematic block diagram illustrating the structure of an embodiment of the review system of the present invention. In this embodiment, the review system for product promotional content includes a text acquisition module, a text segmentation module, a recognition module, a pre-review module, a support detection module, a original text backtracking module, a review module, and a report output module, which operate sequentially. It should be noted that the review system of the present invention identifies risks in promotional text from a single preset review dimension. This single review dimension has clear and adjustable preset rules. Furthermore, multiple such review systems work in parallel to constitute the risk filtering system for product promotional content of the present invention. The number, types, and specific preset rules of these multiple review systems can be adjusted according to actual needs, thereby achieving multi-dimensional review of the violation risks of product promotional content. The technical details of the review system in this embodiment are further illustrated below using a cosmetic product details page on an e-commerce platform as an example.

[0055] The text acquisition module is used to obtain the product's promotional content and extract the text to be reviewed from it. In this example, you can directly input the promotional copy, images, or videos used by the cosmetic product on the e-commerce platform, or you can input the URL (Uniform Resource Locator) of the product details page into the front-end interface of the text acquisition module. The module will then automatically access the page and parse the content after receiving the URL. The acquired promotional content can specifically include, for example, the product title, promotional copy, ingredient list, as well as the superscripts and corresponding small print annotations in the text and images.

[0056] Specifically, when the promotional content is text, the text acquisition module directly uses that text as the document to be reviewed; when the promotional content is an image containing text, the text acquisition module uses an OCR (Optical Character Recognition) model to recognize and extract the OCR text information from the image; when the promotional content is a video containing audio, the text acquisition module uses an ASR (Automatic Speech Recognition) model to recognize the audio in the video and convert it into ASR text information. Both OCR and ASR technologies are conventional techniques and will not be elaborated upon here.

[0057] E-commerce platforms typically use product detail pages that include various types of promotional content. In this case, the text, extracted OCR text information, and ASR text information can be merged according to preset rules to form the final text to be reviewed. In an optional embodiment, the fusion rules can be preset as follows: when the text in the promotional copy can express complete semantic information, the corresponding sentence-level text in the promotional copy is retained first; when the promotional image contains badges, small print descriptions, and / or restrictions not shown in the promotional copy, or when the promotional video contains supplementary information not shown in the promotional copy, the corresponding OCR text information and ASR text information are added to the text to be reviewed. In other scenario examples, when the product's promotional content is only of a single type, the text acquisition module directly uses the promotional copy, OCR text information, or ASR text information as the text to be reviewed.

[0058] Alternatively, since promotional images on e-commerce platforms often have complex layouts, mixed text and images, and diverse fonts and sizes, traditional OCR models often struggle to recognize complete and semantically coherent text information. The OCR model integrated in the text acquisition module can perform text recognition and extraction using the following process:

[0059] First, text detection is performed on the promotional image, generating multiple candidate text boxes containing text. The candidate text boxes are then post-processed as follows to obtain multiple effective text regions: Non-maximum suppression is used to eliminate redundant candidate text boxes with high overlap; candidate text boxes are distinguished into subscript text boxes and regular text boxes, and subscript text boxes are merged with their nearest regular text boxes; paragraph relationships between candidate text boxes are determined, and candidate text boxes belonging to the same paragraph region are merged; nesting relationships between candidate text boxes are determined, and candidate text boxes nested within other candidate text boxes are deleted; borders of candidate text boxes are expanded, optionally with the downward expansion amount greater than the upward expansion amount, to fully preserve text information. Based on this post-processing, the original promotional image is cropped into multiple text sub-images based on the multiple effective text regions. Text recognition is then performed on each of these text sub-images, and the text recognition results are obtained. Finally, these text recognition results are organized in reading order to form the aforementioned OCR text information. The OCR text information obtained in this way can completely preserve the sentence-level content, the order between sentences and sentences, and the local correlation between the main text and the superscript and small text annotations in the promotional image.

[0060] It is important to note that, under the dimensions of efficacy consistency review and quantitative parameter review, the text acquisition module also needs to obtain the product's verified and compliant scope of advertising. For example, in this example, the efficacy claims of the cosmetic product registered with the National Medical Products Administration (NMPA) and the experimental data information corresponding to the quantitative statements appearing in the advertising content. In examples of other product types, the verified scope of claims can also be the registration information of drugs, the registration indications of medical devices, etc., used to subsequently compare and determine whether a certain advertising statement exceeds the legally permitted and compliant scope of advertising for the product. Under other review dimensions, such as the review of absolute terms or prohibited vocabulary, it is not necessary to obtain additional similar registration information.

