Method and apparatus for generating an item sample

Abstract

The present disclosure provides a kind of generation method and device of article sample, belong to network field.In the present disclosure embodiment, by processing multiple seed article samples and detection standard text, the annotation of each article sample is obtained, based on the annotation, the description text in article sample and the attention object in article image are clarified, and the prediction standard item is also clarified, then, when generating the article sample corresponding to seed article sample based on the annotation, each seed article sample and the target detection result of each seed article sample, the description text and the attention object in article image can be varied, and the prediction standard item can also be integrated into seed article sample, and then the article sample obtained contains more rich standard item combination, the generation of article sample is realized, the form of article sample is more rich, the quantity of article sample is improved, and then the timely iteration of risk control model can be realized, and the risk control model has low leakage rate for the article that does not meet the detection standard.

Description

Technical Field

[0001] This disclosure relates to the field of network technology, and in particular to a method and apparatus for generating an item sample. Background Technology

[0002] With the development of network technology, the number of items posted on e-commerce platforms has exploded. Current e-commerce risk control systems use large models (i.e., risk control models) and fixed review rules to intercept items that do not meet the detection standards. However, some merchants, in order to circumvent the review and control of e-commerce platforms, often continuously iterate on the methods of disguising content that does not meet the detection standards in their items. The accuracy of the risk control model in identifying items needs to be trained iteratively, which means it relies on a large number of labeled item samples. In a short period of time after the disguising methods are iterated, it is difficult to obtain enough item samples, resulting in a long iteration lag time for the risk control model. Therefore, within a certain time range, the risk control model has a high miss rate for items that do not meet the detection standards. Thus, a method that can generate item samples is needed. Summary of the Invention

[0003] This disclosure provides a method and apparatus for generating item samples, which realizes the generation of item samples, increases the number of item samples, and enables timely iteration of risk control models.

[0004] According to one aspect of the present disclosure, a method for generating item samples is provided. The method includes: acquiring a plurality of seed item samples, detection standard text, and target detection results for each seed item sample, wherein the seed item sample includes descriptive text and an image of the item, the detection standard text is used to provide standard items for item detection, and the target detection results indicate whether the item sample to be generated conforms to the detection standard text; processing the plurality of seed item samples and the detection standard text to obtain a label for each item sample, wherein the label includes at least one of an object of interest and a predicted standard item in the descriptive text and the item image, the object of interest including an object that hits a standard item in the detection standard text, and the predicted standard item including a standard item that is not hit in the detection standard text; generating an item sample corresponding to each seed item sample based on each seed item sample, the target detection results for each seed item sample, and the label, wherein the item sample conforms to the target detection result of the corresponding seed item sample, and the item sample includes at least one of a first item sample and a second item sample, wherein at least one of the descriptive text and the item image of the first item sample contains information that conforms to the predicted standard item, and the second item sample includes a variant object of the object of interest in the corresponding seed item sample.

[0005] According to another aspect of the embodiments of this disclosure, an apparatus for generating an article sample is provided, the apparatus comprising: The first acquisition module is configured to acquire multiple seed item samples, detection standard text, and target detection results for each seed item sample. The seed item samples include item description text and item images. The detection standard text is used to provide standard items for item detection. The target detection results indicate whether the item sample to be generated conforms to the detection standard text. The processing module is configured to process the plurality of seed item samples and detection standard text to obtain a label for each item sample. The label includes at least one of the object of interest and the predicted standard item in the description text and the item image. The object of interest includes the object that hits the standard item in the detection standard text, and the predicted standard item includes the standard item that is not hit in the detection standard text. The generation module is configured to generate an item sample corresponding to each seed item sample based on each seed item sample, the target detection results of each seed item sample, and the annotation. The item sample conforms to the target detection results of the corresponding seed item sample. The item sample includes at least one of a first item sample and a second item sample. At least one of the descriptive text and item image of the first item sample contains information that conforms to the prediction criteria. The second item sample includes a variant of the object of interest in the corresponding seed item sample.

[0006] In some embodiments, the processing module is configured to extract features from the descriptive text and the detection standard text in any seed item sample to obtain a first text feature; extract features from the item image in the seed item sample to obtain a first image feature; fuse the first text feature and the first image feature based on an attention mechanism to obtain a first fused feature; and map the first fused feature to obtain the annotation of the seed item sample.

[0007] In some embodiments, the annotation includes text annotation and image annotation, wherein the text annotation is an annotation of descriptive text, the image annotation is an annotation of an image of an object, and the generation module includes a text variation submodule, an image variation submodule, and a combination submodule; The text mutation submodule is configured to perform text mutation on the descriptive text in the seed item sample under the constraints of the text prompt word corresponding to any seed item sample, the target detection result of the seed item sample, and the text annotation, to obtain mutated descriptive text. The text prompt word includes the text mutation method and the text mutation process parameters. The image mutation submodule is configured to perform visual mutation on the item image in the seed item sample under the constraints of the descriptive text in the seed item sample, the target detection result of the seed item sample, the image annotation and the visual cue words, to obtain a mutated item image, wherein the visual cue words include the generation parameters of the visual mutation. The combination submodule is configured to combine the mutation description text with the corresponding mutation item image to obtain the item sample corresponding to the seed item sample.

[0008] In some embodiments, the text mutation submodule is configured to extract features from the text prompt words, the descriptive text in the seed item sample, the target detection result of the seed item sample, and the text annotation to obtain text features; perform context aggregation on each word in the text features based on a self-attention mechanism to obtain text aggregation features; and perform autoregressive generation based on the text aggregation features to obtain mutated descriptive text.

[0009] In some embodiments, the image mutation submodule is configured to: extract features from the visual cue words, the descriptive text in the seed item sample, the target detection result of the seed item sample, and the visual annotation to obtain second text features; extract features from multiple sub-images of the item image in the seed item sample to obtain an image sequence, the image sequence including the second image features corresponding to the multiple sub-images respectively; fuse the second text features and the image sequence based on an attention mechanism to obtain a fused sequence; perform autoregressive generation based on the fused sequence to obtain a predicted image sequence; and decode the predicted image sequence to obtain the mutated item image.

[0010] In some embodiments, the apparatus further includes a training module; the training module is configured to train a risk control model based on the item samples to obtain an updated risk control model.

[0011] In some embodiments, the apparatus further includes a second acquisition module, a detection module, and an output module; The second acquisition module is configured to acquire multiple samples of missed items from the risk control model in multiple historical periods, wherein the multiple samples of missed items do not conform to the detection standard text but are identified by the risk control model as conforming to the detection standard text; The detection module is configured to detect the updated risk control model based on the multiple missing item samples and obtain a detection result, wherein the detection result indicates whether the multiple missing item samples meet the recognition status of the detection standard text. The output module is configured to output at least a portion of the multiple missing item samples as seed item samples based on the detection results.

[0012] In some embodiments, the detection module is configured to acquire the target detection results and the detection standard text for each missed item sample, wherein the missed item sample includes descriptive text and an image of the item; process the multiple missed item samples and the detection standard text to obtain a label for each missed item sample; generate a predicted item sample corresponding to each missed item sample based on each missed item sample, the target detection results of each missed item sample, and the label, wherein the predicted item sample conforms to the target detection results of the corresponding missed item sample, and the predicted item sample includes at least one of a first predicted item sample and a second predicted item sample, wherein at least one of the descriptive text and the image of the first predicted item sample contains information that conforms to the prediction standard item, and the second predicted item sample includes a variant of the object of interest in the corresponding missed item sample; and input the multiple missed item samples and the predicted item sample into the updated risk control model to obtain the detection results.

[0013] In some embodiments, the apparatus further includes a screening module; the screening module is configured to screen out multiple third item samples from the multiple missing item samples and the predicted item samples according to a preset rule to obtain multiple fourth item samples, wherein the preset rule indicates the screening rules for the multiple missing item samples and the predicted item samples, the multiple third item samples do not meet the preset rule, and the multiple fourth item samples meet the preset rule; adjust the preset rule according to item samples among the multiple third item samples that do not conform to the detection standard text; perform detection on the updated risk control model according to the multiple missing item samples to obtain detection results, including: perform detection on the updated risk control model according to the multiple fourth item samples to obtain detection results; output at least some item samples among the multiple missing item samples as seed item samples according to the detection results, including: output at least some item samples among the multiple fourth item samples as seed item samples according to the detection results.

[0014] In some embodiments, the output module is configured to obtain a fifth item sample based on the detection result, wherein the fifth item sample is a predicted item sample that is identified by the updated risk control model as conforming to the detection standard text; and output a sixth item sample as a seed item sample, wherein the sixth item sample is an item sample in the fifth item sample that does not conform to the detection standard text, and the target detection result of the sixth item sample indicates that the item sample to be generated does not conform to the detection standard text.

[0015] In some embodiments, the apparatus further includes a third acquisition module; the third acquisition module is configured to acquire a seventh item sample with a scene label, the scene label indicating a standard item that the seventh item sample does not conform to in the detection standard text; acquire a target scene label based on the number of seventh item samples corresponding to each scene label, the target scene label being the scene label with the fewest corresponding seventh item samples; acquire multiple eighth item samples related to the seventh item sample corresponding to the target scene label; output the seventh item sample corresponding to the target scene label and the multiple eighth item samples as seed item samples, the eighth item samples conforming to the detection standard text, and the target detection result of the eighth item samples indicating that the item sample to be generated does not conform to the detection standard text.

[0016] In some embodiments, the apparatus further includes a fourth acquisition module; the fourth acquisition module is configured to acquire a ninth item sample, the ninth item sample conforming to the detection standard text and being identified by the risk control model as not conforming to the detection standard text; acquire a tenth item sample related to the ninth item sample; output the ninth item sample and the tenth item sample as seed item samples; the target detection results of the ninth item sample and the tenth item sample indicate that the item sample to be generated conforms to the detection standard text.

[0017] In some embodiments, the device further includes a text prompt word generation module; the text prompt word generation module is configured to generate process parameter values ​​through a parameter sampler based on a text prompt word template, a preset variety of text variation methods, and a preset range of process parameter values ​​for text variation, and to select a first method from the variety of text variation methods; and to add the process parameter values ​​and the first method to the text prompt word template to obtain multiple text prompt words corresponding to each seed item sample.

[0018] In some embodiments, the apparatus further includes a probability enhancement module; the probability enhancement module is configured to obtain, based on the detection result, the miss rate of the risk control model under the same combination of parameters in the same text prompt word, the miss rate indicating the probability that the item sample obtained under the parameter combination is identified as conforming to the detection standard text; and output the parameter combination with the highest miss rate as the target parameter combination to improve the generation probability of the target parameter combination.

[0019] According to another aspect of the present disclosure, an electronic device is provided, the electronic device comprising: one or more processors; a memory for storing processor-executable program code; wherein the processor is configured to execute the program code to implement the method for generating an article sample as described above.

[0020] According to another aspect of the present disclosure, a computer-readable storage medium is provided that, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform the article sample generation method as described above.

[0021] According to another aspect of the present disclosure, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the above-described method for generating an article sample.

[0022] In the technical solution provided by this disclosure, multiple seed item samples and detection standard text are processed to obtain the annotation of each item sample. Based on the annotation, the descriptive text in the item sample and the objects of interest in the item image are clarified, as well as the prediction standard items. Then, when generating the item sample corresponding to the seed item sample based on the annotation, each seed item sample, and the target detection results of each seed item sample, the objects of interest in the descriptive text and item image can be mutated, or the prediction standard items can be integrated into the seed item sample. The resulting item sample contains a richer combination of standard items, realizing the generation of item samples, enriching the form of item samples, increasing the number of item samples, and enabling timely iteration of the risk control model. The risk control model has a low miss rate for items that do not meet the detection standards.

[0023] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0024] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0025] Figure 1 This is a schematic diagram illustrating an implementation environment according to an exemplary embodiment.

[0026] Figure 2 This is a schematic diagram illustrating the generation principle according to an exemplary embodiment.

[0027] Figure 3 This is a flowchart illustrating a method for generating an article sample according to an exemplary embodiment.

[0028] Figure 4 This is a flowchart illustrating another method for generating an article sample according to an exemplary embodiment.

[0029] Figure 5 This is a schematic diagram illustrating several images of objects according to an exemplary embodiment.

[0030] Figure 6 This is a schematic diagram illustrating several images of objects according to an exemplary embodiment.

[0031] Figure 7 This is a schematic diagram illustrating the detection results of a risk control model according to an exemplary embodiment.

[0032] Figure 8 This is a schematic diagram illustrating the interception rate and leakage rate of a risk control model over several consecutive periods, according to an exemplary embodiment.

[0033] Figure 9 This is a possible operation page diagram illustrating an article sample generation method according to an exemplary embodiment.

[0034] Figure 10 This is a possible operation page diagram illustrating an article sample generation method according to an exemplary embodiment.

[0035] Figure 11 This is a block diagram illustrating an apparatus for generating an article sample according to an exemplary embodiment.

[0036] Figure 12 This is a block diagram illustrating a terminal according to an exemplary embodiment.

[0037] Figure 13 This is a block diagram illustrating a server according to an exemplary embodiment. Detailed Implementation

[0038] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0039] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0040] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this disclosure are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0041] AIGC (Artificial Intelligence Generated Content) is an artificial intelligence technology based on large-scale model algorithms that automatically generates multimodal content such as text and images, enabling intelligent and automated content creation and optimization.

[0042] SOP (Standard Operating Procedure): is a unified standard for reviewing product (item) violations on e-commerce platforms (i.e., testing standards, testing standard text), used to regulate the judgment of objects that do not comply with the testing standard text and to define the standardized execution basis for exemption scenarios.

[0043] NLP (Natural Language Processing) is a branch of artificial intelligence technology used to achieve semantic understanding, feature extraction, content rewriting, and variant generation of text content.

[0044] Computer Vision (CV) is a branch of artificial intelligence technology used to achieve image content parsing, visual feature extraction, compliance judgment, and intelligent image generation and editing.

[0045] MA (Machine Audit): This is an automated content review process on e-commerce platforms, based on risk control models and rule strategies, used to identify and block products (items) that do not meet detection standards in real time.

[0046] RvB (Blue Team vs. Red Team): An adversarial testing technology in the field of e-commerce risk control. It tests vulnerabilities in the machine review process by actively constructing attack item samples, and discovers the risk of missed releases in risk control models and strategies (i.e., preset rules).

[0047] Figure 1 This is a schematic diagram illustrating an implementation environment according to an exemplary embodiment. See also... Figure 1 The implementation environment specifically includes: terminal 101 and server 102. Terminal 101 and server 102 can be connected directly or indirectly through wired or wireless communication, which is not limited herein.

[0048] Terminal 101 is at least one of the following devices: smartphone, smartwatch, desktop computer, laptop, and foldable computer. Terminal 101 runs an application that supports methods for generating item samples. Users can log in to this application through Terminal 101 to access the services provided by the application. For example, users can generate item samples through this application on Terminal 101.

[0049] Terminal 101 generally refers to one of a plurality of terminals; this embodiment uses terminal 101 as an example. Those skilled in the art will understand that the number of terminals can be more or less. For example, there may be several terminals, or dozens or hundreds of terminals, or even more. This disclosure does not limit the number of terminals or the type of device.

[0050] Server 102 is at least one of a single server, multiple servers, a cloud computing platform, and a virtualization center. Server 102 provides backend services for applications supporting methods for generating item samples. Server 102 can connect to terminal 101 via a wireless or wired network. Server 102 can provide the terminal with database services related to the item sample generation method, and of course, other types of services. In some embodiments, the number of servers may be more or fewer, and this disclosure does not limit this. Of course, server 102 may also include other functional servers to provide more comprehensive and diversified services.

[0051] In related technologies, one method for obtaining item samples is through manual mining, editing, and annotation. In this method, risk control personnel inspect the platform for items containing content that does not meet the non-detection criteria. They manually screen for items that are considered offensive (i.e., items that do not meet the detection criteria). This involves manually identifying non-compliant content, manually rewriting the text, and manually modifying the images to generate incremental item samples. Algorithm engineers then import these manually generated samples into the training set of the risk control model to complete iterative training and strategy optimization. This method relies entirely on manual operation; all aspects of item sample mining, filtering, non-compliant content location, text and image mutation, and annotation are done manually. For long-tail scenarios (i.e., scenarios with a small number of non-compliant items), it primarily relies on manual collection of corresponding item samples from the entire network, and cannot actively generate relevant item samples. In other words, attacks against non-compliant items can only manually replicate known non-compliant forms, and cannot generate a large number of item samples containing richer non-compliant content. In summary, the main drawbacks of this method include: (1) Extremely low iteration efficiency: The entire process relies on manual operation. The cycle of single-round item sample mining, generation and labeling is weekly, which is far behind the daily attack iteration speed of merchants. It is impossible to achieve synchronous optimization of attack and defense confrontation and is always in a passive defense state.

[0052] (2) Long-tail scene coverage gap: It is impossible for manuals to obtain low-frequency and niche item samples in batches. There is insufficient training data for scarce scenes, resulting in the risk control model having almost zero ability to identify scarce scenes, and there are a lot of governance blind spots.

[0053] (3) Insufficient adversarial capability of item samples: Manually edited item samples can only replicate known forms that do not meet the detection standards. They cannot simulate the diverse and intelligent variant camouflage methods of merchants. The generated item sample scenarios are simple and the disturbances are simple, making them unable to adapt to new adversarial attacks. Therefore, when using the item samples obtained by this method to perform attack and defense verification on the risk control model, the verification effect is poor. Based on the results of this attack and defense verification, the supplementary defense will make the supplementary defense effect of the risk control model poor.

[0054] (4) High labor costs: The generation of large-scale item samples requires a lot of manpower for risk control operations and algorithms. The error rate of manual annotation is high and it is easy to introduce annotation noise, which affects the training accuracy of the risk control model.

[0055] In related technologies, another method for obtaining item samples is a fixed-rule automated item sample generation method, which is a lightweight intelligent item sample generation method for the industry. In this method, based on preset fixed text and image editing rules, the original item samples are subjected to simple perturbation processing to achieve batch generation of item samples. At the text level, it only supports fixed keyword replacement and fixed character deletion; at the image level, it only supports simple size cropping and brightness adjustment. The main disadvantages of this method include: (1) Lack of precision and controllability: Relying solely on mechanical rule disturbances, it is impossible to accurately identify the core objects of interest in the item sample (that is, in the text or image of the item sample, the content that meets the control items in the testing standard and the content that meets the exemption items in the testing standard), which easily generates invalid item samples or loses the core non-control items features, resulting in low item sample labeling accuracy and a large amount of noisy data.

[0056] (2) Poor multimodal adaptability: It cannot realize the intelligent variation generation of text and image dual modes simultaneously. The image editing dimension is single and it cannot simulate the scene-based camouflage and stylized avoidance methods of real object samples.

[0057] (3) No closed-loop iteration capability: It can only complete the generation of a single item sample, cannot perform refined attribution of the results of machine review and omission, cannot feed back to the item sample generation strategy and model training, and cannot form a governance closed loop of continuous optimization.

[0058] (4) Lack of adaptation to long-tail scenarios: It is impossible to generate scarce item samples (low-frequency, niche item samples) based on item samples that meet the detection standards, and it is impossible to solve the data cold start problem in long-tail scenarios.

[0059] To address the aforementioned problems, this disclosure proposes a method for generating item samples.

[0060] This disclosure proposes a method for generating item samples. The method is achieved through the cooperation of a content understanding model and an item sample generation model (including a text generation model and an image generation model). The content understanding model can be a Claude or GPT series model, typically called via an API (Application Programming Interface). The text generation model can be GPT-5, and the image generation model can be Gemini-3-Pro-Image-Preview. Of course, the content understanding model and item sample generation model can also be other models. This disclosure supports custom model selection via command-line parameters. This disclosure does not specifically limit the type of model. The specific generation principle is as follows: Figure 2 As shown, a content understanding model is used to process multiple seed item samples and detection standard text to obtain the annotation of each item sample. Based on each seed item sample, the target detection results and annotations of each seed item sample, an item sample corresponding to each seed item sample is generated by an item sample generation model, thus realizing the generation of item samples and increasing the number of item samples.

[0061] Figure 3 This is a flowchart illustrating a method for generating an article sample according to an exemplary embodiment. See also: Figure 3 The method in this embodiment can be executed by either a terminal or a server, and the method includes the following steps: In step 301, multiple seed item samples, detection standard text, and target detection results for each seed item sample are obtained. The seed item sample includes the item description text and item image. The detection standard text is used to provide standard items for item detection. The target detection results indicate whether the item sample to be generated conforms to the detection standard text.

[0062] In this step, "item" refers to a product on an e-commerce platform. The description text can be the item's title, or other detailed descriptions or attributes. The item image can be a relevant image, such as a display image of the item or a video frame from a video of the item. For example, if the item is a blind box, the description text could be "Trendy trendy toys, stress-relieving lucky blind boxes, unique and high drop rates, 17 Christmas and New Year gifts," where "item 17" is a high-value item, and the item image can be an image of the blind box's appearance. For blind boxes, the control items included in the full text of the detection standards corresponding to the category to which the item belongs are: 1. High-value blind boxes (any of the following types require control): ① Blind boxes that explicitly or implicitly indicate that they contain high-value items; ② Blind boxes containing protective cases for digital electronic items also belong to high-value blind boxes; ③ Selling or displaying boxes / cartons / bags (such as boxes / gift boxes / gifts / packaging boxes / empty boxes / gift bags, etc.) and explicitly or implicitly indicating that they contain high-value items; ④ Selling or displaying boxes / cartons / bags (such as boxes / gift boxes / gifts / packaging boxes / empty boxes / gift bags, etc.) with the text and images indicating that a QR code needs to be scanned; 2. Jewelry and jade blind boxes; 3. Live pet blind boxes; 4. Virtual items (such as games) The exemptions include: 1. Blind boxes containing props, airline tickets, and travel items; 5. Blind boxes containing alcoholic beverages; 6. Points redemption; 7. Blind boxes containing items sold in the form of blind boxes. 2. If the sale does not involve blind box information but simply sells the packaging box / gift box and displays the contents of the box containing high-value items such as mobile phones, and clearly describes that "only empty boxes are sold, without contents, and not selling / containing high-value items," or displays a shipping list that does not include high-value items, then it is not considered a high-value blind box and does not require control. (Note that this exemption only applies to the control requirements of point ③ for high-value blind boxes); 3. Selling or customizing items or services in the original factory packaging of electronic products is not considered a high-value blind box.