[0061] The text segmentation module divides the aforementioned text to be reviewed into multiple text segments, and transmits these segments to the recognition module for parallel recognition. The number and length of the segmented texts can be flexibly adjusted according to the performance of the recognition module and specific review requirements; this invention does not impose any limitations on this. Compared to directly inputting the full text of the text to be reviewed into the recognition module, this segmentation method can flexibly adapt to the context window limitations of the recognition module and significantly improve recognition efficiency.

[0062] The identification module is used to filter out a first set of structured objects that pose a risk of violation from the text paragraph to be reviewed, based on a preset keyword database and / or a semantic understanding model. The keys of this first set of structured objects include the original sentence and the original words used in the text. For example, in the dimension of absolute term review, absolute terms can be identified from the text paragraph to be reviewed based on a preset keyword database and / or by a semantic understanding model according to preset rules, resulting in one or more first set of structured objects, such as:

[0063] {"groups":[

[0064] "[{\"content\":\"National-level product*\",\"wording\":[\"National-level\"]}]",

[0065] "[{\"content\":\"xxxx Beauty JD.com Self-Operated Flagship Store is the only official genuine beauty flagship store opened by xxxx on JD.com\",\"wording\":[\"only\"]}]"

[0066] ]}

[0067] Under the dimension of efficacy consistency review, efficacy description text can be identified from the text paragraphs to be reviewed based on a preset semantic understanding model, and one or more sets of first structured objects can be obtained, for example:

[0068] {"groups":[

[0069] "[{\"content\":\"Post-workout soothing shower gel*\",\"wording\":[\"Post-workout soothing\"]}]",

[0070] "[{\"content\":\"This product specializes in water-oil balance systems\",\"wording\":[\"water-oil balance\"]}]"

[0071] "[{\"content\":\"Using it morning and night helps brighten the skin\",\"wording\":[\"Brightens skin\"]}]"

[0072] ]}

[0073] Among them, "national-level," "unique," "relief after exercise," "water-oil balance," and "brightening skin" are original words in the text to be reviewed that are identified as having a risk of violating regulations according to a preset keyword database and / or semantic understanding model. "National-level product," "xxxx beauty JD.com self-operated flagship store, is the only official genuine beauty flagship store opened by xxxx on JD.com," "relief shower gel after exercise," "this product specializes in water-oil balance system," and "use morning and evening to brighten skin" are the corresponding paragraph sentences that have a risk of violating regulations. Similarly, in dimensions such as quantitative parameter review and violation word review, different keyword databases and / or semantic understanding models can be preset based on review requirements to accurately identify quantitative descriptions and violation words from the text to be reviewed.

[0074] Understandably, the keyword database and semantic understanding model in this recognition module can be adjusted in real time. For example, adding or removing keywords from the keyword database can change the recognition scope, and redefining the semantic understanding model can change the focus and scale of recognition. Compared with commonly used RAG methods, this effectively reduces the randomness of the output results and improves the flexibility to adapt to different review requirements.

[0075] As an optional embodiment, in the recognition module, the aforementioned semantic understanding model can deconstruct the rule text into the smallest logical unit and construct a structured template object based on this smallest logical unit, representing the applicable condition, behavior pattern, and result constraint. Taking Article 9, Paragraph 3 of the Advertising Law, which states that "advertisements shall not use terms such as 'national level,' 'highest level,' or 'best,'" as an example, a template object can be deconstructed and constructed with the applicable condition being "advertisement of goods or services," the behavior pattern being "use of absolute terms," ​​and the result constraint being "constitutes a risk of using absolute terms." Thus, the recognition module can compare the product with the applicable conditions in each template object and invoke the rules adapted to that product. Further optionally, the rule text may include one or more of the following: laws and regulations, industry standards, platform specifications, and corporate regulations.

[0076] The pre-approval module is an optional supplementary module of the approval system in this invention. In this embodiment, it is used to check multiple identified first structured object sets based on preset rules and obtain a cleaned first structured object set. The pre-approval module has similar functions to the subsequent review module and will be described in detail later.