[0063] Seed item samples may include scene tags (e.g., scene tags were added through manual review). Each scene tag corresponds to a set of detection standard texts. The detection standard text for each scene tag is a subset of all detection standard texts corresponding to the item's category. Furthermore, the detection standard text for a scene tag is the standard item corresponding to that scene tag. For example, if a blind box item's scene tag is "explicitly or implicitly advertising the inclusion of high-value items," the controlled items in the detection standard text corresponding to that scene tag only include point ① of "1. High-value blind boxes," and the exemption items only include point 1. For example, if a blind box item does not include a scene tag, then its detection standard text is all detection standard texts corresponding to the item's category.

[0064] The target detection result indicates whether the item sample to be generated conforms to the detection standard text, that is, whether the item sample corresponding to the seed item sample conforms to the detection standard text, including conforming to the detection standard text and not conforming to the detection standard text. For example, for the blind box mentioned above, if an item sample in the corresponding category hits point ① of "1. High-value item blind box" in the control item and does not hit any exemption item, then the item sample does not conform to the detection standard text. If an item sample in the corresponding category hits point ③ of "1. High-value item blind box" in the control item and hits point 2 of the exemption item, then the item sample conforms to the detection standard text.

[0065] In step 302, multiple seed item samples and detection standard text are processed by the content understanding model to obtain the annotation of each item sample. The annotation includes at least one of the objects of interest in the descriptive text and the predicted standard items in the item image. The objects of interest include objects that hit the standard items in the detection standard text, and the predicted standard items include the standard items that were not hit in the detection standard text.

[0066] In this step, multiple seed item samples and detection standard text are processed using a content understanding model to obtain the annotation for each item sample. A standard item refers to each item in the detection standard text, such as each item in the control items and each item in the exemption items. For example, "2. Jewelry and Jade Blind Boxes" in the control items is a standard item. The object of interest is the feature in the descriptive text and item image, and the object of interest matches the standard items in the detection standard text. For example, in the case of step 301, the descriptive text for the blind box is "Internet celebrity trendy toys, stress relief, lucky blind boxes, surprises, unique, high drop rate, grab 17 Christmas gifts and New Year's gifts." "Item 17" matches the first item in the control item "1. High-value item blind boxes," so the object of interest includes "Item 17." The predicted standard items include at least some of the standard items that are not matched in the detection standard text. The predicted standard items typically include "in the context of the current item (category, form, description logic)." In the above-mentioned blind boxes, other standard items can be reasonably added. Predicted standard items include predicted control items and predicted exemption items. Taking predicted control items as an example, predicted standard items for the above-mentioned blind boxes may include "② Blind box items such as digital electronic item protective cases also belong to high-value blind boxes", "③ Selling or displaying box / box / bag items and explicitly or implicitly indicating that they contain high-value items", and "④ Selling or displaying box / box / bag items, while the graphic information indicates that scanning is required". For example, for point ④, a QR code and the text "Scan the code to view the hidden item" can be added to the item image to achieve natural integration. The objects of concern include controlled objects and exempted objects. Controlled objects refer to objects in the description text and item images that hit the standard items (control items) in the detection standard text. Exempted objects refer to objects in the description text and item images that hit the standard items (exempt items) in the detection standard text. The annotations obtained through this step provide accurate basis for the subsequent generation process, improve the speed of obtaining annotations, and have smaller errors than the annotations obtained by manual analysis.

[0067] The process of obtaining the label for each item sample is described below. Step 302 above includes steps 3021 to 3024.

[0068] In step 3021, feature extraction is performed on the descriptive text and detection standard text in any seed item sample to obtain the first text feature.

[0069] In this step, the text encoder of the content understanding model performs word segmentation on the descriptive text and the detection standard text, and converts the segmentation results into word embedding features, which is the first text feature.

[0070] In step 3022, feature extraction is performed on the item images in the seed item sample to obtain the first image features.

[0071] In this step, the object image is segmented into multiple sub-images by a visual encoder (such as ViT) of the content understanding model, and the segmentation results are transformed into visual features, which is to obtain the first image features.

[0072] In step 3023, the first text features and the first image features are fused based on an attention mechanism to obtain the first fused features.

[0073] In this step, the first text features and the first image features are fused using the backbone network of the content understanding model (such as Transformer) based on an attention mechanism to obtain the first fused features.

[0074] In step 3024, the first fusion feature is mapped to obtain the label of the seed item sample.

[0075] In this step, the output head of the content understanding model outputs annotations based on the first fusion feature. The annotations include text annotations and image annotations. The text annotations are annotations of the descriptive text, and the image annotations are annotations of the object images.

[0076] For each seed item sample, the detection standard text corresponding to the seed item sample can be matched based on at least one of the following: the scene tag annotation of the seed item sample, whether the seed item sample itself conforms to the detection standard text, and the target detection result. Accordingly, in some embodiments, the method further includes: when the seed item sample includes a scene tag annotation, outputting the detection standard text corresponding to the scene tag annotation as the detection standard text of the seed item sample; when the seed item sample does not include a scene tag annotation and the seed item sample itself does not conform to the detection standard text, outputting all the detection standard texts of the category to which the item to be generated belongs as the detection standard text of the seed item sample; when the seed item sample does not include a scene tag annotation and the seed item sample itself conforms to the detection standard text, the scene tag of the target detection result of the seed item sample that indicates the scene tag of the item to be generated is used as the detection standard text of the seed item sample. In this step, to address the issue of varying scene tag annotations for each seed item sample, automatic detection standard text matching was performed, rather than matching all seed item samples with all detection standard texts for the corresponding item category. For seed item samples with scene tag annotations, when generating corresponding item samples, the model can focus on refined variations of the same scene tag. In other words, the item sample generation model can be forced to mutate within the same type of scene tag, allowing for the batch generation of a massive number of item samples under the same scene tag, covering different expressions, without accidentally generating item samples of other categories. The generated item samples are purer, and the generation efficiency of the content understanding model is improved.

[0077] In some embodiments, feature extraction is performed on the descriptive text and detection standard text in any seed item sample to obtain first text features. This includes: extracting features from the descriptive text, detection standard text, and understanding prompts in any seed item sample to obtain first text features. The understanding prompts indicate that the content understanding model outputs a label for each item sample based on the seed item sample and the detection standard text. Of course, at least one of the detection standard text and the seed item sample can also be embedded in the understanding prompt template to obtain understanding prompts.

[0078] In step 303, based on each seed item sample, the target detection results and annotations of each seed item sample, an item sample corresponding to each seed item sample is generated. The item sample conforms to the target detection results of the corresponding seed item sample. The item sample includes at least one of the first item sample and the second item sample. At least one of the descriptive text and item image of the first item sample contains information that conforms to the prediction criteria. The second item sample includes the variant object of the object of interest in the corresponding seed item sample.

[0079] Considering that the descriptive text and the object image do not conform to the detection standard text, there are fundamental differences in their features. For example, in the descriptive text, the features are vocabulary, semantic patterns (inducing WeChat addition), and syntactic structure (homophonic, pinyin, character splitting), etc. The corresponding variation methods are mainly lexical and syntactic level transformations (such as synonym replacement, character insertion, and traditional character conversion). In the object image, the features are visual elements (shape, etc.), composition (subject position, background), and text overlay (such as "+v to view goods"), etc. The corresponding variation methods are mainly pixel-level, style-level, and layout-level transformations (such as changing color, adding metallic texture, and adjusting text density). The variation spaces of the two types of features are almost orthogonal. Therefore, the descriptive text and object image of the object sample are generated separately. Accordingly, the following describes the generation process of the object sample. The above step 303 includes the following steps 3031 to 3033.

[0080] In step 3031, under the constraints of the text prompt words corresponding to any seed item sample, the target detection results of the seed item sample, and the text annotation, the text generation model of the item sample generation model performs text mutation on the descriptive text in the seed item sample to obtain mutated descriptive text. The text prompt words include the text mutation method and the process parameters of text mutation.

[0081] In some embodiments, step 3031 includes: extracting features from text prompts, descriptive text in seed item samples, target detection results of seed item samples, and text annotations to obtain text features; performing context aggregation on each word in the text features based on a self-attention mechanism to obtain text aggregation features; and performing autoregressive generation based on the text aggregation features to obtain the variant descriptive text. In this step, the variant descriptive text is generated through the above process using a text generation model.

[0082] In step 3032, under the constraints of the descriptive text in the seed item sample, the target detection result of the seed item sample, the image annotation and the visual cue words, the image generation model of the item sample generation model performs visual mutation on the item image in the seed item sample to obtain the mutated item image. The visual cue words include the generation parameters of the visual mutation.

[0083] In some embodiments, step 3032 includes: extracting features from visual cue words, descriptive text in seed item samples, target detection results of seed item samples, and visual annotations to obtain second text features; extracting features from multiple sub-images of item images in seed item samples to obtain an image sequence, the image sequence including second image features corresponding to each of the multiple sub-images; fusing the second text features and image sequence based on an attention mechanism to obtain a fused sequence; performing autoregressive generation based on the fused sequence to obtain a predicted image sequence; and decoding the predicted image sequence to obtain a mutated item image. In this step, the generation of mutated item images is achieved through an image generation model and the above process.

[0084] In step 3033, the mutation description text is combined with the corresponding mutated item image to obtain the item sample corresponding to the seed item sample.

[0085] In steps 3031 to 3033, the descriptive text and item images in the seed item samples are mutated to obtain corresponding mutated descriptive text and mutated item images. The descriptive text and item images can be mutated within their respective mutation spaces, resulting in more realistic and plausible mutated descriptive text and mutated item images. Furthermore, during the generation process, text mutation methods and process parameters are added to the text prompts, and visual mutation generation parameters are added to the visual prompts. This achieves pre-process control of the generation process, rather than post-processing after generation. The generated item samples better meet the preset requirements, have no synthetic traces, and highly resemble real items.

[0086] In the technical solution provided by this disclosure, multiple seed item samples and detection standard text are processed to obtain the annotation of each item sample. Based on the annotation, the descriptive text in the item sample and the objects of interest in the item image are clarified, as well as the prediction standard items. Then, when generating the item sample corresponding to the seed item sample based on the annotation, each seed item sample, and the target detection results of each seed item sample, the objects of interest in the descriptive text and item image can be mutated, or the prediction standard items can be integrated into the seed item sample. The resulting item sample contains a richer combination of standard items, realizing the generation of item samples, enriching the form of item samples, increasing the number of item samples, and enabling timely iteration of the risk control model. The risk control model has a low miss rate for items that do not meet the detection standards.

[0087] Figure 4 This is a flowchart illustrating another method for generating an article sample according to an exemplary embodiment, see [link to flowchart]. Figure 4 The method in this embodiment can be executed by either a terminal or a server, and the method includes the following steps: In step 401, multiple samples of items that were missed in multiple historical periods are obtained by the risk control model. These multiple samples of items do not meet the detection standard text but are identified by the risk control model as meeting the detection standard text.

[0088] In this step, the risk control model can be a large model, such as the Qwen3 or Qwen3.5 series, or other models; this disclosure does not specifically limit this. The risk control model is trained in each historical period, therefore the risk control model is different for each historical period.

[0089] In step 402, the first risk control model is tested based on multiple missing item samples to obtain the test results. The test results indicate whether the multiple missing item samples meet the recognition criteria of the test text.

[0090] In this step, the model parameters of the risk control model were adjusted in the previous cycle to obtain the first risk control model. Each missed item sample includes a description text and an image of the item. The first risk control model is then tested based on multiple missed item samples. This can be done by: inputting multiple missed item samples into the first risk control model and using it to identify the samples, obtaining a result indicating whether each sample meets the detection standard text; or by: generating item samples based on the multiple missed item samples, inputting the generated sample and the multiple missed items into the first risk control model, and using it to identify the samples, obtaining a result indicating whether each sample meets the detection standard text.

[0091] The process of testing the first risk control model is described below. Step 402 above includes steps 4021 to 4026.

[0092] In step 4021, the target detection results and detection standard text of each missing item sample are obtained. The missing item sample includes the item description text and item image. The detection standard text is used to provide the standard items for item detection. The target detection results indicate whether the item sample to be generated meets the detection standard text.

[0093] In this step, the principles of the descriptive text, object image, target detection result, and detection standard text are the same as in the above embodiments, and will not be described in detail here.

[0094] In some embodiments, the method further includes: obtaining at least one of the detection results and generation paths for each missing item sample, wherein the detection results indicate whether the missing item sample meets the detection standard text, and the generation path indicates whether the item sample to be generated is generated by the item sample generation model; if generated by the item sample generation model, the description text of the item sample to be generated is generated by the text generation model, and the item image of the item sample to be generated is generated by the image generation model; if not generated by the item sample generation model, the description text of the item sample to be generated remains unchanged, and the item image of the item sample to be generated is generated by the digital graphics processing module.

[0095] For example, the relevant acquisition items for each missing item sample are shown in Table 1 below. Table 1 is the table of relevant acquisition items for missing item samples (that is, the basic input parameters).

[0096]

[0097] Table 1 In step 4022, multiple missing item samples and detection standard texts are processed by a content understanding model to obtain a label for each missing item sample. The label includes at least one of the objects of interest in the descriptive text and the predicted standard items in the item image. The objects of interest include objects that match the standard items in the detection standard text, and the predicted standard items include the standard items that are not matched in the detection standard text.

[0098] The principle of obtaining the annotation in this step is the same as in the above embodiments, and will not be described in detail here.

[0099] In some embodiments, step 4022 includes: processing multiple missing item samples, detection standard text, and understanding prompts using a content understanding model to obtain a label for each missing item sample. The understanding prompts indicate whether the missing item sample conforms to the detection standard text and generate a corresponding label. For example, if the missing item sample conforms to the detection standard text, the generated label includes at least one of exemption objects and predicted standard items. If the missing item sample does not conform to the detection standard text, the generated label includes at least one of control objects and predicted standard items. The predicted standard items include at least one of predicted control items and predicted exemption items. The predicted control items are control items that were not matched in the detection standard text, and the predicted exemption items are exemption items that were not matched in the detection standard text.

[0100] In some embodiments, at least one of the detection standard text and seed item samples can also be embedded into the understanding prompt word template to obtain the understanding prompt words.

[0101] In step 4023, based on each missed item sample, the target detection results and annotations of each missed item sample, a predicted item sample corresponding to each missed item sample is generated. The predicted item sample conforms to the target detection results of the corresponding missed item sample. The predicted item sample includes at least one of the first predicted item sample and the second predicted item sample. At least one of the descriptive text and item image of the first predicted item sample contains information that conforms to the prediction criteria. The second predicted item sample includes the variant object of the object of interest in the corresponding missed item sample.

[0102] In this step, the principle of generating predicted item samples is the same as in the above-described example of generating item samples, and will not be described in detail here.

[0103] In some embodiments, step 4023 includes: under the constraints of the text prompt word corresponding to any missing item sample, the target detection result of the missing item sample, and the text annotation, performing text mutation on the descriptive text in the missing item sample to obtain mutated descriptive text, wherein the text prompt word includes the text mutation method and the process parameters of the text mutation; under the constraints of the descriptive text in the missing item sample, the target detection result of the missing item sample, the image annotation, and the visual prompt word, performing visual mutation on the item image in the missing item sample to obtain mutated item image, wherein the visual prompt word includes the generation parameters of the visual mutation; and combining the mutated descriptive text with the corresponding mutated item image to obtain the item sample corresponding to the missing item sample. The principle of this step is the same as the principle of obtaining the item sample corresponding to the seed item sample in the above embodiments, and will not be repeated here.

[0104] For a single missed item sample, multiple corresponding text prompts are generated, leading to multiple predicted item samples. Each predicted item sample is generated based on a text prompt. Accordingly, in some embodiments, the method further includes: generating process parameter values ​​using a parameter sampler based on a text prompt template, multiple preset text mutation methods, and a preset range of process parameters for text mutation; and selecting a first method from among the multiple text mutation methods. The process parameter values ​​and the first method are then added to the text prompt template to obtain multiple text prompts corresponding to each missed item sample. Through this step, multiple corresponding text prompts can be generated for each missed item sample. Different text prompts can include different text mutation methods and different process parameters, allowing multiple corresponding descriptive texts to be generated based on the descriptive text of the missed item sample.

[0105] In some embodiments, text mutation methods include seven types: similar-looking replacement, synonym replacement, traditional Chinese character replacement, character insertion, character deletion, pinyin replacement, and synonym replacement. Each mutation method can be used alone or in any combination. The application probability, perturbation intensity, and combination logic of each mutation method can be freely adjusted through command-line parameters. Furthermore, the length threshold of the descriptive text can be customized through parameters to ensure it conforms to the specification of an item's title being no more than 60 characters. The range of parameters for the text mutation process can include the range of the number of objects of interest.

[0106] In some embodiments, the method further includes: obtaining the miss rate of the risk control model under the same text prompt word parameter combination based on the detection results, wherein the miss rate indicates the probability that the item sample obtained under the parameter combination is identified as conforming to the detection standard text; and outputting the parameter combination with the highest miss rate as the target parameter combination to improve the generation probability of the target parameter combination. Through this step, the generation ratio of parameter combinations corresponding to high miss rates is improved, realizing the autonomous iterative optimization of the generation strategy.

[0107] For a single missed item sample, multiple corresponding visual cue words are generated, leading to multiple predicted item samples. Each predicted item sample is generated based on a visual cue word. Accordingly, in some embodiments, the method further includes: generating a value for the generation parameters using a parameter sampler based on a visual cue word template and preset visual variation generation parameters; adding the generated parameter value to the visual cue word template to obtain multiple text cue words corresponding to each missed item sample. Through this step, multiple corresponding visual cue words can be generated for each missed item sample. Different visual cue words can include different generation parameters, allowing multiple corresponding item images to be generated based on the item image of the missed item sample.

[0108] In some embodiments, the generation parameters for visual variations include image generation style, image composition, image text density, image object similarity, number of controlled objects contained in the image, and image editing control parameters. For example, the generation parameters are shown in Table 2 below. The image editing control parameters include basic geometric transformation methods and parameters, color and brightness enhancement parameters, and noise and blur parameters. For example, the image editing control parameters are shown in Table 3 below.

[0109]

[0110] Table 2

[0111] Table 3 Among these features, four levels of image-item similarity can be used to generate gradient item images ranging from those retaining the core structure to those with completely different appearances; the number of controlled objects contained in the image allows control over the display quantity and concealment level of controlled objects, and the generated item samples can cover situations ranging from overtly non-compliant text to covertly controlled objects; the text density in the image can be adjusted to control the interference intensity of text in the item image; these parameters, combined with the image generation style and image composition, can change the visual presentation of the item image; and image editing control parameters enable the generation of two levels of item images to simulate real shooting flaws and low-quality transmission effects.

[0112] The mutations corresponding to the image editing control parameters are relatively simple. Visual mutations of the object image can also be performed using a digital image processing module (without an image generation model). Accordingly, in some embodiments, visual mutation of the object image can be performed either by the digital image processing module based on the image editing control parameters in the visual mutation generation parameters, or by performing visual mutation of the object image through an image generation model. The digital image processing module can run on a terminal or server, and can be of various types, such as OpenCV (Open Source Computer Vision Library, a cross-platform computer vision library) or PIL (Python Imaging Library, a Python image processing library).

[0113] In some embodiments, the method further includes: obtaining the miss rate of the risk control model under the same visual cue word parameter combination based on the detection results, wherein the miss rate indicates the probability that the item sample obtained under the parameter combination is identified as text that meets the detection standard; and outputting the parameter combination with the highest miss rate as the target parameter combination to improve the generation probability of the target parameter combination. Through this step, the generation ratio of parameter combinations corresponding to high miss rates is improved, realizing the autonomous iterative optimization of the generation strategy.

[0114] In the above steps, multiple predicted item samples are generated based on each missed item sample (used to attack the risk control model and detect the risk control model). During this generation process, the intent of the core text that does not meet the detection standard and the object of interest are not lost through the relevant parameters of the above text variation and visual variation. At the same time, all parameters are fully invoked, and the full-level attack methods from basic to advanced are accurately simulated through arbitrary combination of parameters.

[0115] In step 4024, multiple third item samples are screened out from multiple missing item samples and predicted item samples according to preset rules to obtain multiple fourth item samples. The preset rules indicate the screening rules for multiple missing item samples and predicted item samples. Multiple third item samples do not meet the preset rules, while multiple fourth item samples meet the preset rules.

[0116] In this step, considering that the actual process of detecting items on the content platform is as follows: first, the items on the content platform are screened out (also called coarse recall) according to preset rules. The items obtained after screening are input into the risk control model and identified by the risk control model. Furthermore, the predicted item samples generated by the item sample generation model are not 100% item samples that do not meet the detection standard text (approximately 80% to 90% do not meet the detection standard text). Based on this, the embodiments of this disclosure also first screen out the items according to preset rules, and then use the screened item samples (that is, multiple fourth item samples) to detect the risk control model.