[0077] The support detection module is also an optional supplementary module to the review system. In dimensions such as efficacy consistency review and quantitative parameter review, it compares each structured object in the cleaned first structured object set with the previously obtained approved compliant advertising scope using a preset semantic understanding model, thereby obtaining a first structured object set that exceeds the compliant support scope. Specifically, in the efficacy consistency review dimension, the support detection module may first classify the efficacy description text "water-oil balance" obtained by the identification module into a specific efficacy dimension using a preset semantic classification strategy. In this example, it falls under the "oil control" efficacy dimension, which is one of the efficacy categories allowed by the drug regulatory authority for cosmetic claims. Next, this efficacy dimension is compared with the product's approved compliant advertising scope, i.e., the efficacy claims filed with the drug regulatory authority, through field matching and / or preset rules. If the product only files for efficacies such as "moisturizing" and "soothing" but not "oil control," then the efficacy description text is determined to exceed the approved compliant advertising scope, posing a definite risk of violation. The corresponding first structured object set is the first structured object set that exceeds the compliant support scope. The preset semantic classification strategy can be a prompt word reasoning method based on a large language model, a mapping dictionary, etc. This invention does not limit this, as long as it can achieve accurate classification at the semantic level.

[0078] The following uses the phrase "brighten skin" as an example to illustrate the difference between this functional module of the present invention and existing technical solutions. First, existing open-source or closed-source commercial general-purpose language models, in their initial state without specialized training, have very limited knowledge about "brighten skin." Taking publicly available general-purpose models as an example, the most frequently learned word in public corpora is the high-frequency co-occurrence relationship between "brighten" and "whiten," thus defaulting to considering them highly correlated or even equivalent. At this point, existing models do not naturally know that "brighten skin" can also be achieved through multiple efficacy pathways such as moisturizing, anti-wrinkle, cosmetic enhancement (concealing), and nourishing—this knowledge is not an explicit consensus in general public corpora, but rather domain knowledge that requires combining specific regulations, efficacy evaluation standards, and product claim practices within the cosmetics industry. In other words, general-purpose language models are inherently lacking in knowledge.

[0079] To address this deficiency, companies need to inject additional knowledge into their existing models. This involves methods such as meticulous prompt word engineering, fine-tuning training with industry-specific corpora, or few-shot learning with numerous manually labeled examples. The goal is for the existing models to learn the relationships supporting multi-dimensional efficacy claims, such as "moisturizing increases light reflectivity by raising the stratum corneum's water content, thus supporting the brightening claim; anti-wrinkle improves skin texture, making the skin smoother, thus visually brightening the effect; and concealing blemishes achieves visual brightening by covering imperfections." However, this knowledge injection process itself requires significant time investment, such as repeatedly debugging prompt words or collecting labeled data; economic costs, such as the computational resources or API call fees required for model fine-tuning; and professional expertise, such as requiring personnel with both cosmetic compliance knowledge and experience in large-scale language model engineering. These costs are not affordable for all companies, nor are they readily available from any existing general-purpose model.

[0080] Even assuming that companies overcome the aforementioned costs and successfully equip their models with the knowledge that "bright skin can be supported by multiple effects" through pre-training or refined prompting word engineering, the application-side problems remain unresolved. The essence of corporate compliance review is making definitive compliance judgments for each specific product: Is it permissible for a face cream with only moisturizing efficacy data to claim "bright skin"? Is it permissible for a foundation with only concealing efficacy data to claim "bright skin"? These judgment rules are definitive at the product level—yes or no, must be clear, fixed, and error-free. While existing large language models, after knowledge injection, can understand that "multiple effects may support brightening," they still cannot make a definitive, repeatable attribution judgment for any given product based on its specific, limited efficacy data. Their probabilistic output mechanism means that even with clearly defined judgment rules, the model may still be biased due to contextual interference, ambiguous logical boundaries, or randomness. In other words, compliance review requires absolute certainty, while large language models provide statistical guesses.

[0081] A deeper problem lies in the fact that even assuming a large language model can achieve near-zero error for a fixed set of products at a given moment, companies still face unacceptable iteration costs. A company's products are dynamically changing. For example, a face cream might only have moisturizing efficacy in its first generation, add concealing efficacy in the second, and completely change its formula and claim strategy in the third. With each product iteration, the compliance claim rules need to be adjusted accordingly: previously allowed claims may no longer apply, and previously prohibited claims may become acceptable. Faced with such rule changes, existing models cannot reliably adapt through simple rule replacement. To adapt the model to a completely new claim rule system for different products, companies need to reinvest all of the above costs: re-tuning prompts, reconstructing examples, re-verifying output stability, and even fine-tuning in some cases. This means that each product iteration is accompanied by a new round of high costs and uncontrollable error risks. In a business environment of rapid product iteration, this "reinvestment every time there's a change" model is unreasonable both in terms of time and economy. Companies cannot afford to bear the cost of almost retraining from scratch for every rule change for every product.