[0117] In some embodiments, before step 4024, the method further includes: performing difficulty stratification on multiple missed item samples and predicted item samples. There are various methods for difficulty stratification. For example, features that do not conform to the detection standard text can be extracted from the item samples (descriptive text, item images), and the features can be numerically quantified according to at least one of the feature complexity, concealment level, and mutation depth to obtain a comprehensive score. The higher the comprehensive score, the easier it is for the corresponding item sample to be misidentified; the lower the comprehensive score, the easier it is for the corresponding item sample to be correctly identified. Alternatively, multiple missed item samples and predicted item samples can be input into the current risk control model (i.e., the first risk control model). The current risk control model outputs the violation probability of each item sample (e.g., a score between 0 and 1). The closer the score is to 0 and 1, the easier it is for the corresponding item sample to be correctly identified; a score between 0.4 and 0.6 indicates that the corresponding item sample is more likely to be misidentified. For example, difficulty stratification can be performed through manual review. Of course, other difficulty stratification methods are also possible, and this disclosure does not specifically limit these methods. By stratifying by difficulty, multiple missed and predicted item samples can be divided into several difficulty levels, such as S, A, B, and C. S-level items are more likely to be misidentified, while C-level items are the easiest to correctly identify. Items more prone to misidentification are subject to focused human-machine sampling or manual verification to ensure the accuracy of the annotations.

[0118] In step 4025, the preset rules are adjusted based on the item samples among the multiple third item samples that do not conform to the detection standard text.

[0119] In this step, if there is an item sample among multiple third item samples that does not conform to the detection standard text, the reason for the omission of the item sample that does not conform to the detection standard text is the pre-screening rule. At this time, the preset rule is adjusted. Through this process, the pre-screening rule is iterated and the accuracy of the preset rule is improved.

[0120] In step 4026, multiple fourth item samples are input into the first risk control model to obtain detection results. The detection results indicate whether the first risk control model recognizes the multiple fourth item samples according to the detection standard text.

[0121] In this step, the first risk control model is used to detect whether each fourth item sample meets the detection standard text, and the detection results are obtained. The detection results include item samples that are identified by the first risk control model as meeting the detection standard text and item samples that are identified by the first risk control model as not meeting the detection standard text.

[0122] In some embodiments, the method further includes: inputting multiple fourth item samples into a first risk control model to obtain a scene label for each fourth item sample. The scene label includes a first label and a second label. The first label indicates that the fourth item sample conforms to the detection standard text, and the second label indicates that the fourth item sample does not conform to the control items in the detection standard text. In this step, the risk control model not only identifies whether the fourth item sample conforms to the detection standard text, but also outputs the corresponding scene label. This improves the efficiency of manual review of the detection results and reduces the workload of manually labeling scene tags.

[0123] In some implementations, the first risk control model includes a coarse-ranking model and a fine-ranking model. Accordingly, step 4026 includes: inputting multiple fourth-item samples into the coarse-ranking model to obtain a first detection result, which indicates whether the coarse-ranking model recognizes the multiple fourth-item samples as conforming to the detection standard text; inputting recalled item samples into the fine-ranking model to obtain a second detection result, where the recalled item samples are those recalled by the coarse-ranking model from the multiple fourth-item samples, and the second detection result indicates whether the fine-ranking model recognizes the recalled item samples as conforming to the detection standard text. The coarse-ranking model is used to further filter out item samples that may not conform to the detection standard text, and then the fine-ranking model performs refined recognition, improving detection and recognition efficiency.

[0124] In step 403, based on the detection results, at least some of the fourth item samples are output as seed item samples.

[0125] This step includes several methods, such as outputting item samples that are identified by the first risk control model as conforming to the detection standard text but actually do not conform to the detection standard text as seed item samples, and outputting item samples that are identified by the first risk control model as not conforming to the detection standard text but actually conform to the detection standard text as seed item samples, etc. The following will introduce the two methods mentioned separately: The first method involves obtaining a fifth item sample based on the detection results. The fifth item sample is the fourth item sample that is identified by the first risk control model as conforming to the detection standard text. The sixth item sample is output as a seed item sample. The sixth item sample is the item sample in the fifth item sample that does not conform to the detection standard text. The target detection result of the sixth item sample indicates that the item sample to be generated does not conform to the detection standard text. In this step, since the predicted item samples generated by the item sample generation model are not all 100% of the item samples that do not conform to the detection standard text, and the multiple fourth item samples obtained after filtering by preset rules are also not 100% of the item samples that do not conform to the detection standard text, the fourth item samples identified by the first risk control model as conforming to the detection standard text may include fourth item samples that conform to the detection standard text. At this time, the fourth item samples that do not conform to the detection standard text are filtered out by manual screening and other methods to obtain the sixth item sample. The sixth item sample is the item sample that the first risk control model missed in this period. The reason for the missed release of the sixth item sample is the first risk control model. The sixth item sample is output as the seed item sample. The first risk control model is trained based on the sixth item sample to achieve accurate supplementary defense.

[0126] The second approach involves obtaining a ninth item sample. This ninth item sample conforms to the detection standard text but is identified by the first risk control model as not conforming to the detection standard text. Then, a tenth item sample related to the ninth item sample is obtained. The ninth and tenth item samples are output as seed item samples. The target detection results of the ninth and tenth item samples indicate that the item sample to be generated conforms to the detection standard text. In this step, the ninth item sample is often an error caused by the first risk control model's "oversensitivity" or "feature learning bias," and its sample size is usually small. The "relatedness" in the tenth item sample refers to items belonging to the same category as the ninth item sample or having the same scene label. Outputting the ninth and tenth item samples as seed item samples generates more similar item samples that conform to the detection standard text. This improves the first risk control model's ability to identify these "difficult item samples" (boundary item samples), thereby reducing the false positive rate and increasing the accuracy of the first risk control model's identification.

[0127] The above example, using a first risk control model comprising one model, describes two possible ways to output at least some of the fourth item samples as seed item samples. Both of these cases involve outputting incorrectly identified item samples by the first risk control model as seed item samples. However, the first risk control model may also include two models, namely a coarse-ranking model and a fine-ranking model. The following example uses this case to describe another possible way to output at least some of the fourth item samples as seed item samples: The first method involves obtaining an eleventh item sample based on the detection results. The eleventh item sample is the fourth item sample identified by the coarse-ranking model as conforming to the detection standard text. The twelfth item sample is output as the first seed item sample. The twelfth item sample is the item sample in the eleventh item sample that does not conform to the detection standard text. The target detection result of the twelfth item sample indicates that the item sample to be generated does not conform to the detection standard text. The thirteenth item sample is obtained. The thirteenth item sample is the recalled item sample identified by the fine-ranking model as conforming to the detection standard text. The fourteenth item sample is output as the second seed item sample. The fourteenth item sample is the item sample in the thirteenth item sample that does not conform to the detection standard text. The target detection result of the fourteenth item sample indicates that the item sample to be generated does not conform to the detection standard text.

[0128] The second method involves obtaining a fifteenth item sample, which meets the detection standard text but is identified as not meeting the detection standard text by the coarse-ranking model; obtaining a sixteenth item sample related to the fifteenth item sample, and outputting the fifteenth and sixteenth item samples as the first seed item sample, whose target detection results indicate that the item sample to be generated meets the detection standard text; obtaining a seventeenth item sample, which meets the detection standard text but is identified as not meeting the detection standard text by the fine-ranking model; obtaining an eighteenth item sample related to the seventeenth item sample, and outputting the seventeenth and eighteenth item samples as the second seed item sample, whose target detection results indicate that the item sample to be generated meets the detection standard text. Here, the fifteenth item sample is an item sample from multiple fourth item samples, and the seventeenth item sample is an item sample from multiple recalled item samples.

[0129] Steps 401 to 403 above are the process of detecting the risk control model and preset rules based on multiple samples of missing items, and performing attribution analysis based on the detection results of the risk control model and the screening results of the preset rules, which is the detection stage of the first risk control model. The following is the stage of supplementing the first risk control model based on the results of the detection stage.

[0130] In step 404, multiple seed item samples, detection standard text, and target detection results for each seed item sample are obtained. The seed item sample includes the item description text and item image. The detection standard text is used to provide standard items for item detection. The target detection results indicate whether the item sample to be generated conforms to the detection standard text.

[0131] The implementation principle of this step is the same as that of the above embodiments, and will not be repeated here.

[0132] For all item samples with scene labels (including item samples during the risk control model training phase and seed item samples for each period), some scene labels correspond to a very small number of item samples (i.e., long-tail rare item samples). If only these few item samples are used to train the first risk control model, the model will struggle to learn relevant patterns due to the limited sample size, resulting in poor item recognition capabilities for scene-label-related items. To address this issue, we obtain item samples with text conforming to the detection criteria that are related to the item samples corresponding to the scene label. Based on these item samples, we generate item samples with text that does not conform to the detection criteria, thereby increasing the number of item samples. Accordingly, in some embodiments, the method further includes: acquiring a seventh item sample with a scene label, the scene label indicating the standard item that the seventh item sample does not conform to in the detection standard text; acquiring a target scene label based on the number of seventh item samples corresponding to each scene label, the target scene label being the scene label with the fewest corresponding seventh item samples; acquiring multiple eighth item samples related to the seventh item sample corresponding to the target scene label, outputting the seventh item sample corresponding to the target scene label and the multiple eighth item samples as seed item samples, the eighth item samples conforming to the detection standard text, and the target detection result of the eighth item samples indicating that the item sample to be generated does not conform to the detection standard text. In this step, the eighth item sample can be an item sample belonging to the same category as the seventh item sample corresponding to the target scene label, or an item sample with similar mutation methods, etc. This step increases the number of item samples (that do not conform to the detection standard text) corresponding to the target scene label.

[0133] In some embodiments, when the first risk control model includes a coarse-ranking model and a fine-ranking model, the seventh item sample and multiple eighth item samples are output as the first seed item sample, and the seventh item sample and multiple eighth item samples are also output as the second seed item sample.

[0134] In step 405, the content understanding model processes multiple seed item samples and detection standard text to obtain the annotation for each item sample. The annotation includes at least one of the objects of interest in the descriptive text and the predicted standard items in the item image. The objects of interest include objects that match the standard items in the detection standard text, and the predicted standard items include the standard items that are not matched in the detection standard text.

[0135] The implementation principle of this step is the same as that of the above embodiments, and will not be repeated here.

[0136] In step 406, the item sample generation model generates an item sample corresponding to each seed item sample based on each seed item sample, the target detection results and annotations of each seed item sample. The item sample conforms to the target detection results of the corresponding seed item sample. The item sample includes at least one of the first item sample and the second item sample. At least one of the descriptive text and item image of the first item sample contains information that conforms to the prediction criteria. The second item sample includes the variant object of the object of interest in the corresponding seed item sample.

[0137] The implementation principle of this step is the same as that of the above embodiments, and will not be described in detail here.

[0138] In some embodiments, step 406 includes: under the constraints of the text prompt words corresponding to any seed item sample, the target detection results of the seed item sample, and the text annotation, performing text mutation on the descriptive text in the seed item sample to obtain mutated descriptive text, wherein the text prompt words include the text mutation method and the process parameters of text mutation; under the constraints of the descriptive text in the seed item sample, the target detection results of the seed item sample, the image annotation, and the visual prompt words, performing visual mutation on the item image in the seed item sample to obtain a mutated item image, wherein the visual prompt words include the generation parameters of visual mutation; and combining the mutated descriptive text with the corresponding mutated item image to obtain the item sample corresponding to the seed item sample. The principle of this step is the same as that in the above embodiments, and will not be described in detail here. The generation principle of the text prompt words and the visual prompt words is the same as that in step 4023 above, and will not be described in detail here.

[0139] For example, the following example illustrates the item sample (item sample that does not conform to the detection standard text) generated based on the seed item sample (such as the sixth item sample), as shown in Table 4. Table 4 is a schematic table of the seed item sample and the corresponding item sample.

[0140]

[0141] Table 4 For example, the following provides examples of item samples (item samples conforming to the detection standard text) generated based on seed item samples conforming to the detection standard text. The first item sample in Table 5 below is an item sample identified by the risk control model as not conforming to the detection standard text but actually conforming to it; this is the ninth item sample, and it does not contain any controlled objects. The remaining seed item samples below are all ordinary fishing tools (neither the description text nor the item image of each seed item sample contains any controlled objects), and all conform to the corresponding detection standard text. Their conforming category is (no liability) B-16 Animal Hunting Tools (other categories of animal hunting tools may not conform to the corresponding detection standard text, such as electrofishing tools). Based on the above item samples, under the constraints of the text mutation method and the generation parameters of visual mutation, the corresponding item samples conforming to the detection standard text are obtained, as shown in Table 5. Table 5 is a schematic table of seed item samples and corresponding item samples.

[0142]

[0143] Table 5 For example, when generating a corresponding seed item sample based on the seventh item sample (text that does not meet the detection standard) corresponding to the target scene label, the resulting seed item sample includes the prediction exemption item corresponding to the seventh item sample corresponding to the target scene label. For instance, the description text of the seventh item sample corresponding to the target scene label is "trendy internet celebrity toys, stress relief, lucky blind surprises, unique high drop rate items, 17 Christmas gifts and New Year's gifts", and its item image includes ( Figure 6 8-1 in the middle includes the image of "Item 17" and the text "Item 17 is waiting for you to grab", which hits the control item "① blind box items that explicitly or implicitly indicate that they contain high-value goods" in the corresponding detection standard text. The description text of the generated seed item sample is "trendy internet celebrity toys, stress relief, lucky blind box surprises, unique, high drop rate, grab Christmas gifts and New Year gifts", and the item image of the generated seed item sample ( Figure 6 In example 8-2), the graphic of "Item 17" and the text "Item 17 is waiting for you to grab" were removed, and "Officially licensed doll / figure" and "Solemn declaration: Does not include mobile phones, tablets, or any other digital electronic products" were added. This matched the exemption item in the corresponding detection standard text: "The product type is a legitimate product sold in the market in blind box form." In this case, by implanting the predicted exemption item, it was possible to generate a product sample that meets the detection standard text based on a product sample that does not meet the detection standard text. Accordingly, when generating a product sample that meets the detection standard text from a product sample that meets the detection standard text, predicted control points and corresponding predicted exemption points can be implanted into the product sample of the detection standard text.

[0144] In this step, the item samples generated based on the sixth item sample are similar missing item samples. During the generation process, lightweight and targeted controllable perturbations are performed based on the sixth item sample, fully utilizing all parameters (including those included in the text and visual cue templates). By adjusting the perturbation amplitude and range of the parameters, the similarity between the generated item samples and the original sixth item sample is precisely controlled. Only non-core features are modified, while features that do not meet the detection standard's control items are strictly retained, generating a batch of similar missing item samples. For example, by freely combining text mutation methods and controlling the perturbation intensity of different text mutation methods through command-line parameters, multiple descriptive texts with gradient mutations (from fine-tuning to moderate rewriting) can be generated. For example, by using image item similarity, the similarity between the item and the item indicated by the generated item sample can be adjusted, generating fully gradient similar item images from different angles of the same item to different models of the same category. For example, by using other generation parameters, the degree of mutation of non-core visual features in the item image can be flexibly adjusted. All parameters support custom configuration, and can generate item samples with different similarity levels to the seed item samples according to needs. The generated item samples are used for model adversarial training, which can quickly fill the data gaps of similar missed release scenarios and improve the risk control model's ability to generalize and identify known attack variants.

[0145] In this step, based on the seventh and eighth item samples corresponding to the target scene labels, the generated item samples are long-tail rare item samples. During the generation process, all parameters can be fully invoked. Through the full combination of parameters, a sufficient number and variety of item images can be generated in a traversal manner. Through text prompts and visual prompts, the implantation position and intensity of prediction standard items can be precisely controlled. The generation process strictly aligns with the detection standard text, avoiding noise from manual annotation at the source. Through image generation style and image composition, the generated item images can cover all mainstream item display formats on the content platform. Through image item similarity, full-dimensional item images are generated from the same item to the same type to similar in shape. Through the number of controlled objects contained in the image, item images with different intensities of non-compliance with detection standard text can be generated. By combining image text density and text variation, the diversity of features in descriptive text and item images is enriched, completely solving the cold start problem caused by the small number of item samples in long-tail scenes, and realizing the transformation from "passively collecting item samples" to "actively creating item samples".

[0146] In step 407, the first risk control model is trained based on the item samples to obtain the second risk control model.

[0147] In this step, during the training of the first risk control model based on the item samples, the model parameters of the first risk control model were adjusted to obtain the second risk control model (that is, the updated risk control model).

[0148] In some embodiments, when the first risk control model includes a coarse-ranking model and a fine-ranking model, the coarse-ranking model is trained based on the item samples generated based on the first seed item samples to obtain the updated coarse-ranking model; the fine-ranking model is trained based on the item samples generated based on the second seed item samples to obtain the updated fine-ranking model.

[0149] In step 408, steps 401 to 407 are repeated.

[0150] One repetition of steps 401 to 407 constitutes one detection and remedial cycle. In each cycle, a closed-loop iterative remedial process for the risk control model is formed.

[0151] For example, taking a first risk control model that includes a coarse-grained ranking model and a fine-grained ranking model as an example, after step 402 above, the detection results of the first risk control model are displayed, along with the corresponding attribution results. For instance, referring to... Figure 7 A total of 681 item samples were delivered, 571 were intercepted, and 110 were missed. The main scene tags that were missed were scene tag 1 and scene tag 2.

[0152] The technical solution of this disclosure, based on a large model, generates multimodal controllable AIGC data, and combines this with risk control model detection and refined attribution iteration to form a fully automated risk control model supplementation and defense closed loop. The method mainly includes: sequentially executing multimodal item sample generation, detecting the risk control model, and performing attribution and iterative optimization based on the detection results. The entire process is automated, and the batch generation of item samples is achieved based on the multimodal item sample generation step.

[0153] The technical solution of this disclosure addresses the problems of inefficiency and high cost in traditional manual item sample generation iteration. Through AIGC intelligent generation, the item sample generation cycle is compressed from weekly to hourly, achieving automated attack and defense iteration to match the rapid iteration rhythm of merchant masquerading methods. It solves the problem of low accuracy in risk control model recognition of corresponding scenarios due to the scarcity of training data for long-tail scenarios. By embedding the features corresponding to the prediction standard items into seed item samples and scenario expansion (i.e., combinations of different control objects and prediction standard items (predicted exemption objects) to form different scenarios), it can generate item samples that conform to the detection standard text in batches, filling data gaps. It also solves the problems of insufficient adversarial capabilities and limited scenario diversity in item samples generated by traditional fixed rules. Through multimodal controllable mutation, it accurately simulates masquerading methods, generating highly realistic item samples and improving the anti-attack capability of the risk control model. This disclosure constructs a fully closed-loop mechanism of "item sample generation - machine review link verification - omission attribution iteration," realizing continuous autonomous iteration of item sample generation, model training, and strategy optimization. The method of this disclosure can automatically complete the import of item samples, the parsing of item samples through a content understanding model (output annotation), multimodal generation, machine verification (inputting item samples into the risk control model to obtain detection results), and attribution iteration (training the risk control model based on the detection results), without manual intervention. The single iteration cycle is compressed to the hour level, and the risk of missed releases in the risk control model is gradually reduced during multiple iterations.

[0154] The method in this embodiment of the disclosure generates item samples with good coverage. For example, the item samples (similar missing item samples) generated based on the sixth item sample have a coverage rate of ≥95%, achieving full coverage defense against similar attacks. The rare long-tail item samples (generated based on the seventh and eighth item samples corresponding to the target scene label) have a coverage rate of ≥90%, solving the problem of blind spots in long-tail scene recognition. The coverage rate of similar missing item samples is the proportion of the scene label corresponding to the similar missing item sample to all scene labels of the historical missing item samples. The coverage rate of rare long-tail item samples is the proportion of the scene label corresponding to the rare long-tail item sample to all scene labels corresponding to the target scene label.

[0155] The method described in this embodiment of the present disclosure demonstrates a significant supplementary defense effect on the risk control model in each cycle, as shown in the reference. Figure 8The overall success rate of remediation is ≥80%, meaning that the risk control model can significantly improve the interception rate of attack item samples after iteration. For several consecutive periods, the platform's high-hot leak rate decreased by 83% to 0.19% (a decrease of 83%, or from 1.12% to 0.19%), and the S-class and A-class risks discovered by front-end inspection decreased by 99% to 0.01% (a decrease of 99%). Among them, the remediation success rate refers to the proportion of historically leaked item samples that can be correctly identified and intercepted by training the risk control model. "High-hot" refers to key scenarios (scenario tags) with high traffic, high risk, and high attention on the content platform. The high-hot leak rate is the proportion of item samples that the risk control model misses in high-hot scenarios out of all item samples that do not meet the test standard text. Front-end inspection refers to the sampling inspection of online items by human reviewers or automated tools. S-class and A-class refer to seriously non-compliant items.

[0156] Among them, Figure 8 Figure (a) illustrates the interception status of the risk control model during five consecutive periods. The delivery volume is the number of item samples (all texts that do not meet the detection criteria) input to the risk control model, and the interception volume is the number of item samples (identified as texts that do not meet the detection criteria) intercepted by the risk control model. See Table 6, which shows the interception status of the risk control model. Figure 8 In Figure (b), the missed release amount is the item sample that the risk control model missed (text that does not meet the detection standard but is identified by the risk control model as text that meets the detection standard), see Table 7, Table 7 is the interception table of the risk control model.

[0157]

[0158] Table 6

[0159] Table 7 The method in this embodiment significantly improves the iteration efficiency of the risk control model, compressing the traditional weekly manual sample iteration cycle to the hourly level, improving the overall governance efficiency by more than 10 times, and greatly reducing the cost of human operation and maintenance; the convergence effect of the risk control model is outstanding, and the missed release rate is reduced by more than 50%.

[0160] After the item samples obtained by the method in this embodiment of the disclosure are used to train the risk control model, the changes in the Acc (Accuracy) and P / R (Precision / Recall) of the risk control model are shown in Table 8 below. Table 8 is a table of Acc and P / R before and after the risk control model is trained.

[0161]

[0162] Table 8 The following provides exemplary formats for comprehension cue templates, text cue templates, and visual cue templates.

[0163] For example, the understanding prompt template can be: "role: You are an experienced product security auditing expert, especially skilled at identifying text content that does not meet the inspection standards for products on e-commerce platforms. Your task is to analyze and locate controlled and exempted items based on the provided product information and strictly follow the 'audit standards'."