[0082] In contrast, the technical solution of this invention can flexibly adjust the focus and standards of review by simply adjusting or replacing the preset database, semantic understanding model, rule text, etc. in each module. That is, while effectively utilizing the semantic understanding capabilities of the general large language model, it avoids the random uncertainty of the output of the large language model and the high cost of reinforcement training or fine-tuning.

[0083] Similarly, under the quantitative parameter review dimension, the support detection module compares the quantitative description text obtained by the identification module with the boundary values ​​of the quantitative parameters within the approved compliant advertising scope of the product, thereby obtaining the first set of structured objects exceeding the boundary values ​​of the quantitative parameters. For example, in the cosmetic registration information published by the State Drug Administration, a certain cosmetic product includes oil-control test data and consumer survey data regarding its oil-control efficacy. For example: "The number of valid test participants was 30 female subjects... After using the product for 24 hours, the amount of oil on the skin in the test area was significantly reduced compared with the blank control. The results show that the product has an oil-control effect that can last up to 24 hours." and "The number of valid consumer survey participants was 112 women, aged 19-55, who used the product at least once a day for 7 consecutive days. The survey results show that... 93.7% of the respondents agreed that the product controlled shine when wearing makeup; 81.2% agreed that the product could permanently correct skin tone; and 90.2% agreed that the product could maintain its color without caking."

[0084] Based on the aforementioned approved and compliant advertising scope, and using a pre-defined semantic understanding model, if the quantitative description of the cosmetic product in its advertising content is "oil control for 24 hours" or "90.2% of consumers believe it is long-lasting and does not cake," then it can be determined that the quantitative value and efficacy of this description are supported by data within the approved and compliant advertising scope, and can be advertised compliantly. However, if the quantitative description is "oil control for 48 hours" or "80% of consumers believe it improves hair color" which is not mentioned in consumer survey data, then it can be determined that the quantitative value and efficacy of this description are not matched by data within the approved and compliant advertising scope, posing a risk of violation. The corresponding first structured object set is thus outside the scope of compliant support.

[0085] The aforementioned two review dimensions correspond to high-frequency professional whistleblowing scenarios in actual business situations: claims of efficacy exceeding the scope and false efficacy figures. Therefore, the review system of this invention can accurately identify and detect high-risk violation types from the source, targeting them with a high probability of being prosecuted. It is understandable that, in dimensions such as the review of absolute terms and prohibited words, when the identification module identifies original text containing absolute terms or prohibited words, it can determine that the original text contains a definite risk of violation, without requiring additional supporting detection modules for comparison.

[0086] The original text backtracking module is used to combine each structured object in the first set of structured objects that exceeds the scope of compliance support with the full text of the text to be reviewed as a complete context, and to filter out the second set of structured objects based on preset backtracking rules. The keys of the second set of structured objects may include one or more of the following: original text content, original text wording, relevant legal provisions, review conclusion, and reason. Its specific form can be adjusted according to different review dimensions or review needs. Since the identification module is based on segmented text paragraphs to be reviewed, semantic information will inevitably be missed in the obtained original sentences. For example, when the same original word appears in multiple places in the full text of the text to be reviewed, the overall understanding based on multiple full-text original sentences and the individual understanding based on a single paragraph original sentence are prone to discrepancies, leading to reduced accuracy in risk identification. Another example is when the full text of the text to be reviewed involves multiple independent products in a product portfolio, where the main product is allowed to be advertised while the subsidiary products are prohibited from advertising "anti-wrinkle" effects. The original word "anti-wrinkle" appearing in a certain paragraph of the text to be reviewed needs to be confirmed in conjunction with the full text to determine its actual corresponding independent product. The original text backtracking module is used to correct this discrepancy, further improving the reliability and stability of the review system of this invention.

[0087] Based on the various scenarios listed above, the backtracking rules can be as follows: Analyze the semantic attributes of the same original words in multiple full-text sentences, and merge the results of each semantic attribute to arrive at the review conclusion; construct the full-text sentences corresponding to each independent product based on the full text of the text to be reviewed, determine the independent products corresponding to the original words based on the construction results, and analyze the compliant promotional scope of each independent product to arrive at the review conclusion. It is important to note that these backtracking rules in the original text backtracking module are customizable and flexibly adjusted by the user based on actual review needs. The "preset" backtracking rules here do not refer to a fixed specific rule or the reasoning logic within the model, but rather allow the user to dynamically configure them through one or more methods, such as loading configuration files, inputting in a visual rule editor, configuring natural language prompts, or defining code scripts. For example, the output of the example in the aforementioned recognition module in the original text backtracking module could be:

[0088] {"output":[

[0089] {"output":[{\"conclusion\":\"Absolute words that need to be indicated in the context of advertising\",\"content\":\"National-level product*\",\"provision\":\"No applicable exemption\",\"reason\":\"Does not meet the exemption\",\"risk_level\":\"high\",\"time\":\"\",\"wording\":[\"National-level\"]}]},

[0090] {"output":[{\"conclusion\":\"Please provide supporting materials or verify the authenticity. If there is solid supporting material, it does not constitute an absolute term in the context of advertising language\",\"content\":\"xxxx beauty JD self-operated flagship store is the only official beauty genuine flagship store opened by xxxx on JD\",\"provision\":\"Article 6 (6) of the "Guidelines for the Enforcement of Absolute Terms in Advertising" states that under the condition of limiting specific time, region, etc., it describes the objective situation of time and space order or promotes the factual information such as product sales volume, sales revenue, and market share\",\"reason\":\"Objective factual information under the condition of limiting time, region, or channel\",\"time\":\"\",\"wording\":[\"unique\"]}]}

[0091] ]}

[0092] The review module, similar to the aforementioned pre-approval module, checks the second set of structured objects based on preset review rules to obtain a cleaned set of structured objects. Optionally, the review module, like the pre-approval module, can not only clean key-value pairs in the second set of structured objects that do not conform to the review rules, but also add risk level fields to key-value pairs that do conform to the review rules. Similar to the preset backtracking rules in the original document backtracking module, the preset review rules here are also customized and flexibly adjusted by the user based on actual review needs.

[0093] It is important to note that in the review system of this invention, the upstream and downstream cooperation between the original text backtracking module and the review module is a fixed system mechanism. That is, the original text backtracking module performs full-text backtracking on structured objects and attaches a judgment mark such as "review result," while the review module cleans the text based on the judgment mark. However, the specific backtracking and review rules used are customizable by the user according to actual business needs. This design, with a fixed mechanism and variable standards, allows the review system of this invention to maintain both the determinism and repeatability of the workflow while possessing the flexibility to adapt to different business scenarios and rule changes.

[0094] Continuing with the previous example, after review, objects that do not constitute violations can be cleaned up and replaced with empty objects to avoid additional manual review workload. For example:

[0095] {"output":[

[0096] {"output":[{\"conclusion\":\"Absolute words that need to be indicated in the context of advertising\", \"content\":\"National-level product*\",\"provision\":\"No applicable exemption\",\"reason\":\"Does not meet the exemption\",\"risk_level\":\"high\",\"time\":\"\",\"wording\":[\"National-level\"]}]},

[0097] {"output":[]}

[0098] ]}

[0099] In other examples, when the original text word identified by the identification module is "first place", while the corresponding paragraph sentence or full text sentence is "the famous person who is first place", the corresponding second structured object set will also be cleaned up by the pre-review module or review module because the absolute term is not related to the product being promoted.

[0100] The report output module transforms the cleaned second structured object set into an audit report in human natural language. Based on the structural characteristics of the second structured object set, the audit report not only contains simple conclusions or labels indicating whether a risk exists, but also allows for tracing the conclusion or label back to the original text to be audited, the rules and regulations on which the judgment was based, and the complete record of the audit process, thereby improving the certainty and efficiency of manual review and result auditing.

[0101] By understanding the workflow of each module in the review system of this invention, it can be understood that by flexibly adjusting the keyword database, semantic understanding model, and rule text under different review dimensions, the review system can achieve identification and judgment under stricter standards. This results in an review report that only includes definitive risks such as claims of efficacy exceeding the permitted scope, inaccurate efficacy values, absolute terms, and prohibited vocabulary, avoiding excessive rectification and reducing the workload of subsequent manual verification. In other words, the review system of this invention has a review focus and review scale that can be adjusted in real time, which is the core distinguishing feature from existing full-risk warning review systems.