[0164] Background: On e-commerce platforms, it is necessary to control the items sold by influencers to protect the platform's legitimate rights and interests. This scenario involves strong adversarial behavior, and you need to accurately identify and locate the controlled and exempted items based on the item description text and item image content.

[0165] Goals: Carefully analyze the item description text and item image content, compare them with the audit standards, check all possible controllable objects item by item, and identify whether there are any exempted objects (such as exception clauses in the audit standards) that make the item conform to the inspection standard text.

[0166] input info: Input information: Item description text: {text_data} Item image content: Audit criteria: {audit_sop} Does the text conform to the detection criteria: {input_data_violate} Workflow: 1. If the "Does it meet the detection standard text?" is determined to be "No" (responsible), then directly, clearly, and item by item list all the specific controlled objects included in the item, as well as the exempted objects that need to be added to convert it into a case that meets the detection standard text (not responsible). (For example: Controlled objects included: [content]; Exempted objects to be added: [content]; Controlled objects to be added: [content]).

[0167] 2. If "Does it meet the detection standard text?" is "Yes" (data is not liable), then directly, clearly, and item by item list all included exemption objects, as well as control objects that could potentially lead to misjudgment if the data is modified to be liable data. (For example: Excluded exemption objects: [content]; Control objects to be added: [content]; Exemption objects to be added: [content]).

[0168] Output: Output requirements: List only the points, without adding any prefixes, suffixes, summaries, or explanatory text.

[0169] For example, the text prompt template can be: "role: You are an adversarial testing expert who is proficient in e-commerce platform rules and merchant avoidance strategies, and is especially good at generating corresponding content outputs according to workflow requirements based on existing input information."

[0170] Goals: Based on an item's description text, the reason why it does not meet the detection criteria, the specified mutation method, and whether the data to be generated meets the detection criteria, generate a new item title that conforms to the following logic: Workflow: 1. If the data to be generated is responsible (i.e. does not actually conform to the detection standard text, and the "Does the generated data conform to the detection standard text?" is "No"), then the generated item description text must include and add a clear control object, and must not include any exempt objects.

[0171] 2. If the data to be generated is liability-free (i.e., actually conforms to the detection standard text, and the "Does the generated data not conform to the detection standard text?" is "conforms"), then the generated item description text should not contain and should remove all "can be added control objects" and seemingly non-compliant expressions (potential control objects). However, it should also embed reasonable exemption objects (but should not be directly and explicitly mentioned, and should not include such as limitation instructions, usage markings, disclaimers, etc.) to ensure that it is actually liability-free and to reflect situations that are easily misjudged in reality.

[0172] constraints: 1. Consistency of core item: The core item described in the new item description text must be completely consistent with the original title.

[0173] 2. Adhere to the mutation method: The specified text mutation method must be strictly applied.

[0174] 3. Natural and fluent: The new title should be semantically fluent and conform to the language habits of e-commerce product titles.

[0175] 4. Length limit: The generated text content must be limited to 60 characters.

[0176] input info: Input information: Original product description text: {text_data} Controlled objects and exempted objects: {violation_and_exemption_points} Specified mutation method: {text_edit_method} Does the generated data meet the detection criteria text: {gen_data_violate} output: Output requirements: Output the modified item description text directly. Do not add any prefixes, quotation marks, explanations, or any other irrelevant characters such as "Modified title:".

[0177] For example, a visual cue template can be: "role: You are a professional e-commerce product image generation expert. Your task is to generate a new main image of a product that conforms to specific compliance logic based on instructions."

[0178] Goals: Based on a reference image in the input information and whether it meets the detection criteria text and core requirements, generate an item avatar that meets the generation control parameters.

[0179] Constraints: 1. Category Consistency: The categories of items displayed in the images should meet the requirements of [Image Product Similarity].

[0180] 2. Originality: Generate a completely new image, without directly copying or plagiarizing the composition details and specific product appearance of the reference image.

[0181] workflow: Core generation logic: 1. If the generated data does not meet the detection standard text: The generated image must contain a clear and identifiable controlled object, and must not contain any exempted objects.

[0182] 2. If the generated data conforms to the detection standard text: The generated image should not contain and should remove all [addable controllable objects] as well as visual features (potential controllable objects) that appear to not conform to the detection standard text on the surface. However, it should also embed reasonable exemption objects to ensure that they are actually not responsible and to reflect situations that are easily misjudged in reality.

[0183] input info: Input information: 1. Reference item image 2. Reference item text information: {text_data} 3. Does the generated data meet the detection standard text: {gen_data_violate} 4. Core requirements, including the controllable and exempted entities: {violation_and_exemption_points} 5. Generate control parameters: Image generation / editing style: {image_gen_edit_style} Image composition: {image_composition} Image text density: {image_text_density} Image product similarity: {image_product_similarity} The number of controlled / exempted objects contained in the image: {image_violation_count} style: Visual requirements: The style, image quality, and display format must meet the requirements of [Image Generation / Editing Style] and [Image Composition].

[0184] The images present a realistic and natural photographic effect, with normal lighting and shadows, and no obvious traces of AI synthesis.

[0185] Text layout must conform to the setting of [Image Text Density], and the text content and position of controlled and exempted objects should be reasonably arranged according to the value of [Whether it meets the detection standard text (responsible / not responsible)].

[0186] The text content in the image must be in Chinese and must not contain English or other languages.

[0187] Compliance logic implementation: If the "Does the generated data meet the detection standard text?" is "No" (responsible), the image must clearly show the controlled objects described in "Controlled Objects and Exempted Objects" (the number is controlled by "Number of Controlled Objects in the Image"), and no exempt text, icons or visual hints may appear.

[0188] If the "Does the generated data meet the detection criteria text?" is "Yes" (no liability), the image must not contain the following content: 1. The controlled objects in

Controlled Objects and Exempted Objects

[0189] In some embodiments, the method further includes: responding to a selection operation generated on an item sample (i.e., to...) Figure 9 (See Figure (a) for the selection operation of the general item sample data generation tool). The item sample generation page is displayed, and a new task control is displayed on the item sample generation page. Figure 9In Figure (b), the newly created task control is used to import seed item samples and configure parameters for item sample generation. In response to triggering the newly created task control, the new task page is displayed, showing multiple selectable task types (such as seed item sample import task, item sample generation task, etc.). In response to selecting the seed item sample import task, the data upload entry is displayed on the new task page. In response to the import of seed item samples through the data upload entry, a details page (such as...) is displayed. Figure 10 (Figure (a) in the figure), the details page is used to display the imported seed item samples and to select the content understanding model; in response to the selection operation of the item sample generation task in the task type, the advanced configuration controls (such as) are displayed on the new task page. Figure 10 (Figure (b) in the diagram) responds to a trigger action on the advanced configuration control, displaying the command-line argument edit box (e.g., ...). Figure 10 (Figure (c)) The command-line parameter edit box is used to edit relevant parameters such as the detection results of seed item samples and the target detection results.

[0190] Figure 11 This is a block diagram illustrating an apparatus for generating an article sample according to an exemplary embodiment. See also... Figure 11 The device includes: The first acquisition module 1101 is configured to acquire multiple seed item samples, detection standard text, and target detection results for each seed item sample. The seed item samples include item description text and item images. The detection standard text is used to provide standard items for item detection. The target detection results indicate whether the item sample to be generated conforms to the detection standard text. Processing module 1102 is configured to process the plurality of seed item samples and detection standard text to obtain a label for each item sample. The label includes at least one of the object of interest and the predicted standard item in the description text and the item image. The object of interest includes the object that hits the standard item in the detection standard text, and the predicted standard item includes the standard item that is not hit in the detection standard text. The generation module 1103 is configured to generate an item sample corresponding to each seed item sample based on each seed item sample, the target detection results of each seed item sample, and the annotation. The item sample conforms to the target detection results of the corresponding seed item sample. The item sample includes at least one of a first item sample and a second item sample. At least one of the descriptive text and item image of the first item sample contains information that conforms to the prediction criteria. The second item sample includes a variant of the object of interest in the corresponding seed item sample.

[0191] In some embodiments, the processing module is configured to extract features from the descriptive text and the detection standard text in any seed item sample to obtain a first text feature; extract features from the item image in the seed item sample to obtain a first image feature; fuse the first text feature and the first image feature based on an attention mechanism to obtain a first fused feature; and map the first fused feature to obtain the annotation of the seed item sample.

[0192] In some embodiments, the annotation includes text annotation and image annotation, wherein the text annotation is an annotation of descriptive text, the image annotation is an annotation of an image of an object, and the generation module includes a text variation submodule, an image variation submodule, and a combination submodule; The text mutation submodule is configured to perform text mutation on the descriptive text in the seed item sample under the constraints of the text prompt word corresponding to any seed item sample, the target detection result of the seed item sample, and the text annotation, to obtain mutated descriptive text. The text prompt word includes the text mutation method and the text mutation process parameters. The image mutation submodule is configured to perform visual mutation on the item image in the seed item sample under the constraints of the descriptive text in the seed item sample, the target detection result of the seed item sample, the image annotation and the visual cue words, to obtain a mutated item image, wherein the visual cue words include the generation parameters of the visual mutation. The combination submodule is configured to combine the mutation description text with the corresponding mutation item image to obtain the item sample corresponding to the seed item sample.

[0193] In some embodiments, the text mutation submodule is configured to extract features from the text prompt words, the descriptive text in the seed item sample, the target detection result of the seed item sample, and the text annotation to obtain text features; perform context aggregation on each word in the text features based on a self-attention mechanism to obtain text aggregation features; and perform autoregressive generation based on the text aggregation features to obtain mutated descriptive text.

[0194] In some embodiments, the image mutation submodule is configured to: extract features from the visual cue words, the descriptive text in the seed item sample, the target detection result of the seed item sample, and the visual annotation to obtain second text features; extract features from multiple sub-images of the item image in the seed item sample to obtain an image sequence, the image sequence including the second image features corresponding to the multiple sub-images respectively; fuse the second text features and the image sequence based on an attention mechanism to obtain a fused sequence; perform autoregressive generation based on the fused sequence to obtain a predicted image sequence; and decode the predicted image sequence to obtain the mutated item image.

[0195] In some embodiments, the apparatus further includes a training module; the training module is configured to train a risk control model based on the item samples to obtain an updated risk control model.

[0196] In some embodiments, the apparatus further includes a second acquisition module, a detection module, and an output module; The second acquisition module is configured to acquire multiple samples of missed items from the risk control model in multiple historical periods, wherein the multiple samples of missed items do not conform to the detection standard text but are identified by the risk control model as conforming to the detection standard text; The detection module is configured to detect the updated risk control model based on the multiple missing item samples and obtain a detection result, wherein the detection result indicates whether the multiple missing item samples meet the recognition status of the detection standard text. The output module is configured to output at least a portion of the multiple missing item samples as seed item samples based on the detection results.