[0102] A second aspect of the present invention provides a risk filtering system for product promotional content based on the aforementioned review system. Figure 2 This is a schematic block diagram illustrating the structure of an embodiment of the risk filtering system of the present invention. Figure 2 As shown, the risk filtering system includes multiple concurrent risk review workflows, all of which filter promotional content through the aforementioned review system. In this embodiment, the risk review workflows include efficacy consistency review, absolute terminology review, prohibited vocabulary review, and quantitative parameter review. In other optional embodiments, different risk review workflows can be expanded or adjusted based on product type and industry regulations, or on other review dimensions, such as reviewing medical terminology for non-medical products. These risk review workflows execute synchronously and in parallel, without interdependence, and each outputs multiple sets of second structured objects obtained under different review dimensions.

[0103] The risk filtering system also includes a report fusion workflow, which merges these multiple cleaned second structured object sets into a unified second structured object set and transforms it into a final audit report in human natural language. In an optional embodiment, when multiple cleaned second structured object sets contain the same original text terms under different audit dimensions, such as the original text term "treatment of melasma" in non-medical, pharmaceutical, or medical device advertisements, which may pose risks under both the efficacy consistency audit dimension and the medical terminology audit dimension, the report fusion workflow merges the cleaned second structured object sets corresponding to the original text term and displays the corresponding risk conclusions for that original text term in the final audit report.

[0104] Those skilled in the art will understand that these risk review workflows can be as follows: Figure 2 As shown, the same text acquisition module is used, while each text segmentation module and recognition module are executed separately. Alternatively, the text acquisition module and text segmentation module can be used simultaneously, or each can have its own independent text acquisition module and text segmentation module. Such simple variations in architecture fall within the protection scope of this invention.

[0105] This multi-dimensional, concurrent review system architecture for risk filtering significantly improves review efficiency compared to the commonly used serial processing solutions. Furthermore, it can be flexibly adapted to different product types, such as cosmetics, medical devices, electronic products, and food and health products, by adding or deleting risk review workflows, demonstrating good versatility and scalability.

[0106] Based on the aforementioned review system and risk filtering system, the third aspect of this invention discloses a risk filtering method for product promotional content. This method includes the following steps: First, constructing multiple concurrent risk review workflows based on different risk review dimensions. These workflows are adapted to the product type and may include one or more of the following: efficacy consistency review workflow, absolute terminology review workflow, prohibited vocabulary review workflow, and quantitative parameter review workflow. Second, acquiring the product's promotional content, such as promotional copy, promotional images containing text, and / or promotional videos containing audio, and extracting the text to be reviewed from the promotional content. Under the dimensions of efficacy consistency review and quantitative parameter review, the product's verified and compliant promotional scope also needs to be obtained simultaneously.

[0107] Then, the text to be reviewed is segmented into multiple paragraphs. Based on a pre-defined keyword database and / or semantic understanding model, a first set of structured objects is selected from these paragraphs. The keys of this first set of structured objects include the original sentences and the original words used in the paragraphs. Under different review dimensions, the original words can be, for example, efficacy statements, quantitative statements, absolute terms, and prohibited vocabulary. Specifically, under the dimensions of efficacy consistency review and quantitative parameter review, each structured object in the first set of structured objects needs to be compared with the approved compliant scope of publicity using a pre-defined semantic understanding model. A first set of structured objects that exceeds the scope of compliance support is then obtained for subsequent steps.

[0108] Each structured object in the aforementioned first set of structured objects, or a set exceeding the scope of compliance support, is combined with the text to be reviewed as a complete context for unified contextual processing. Based on preset backtracking rules, a second set of structured objects is obtained. The keys of this second set of structured objects include one or more of the following: original sentence, original wording, relevant legal provisions, review conclusion, and reasons for judgment. Subsequently, the second set of structured objects is checked based on preset review rules to obtain a cleaned second set of structured objects. Finally, the multiple cleaned second set of structured objects obtained from multiple parallel risk review workflows are merged to obtain a unified second set of structured objects, which is then transformed into a review report in human natural language.

[0109] The product promotion content review system, risk filtering system, and risk filtering method of this invention, after data verification in a real production environment, show a 97% consistency rate between their review results and the manual review results of senior compliance experts. This data fully demonstrates that this invention, while reducing labor costs and improving review efficiency, possesses extremely high review accuracy.

[0110] The technical scope of this invention is not limited to the contents of the above specification. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the scope of this invention.

Claims

1. A product promotion content review system, characterized in that, The auditing system comprises components connected in a sequential series: The text acquisition module is used to acquire the product's promotional content and extract the text to be reviewed from the promotional content. The promotional content includes one or more of the following: promotional copy, promotional images containing text, and promotional videos containing audio. The text segmentation module is used to segment the text to be reviewed into multiple text segments. The identification module is used to filter out a first set of structured objects that pose a risk of violation from the text paragraph to be reviewed based on a preset keyword database and / or semantic understanding model. The keys of the first set of structured objects include the original sentence of the paragraph and the words used in the original text. The original text backtracking module is used to combine each structured object in the first structured object set with the text to be reviewed as a complete context, and derive a second structured object set based on preset backtracking rules. The keys of the second structured object set include one or more of the following: the original sentence in the full text, the words used in the original text, relevant legal and regulatory provisions, review conclusions, and reasons for judgment. The backtracking rules are user-defined and configured. The review module is used to check the second structured object set based on preset review rules to obtain a cleaned second structured object set. The review rules are user-defined and configured. The report output module is used to convert the cleaned second set of structured objects into an audit report in human natural language.

2. The auditing system as described in claim 1, characterized in that, The auditing system also includes a supportive detection module, wherein, The text acquisition module is also used to acquire the approved and compliant advertising scope of the product; The support detection module is used to compare each structured object in the first set of structured objects with the approved compliant publicity scope using a preset semantic understanding model, so as to obtain a first set of structured objects that exceeds the compliant support scope. The original text backtracking module is used to combine each structured object in the first set of structured objects that exceeds the scope of compliance support with the text to be reviewed as a complete context, and obtain the second set of structured objects based on the backtracking rules.

3. The auditing system as described in claim 1, characterized in that, The review system also includes a pre-review module, wherein... The pre-approval module is used to check the first set of structured objects based on preset rules to obtain a cleaned set of structured objects. The original text backtracking module is used to combine each structured object in the first set of cleaned structured objects with the text to be reviewed as a complete context, and obtain the second set of structured objects based on the backtracking rules.

4. The auditing system as described in claim 1, characterized in that, In the recognition module, the semantic understanding model deconstructs the rule text into the smallest logical unit and constructs a structured template object based on the smallest logical unit, which includes applicable conditions, behavior patterns, and result constraints. The rule text includes one or more of the following: laws and regulations, industry standards, platform specifications, and corporate regulations.

5. The auditing system as described in claim 1, characterized in that, In the text acquisition module, when the promotional content is a promotional image containing text, the text acquisition module is used to recognize and extract the OCR text information from the promotional image using an OCR model; when the promotional content is a promotional video containing audio, the text acquisition module is used to recognize the audio in the promotional video using an ASR model and convert the audio into ASR text information; simultaneously... When the promotional content includes one type, the text acquisition module is used to use the promotional copy, the OCR text information, or the ASR text information as the text to be reviewed; when the promotional content includes multiple types, the text acquisition module is used to merge the promotional copy, the OCR text information, and / or the ASR text information according to preset rules to obtain the text to be reviewed.

6. The auditing system as described in claim 5, characterized in that, When recognizing and extracting OCR text information from the promotional image using an OCR model: The text acquisition module is used to perform text detection on the promotional image and generate multiple candidate text boxes containing text; post-process the multiple candidate text boxes to obtain multiple effective text regions; based on the multiple effective text regions, crop the promotional image into multiple text sub-images; and perform text recognition on the multiple text sub-images to obtain text recognition results. The text recognition results are organized into the OCR text information according to the reading order; wherein... The post-processing includes: non-maximum suppression processing; classifying the candidate text boxes into subscript text boxes and text text boxes, merging the subscript text boxes with the nearest text text box; determining the paragraph relationship between multiple candidate text boxes, merging candidate text boxes belonging to the same paragraph region; determining the nesting relationship between multiple candidate text boxes, deleting candidate text boxes nested inside other candidate text boxes; and expanding the borders of the candidate text boxes.

7. A risk filtering system for product promotional content, implemented based on the review system as described in claim 1, characterized in that, The risk filtering system includes multiple concurrent risk review workflows, all of which filter the promotional content through the system. These risk review workflows include one or more of the following: efficacy consistency review workflow, absolute terminology review workflow, prohibited vocabulary review workflow, and quantitative parameter review workflow. The risk filtering system also includes a report fusion workflow, which merges multiple cleaned second structured object sets obtained from the risk audit workflow to obtain a unified second structured object set, and transforms the unified second structured object set into a final audit report in human natural language.

8. The risk filtering system as described in claim 7, characterized in that, In the efficacy consistency review workflow, the review system further includes a support detection module. The text acquisition module is used to acquire the product's promotional content and the approved compliant promotional scope, and convert the promotional content into the text to be reviewed. The recognition module is used to filter out a first set of structured objects from the text paragraphs to be reviewed based on a preset semantic understanding model, where the original text uses efficacy descriptions. The support detection module is used to compare each structured object in the first set of structured objects with the approved compliant promotional scope to obtain a first set of structured objects that exceeds the approved compliant promotional scope. The original text backtracking module is used to combine each structured object in the first set of structured objects that exceeds the approved compliant promotional scope with the text to be reviewed as a complete context, and obtain a second set of structured objects based on the backtracking rules. In the absolute term review workflow, the identification module is used to filter out a first set of structured objects from the text paragraph to be reviewed based on a preset keyword database and / or by a semantic understanding model according to preset rules, wherein the original text uses absolute terms; the original text backtracking module is used to combine each structured object in the first set of structured objects with the text to be reviewed as a complete context, exclude structured objects that are obviously not related to product promotional content, and obtain a second set of structured objects; the review module is used to check the second set of structured objects based on the review rules, and obtain a second set of structured objects after cleaning up special exempted words that are permitted by laws and regulations, and use this as the cleaned second set of structured objects; In the aforementioned workflow for reviewing prohibited words, the identification module is used to filter out a first set of structured objects from the text paragraph to be reviewed based on a preset keyword database and a semantic understanding model, wherein the words used in the original text are prohibited words. In the quantitative parameter review workflow, the review system further includes a support detection module. The text acquisition module is used to acquire the product's promotional content and the verified compliant promotional scope, and convert the promotional content into the text to be reviewed. The recognition module is used to filter out a first set of structured objects from the text paragraphs to be reviewed based on a preset semantic understanding model, where the original text uses quantitative descriptions. The support detection module is used to compare each structured object in the first set of structured objects with the boundary values ​​of the quantitative parameters in the verified compliant promotional scope to obtain a first set of structured objects that exceed the boundary values ​​of the quantitative parameters. The original text backtracking module is used to combine each structured object in the first set of structured objects that exceed the boundary values ​​of the quantitative parameters with the text to be reviewed as a complete context, and obtain a second set of structured objects based on the backtracking rules.

9. A method for risk filtering of product promotional content, implemented based on the risk filtering system as described in claim 7, characterized in that, The risk filtering method includes: Step S1: Construct a risk audit workflow with multiple concurrent tasks based on different risk audit dimensions. The risk audit workflow includes one or more of the following: efficacy consistency audit workflow, absolute term audit workflow, illegal word audit workflow, and quantitative parameter audit workflow. Step S2: Obtain the product's promotional content and extract the text to be reviewed from the promotional content. The promotional content includes one or more of the following: promotional copy, promotional images containing text, and promotional videos containing audio. Step S3: Divide the text to be reviewed into multiple text segments; Step S4: In the text paragraph to be reviewed, a first set of structured objects is selected based on a preset keyword database and / or semantic understanding model. The keys of the first set of structured objects include the original sentence of the paragraph and the words used in the original text. Step S5: Combine each structured object in the first set of structured objects with the text to be reviewed as a complete context, and derive a second set of structured objects based on preset backtracking rules. The keys of the second set of structured objects include one or more of the following: the original sentence in the full text, the words used in the original text, relevant legal provisions, review conclusions, and reasons for judgment. Step S6: Check the second structured object set based on the preset review rules to obtain the cleaned second structured object set; Step S7: Merge the multiple cleaned second structured object sets obtained from the risk audit workflow to obtain a unified second structured object set; Step S8: Transform the unified second structured object set into an audit report in human natural language.

10. The risk filtering method as described in claim 9, characterized in that, In step S2: the approved and compliant advertising scope of the product is obtained simultaneously; in step S4: each structured object in the first set of structured objects is compared with the approved and compliant advertising scope using a preset semantic understanding model to obtain a first set of structured objects that exceeds the scope of compliance support. In step S5: Each structured object in the first set of structured objects that exceeds the scope of compliance support is combined with the text to be reviewed as a complete context, and the second set of structured objects is obtained based on the backtracking rules.

Images