[0197] In some embodiments, the detection module is configured to acquire the target detection results and the detection standard text for each missed item sample, wherein the missed item sample includes descriptive text and an image of the item; process the multiple missed item samples and the detection standard text to obtain a label for each missed item sample; generate a predicted item sample corresponding to each missed item sample based on each missed item sample, the target detection results of each missed item sample, and the label, wherein the predicted item sample conforms to the target detection results of the corresponding missed item sample, and the predicted item sample includes at least one of a first predicted item sample and a second predicted item sample, wherein at least one of the descriptive text and the image of the first predicted item sample contains information that conforms to the prediction standard item, and the second predicted item sample includes a variant of the object of interest in the corresponding missed item sample; and input the multiple missed item samples and the predicted item sample into the updated risk control model to obtain the detection results.

[0198] In some embodiments, the apparatus further includes a screening module; the screening module is configured to screen out multiple third item samples from the multiple missing item samples and the predicted item samples according to a preset rule to obtain multiple fourth item samples, wherein the preset rule indicates the screening rules for the multiple missing item samples and the predicted item samples, the multiple third item samples do not meet the preset rule, and the multiple fourth item samples meet the preset rule; adjust the preset rule according to item samples among the multiple third item samples that do not conform to the detection standard text; perform detection on the updated risk control model according to the multiple missing item samples to obtain detection results, including: perform detection on the updated risk control model according to the multiple fourth item samples to obtain detection results; output at least some item samples among the multiple missing item samples as seed item samples according to the detection results, including: output at least some item samples among the multiple fourth item samples as seed item samples according to the detection results.

[0199] In some embodiments, the output module is configured to obtain a fifth item sample based on the detection result, wherein the fifth item sample is a predicted item sample that is identified by the updated risk control model as conforming to the detection standard text; and output a sixth item sample as a seed item sample, wherein the sixth item sample is an item sample in the fifth item sample that does not conform to the detection standard text, and the target detection result of the sixth item sample indicates that the item sample to be generated does not conform to the detection standard text.

[0200] In some embodiments, the apparatus further includes a third acquisition module; the third acquisition module is configured to acquire a seventh item sample with a scene label, the scene label indicating a standard item that the seventh item sample does not conform to in the detection standard text; acquire a target scene label based on the number of seventh item samples corresponding to each scene label, the target scene label being the scene label with the fewest corresponding seventh item samples; acquire multiple eighth item samples related to the seventh item sample corresponding to the target scene label; output the seventh item sample corresponding to the target scene label and the multiple eighth item samples as seed item samples, the eighth item samples conforming to the detection standard text, and the target detection result of the eighth item samples indicating that the item sample to be generated does not conform to the detection standard text.

[0201] In some embodiments, the apparatus further includes a fourth acquisition module; the fourth acquisition module is configured to acquire a ninth item sample, the ninth item sample conforming to the detection standard text and being identified by the risk control model as not conforming to the detection standard text; acquire a tenth item sample related to the ninth item sample; output the ninth item sample and the tenth item sample as seed item samples; the target detection results of the ninth item sample and the tenth item sample indicate that the item sample to be generated conforms to the detection standard text.

[0202] In some embodiments, the device further includes a text prompt word generation module; the text prompt word generation module is configured to generate process parameter values ​​through a parameter sampler based on a text prompt word template, a preset variety of text variation methods, and a preset range of process parameter values ​​for text variation, and to select a first method from the variety of text variation methods; and to add the process parameter values ​​and the first method to the text prompt word template to obtain multiple text prompt words corresponding to each seed item sample.

[0203] In some embodiments, the apparatus further includes a probability enhancement module; the probability enhancement module is configured to obtain, based on the detection result, the miss rate of the risk control model under the same combination of parameters in the same text prompt word, the miss rate indicating the probability that the item sample obtained under the parameter combination is identified as conforming to the detection standard text; and output the parameter combination with the highest miss rate as the target parameter combination to improve the generation probability of the target parameter combination.

[0204] It should be noted that the above embodiments of the item sample generation apparatus are only illustrated by the division of the above functional units when generating item samples. In practical applications, the above functions can be assigned to different functional units as needed, that is, the internal structure of the electronic device can be divided into different functional units to complete all or part of the functions described above. In addition, the item sample generation apparatus and the item sample generation method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0205] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0206] When an electronic device is provided as a terminal, Figure 12This is a block diagram illustrating a terminal according to an exemplary embodiment. The diagram shows a structural block diagram of a terminal 800 provided in an exemplary embodiment of this disclosure. The terminal 800 may be a smartphone, tablet computer, MP3 player (Moving Picture Experts Group Audio Layer III), MP4 player (Moving Picture Experts Group Audio Layer IV), laptop computer, or desktop computer. The terminal 800 may also be referred to as a user device, portable terminal, laptop terminal, desktop terminal, or other names.

[0207] Typically, terminal 1200 includes a processor 1201 and a memory 1202.

[0208] Processor 1201 may include one or more processing cores, such as a quad-core processor, an octa-core processor, etc. Processor 1201 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 1201 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 1201 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 1201 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0209] The memory 1202 may include one or more computer-readable storage media, which may be non-transitory. The memory 1202 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 1202 are used to store at least one computer program, which is executed by the processor 1201 to implement the method provided in the method embodiments of this application.

[0210] In some embodiments, the terminal 1200 may also optionally include a peripheral device interface 1203 and at least one peripheral device. The processor 1201, memory 1202, and peripheral device interface 1203 can be connected via a bus or signal line. Each peripheral device can be connected to the peripheral device interface 1203 via a bus, signal line, or circuit board. Specifically, the peripheral device includes at least one of the following: radio frequency circuitry 1204, display screen 1205, camera assembly 1206, audio circuitry 1207, and power supply 1208.

[0211] Peripheral device interface 1203 can be used to connect at least one I / O (Input / Output) related peripheral device to processor 1201 and memory 1202. In some embodiments, processor 1201, memory 1202 and peripheral device interface 1203 are integrated on the same chip or circuit board; in some other embodiments, any one or two of processor 1201, memory 1202 and peripheral device interface 1203 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0212] The radio frequency (RF) circuit 1204 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The RF circuit 1204 communicates with communication networks and other communication devices via electromagnetic signals. The RF circuit 1204 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals back into electrical signals. In some embodiments, the RF circuit 1204 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, etc. The RF circuit 1204 can communicate with other terminals via at least one wireless communication protocol. This wireless communication protocol includes, but is not limited to: the World Wide Web, metropolitan area networks, intranets, various generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area networks, and / or WiFi (Wireless Fidelity) networks. In some embodiments, the RF circuit 1204 may also include circuitry related to NFC (Near Field Communication), which is not limited in this application.

[0213] Display screen 1205 is used to display a UI (User Interface). This UI may include graphics, text, icons, videos, and any combination thereof. When display screen 1205 is a touch display screen, it also has the ability to collect touch signals on or above its surface. These touch signals can be input as control signals to processor 1201 for processing. In this case, display screen 1205 can also be used to provide virtual buttons and / or a virtual keyboard, also known as soft buttons and / or a soft keyboard. In some embodiments, there may be one display screen 1205, disposed on the front panel of terminal 1200; in other embodiments, there may be at least two display screens, disposed on different surfaces of terminal 1200 or in a folded design; in still other embodiments, display screen 1205 may be a flexible display screen, disposed on a curved or folded surface of terminal 1200. Furthermore, display screen 1205 may also be configured as a non-rectangular, irregular shape, i.e., a non-rectangular screen. The display screen 1205 can be made of materials such as LCD (Liquid Crystal Display) and OLED (Organic Light-Emitting Diode).

[0214] The camera assembly 1206 is used to acquire images or videos. In some embodiments, the camera assembly 1206 includes a front-facing camera and a rear-facing camera. Typically, the front-facing camera is located on the front panel of the terminal, and the rear-facing camera is located on the back of the terminal. In some embodiments, there are at least two rear-facing cameras, which are any one of a main camera, a depth-sensing camera, a wide-angle camera, and a telephoto camera, to achieve background blurring by fusion of the main camera and the depth-sensing camera, panoramic shooting by fusion of the main camera and the wide-angle camera, VR (Virtual Reality) shooting, or other fusion shooting functions. In some embodiments, the camera assembly 1206 may also include a flash. The flash can be a single-color temperature flash or a dual-color temperature flash. A dual-color temperature flash is a combination of a warm-light flash and a cool-light flash, which can be used for light compensation at different color temperatures.

[0215] The audio circuit 1207 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, converting the sound waves into electrical signals that are input to the processor 1201 for processing, or input to the radio frequency circuit 1204 for voice communication. For stereo sound acquisition or noise reduction purposes, multiple microphones may be used, each positioned at a different location on the terminal 1200. The microphone may also be an array microphone or an omnidirectional microphone. The speaker is used to convert electrical signals from the processor 1201 or the radio frequency circuit 1204 into sound waves. The speaker may be a conventional diaphragm speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can convert electrical signals not only into audible sound waves but also into inaudible sound waves for purposes such as distance measurement. In some embodiments, the audio circuit 1207 may also include a headphone jack.

[0216] Power supply 1208 is used to power the various components in terminal 1200. Power supply 1208 can be AC ​​power, DC power, a disposable battery, or a rechargeable battery. When power supply 1208 includes a rechargeable battery, the rechargeable battery can be a wired rechargeable battery or a wireless rechargeable battery. A wired rechargeable battery is a battery that is charged via a wired line, and a wireless rechargeable battery is a battery that is charged via a wireless coil. The rechargeable battery can also be used to support fast charging technology.

[0217] Those skilled in the art will understand that Figure 12 The structure shown does not constitute a limitation on terminal 1200 and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0218] The aforementioned electronic devices can also be implemented as servers. The structure of a server is described below: Figure 13 This is a schematic diagram of a server structure provided in an embodiment of this application. The server 1300 can vary significantly due to different configurations or performance. It may include one or more Central Processing Units (CPUs) 1301 and one or more memories 1302. The one or more memories 1302 store at least one computer program, which is loaded and executed by the one or more processors 1301 to implement the methods provided in the various method embodiments described above. Of course, the server 1300 may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. The server 1300 may also include other components for implementing device functions, which will not be elaborated upon here.

[0219] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including a computer program that can be executed by a processor to perform the methods in the above embodiments. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage device, etc.

[0220] In an exemplary embodiment, a computer program product or computer program is also provided, which includes program code stored in a computer-readable storage medium. A processor of an electronic device reads the program code from the computer-readable storage medium and executes the program code, causing the electronic device to perform the method described above.

[0221] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0222] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A method for generating an item sample, characterized in that, The method includes: The process involves acquiring multiple seed item samples, detection standard text, and target detection results for each seed item sample. The seed item samples include descriptive text and images of the items. The detection standard text provides standard items for item detection. The target detection results indicate whether the item sample to be generated conforms to the detection standard text. The multiple seed item samples and detection standard text are processed to obtain the annotation of each item sample. The annotation includes at least one of the object of interest and the predicted standard item in the description text and item image. The object of interest includes the object that hits the standard item in the detection standard text, and the predicted standard item includes the standard item that is not hit in the detection standard text. Based on each seed item sample, the target detection results of each seed item sample, and the annotation, an item sample corresponding to each seed item sample is generated. The item sample conforms to the target detection results of the corresponding seed item sample. The item sample includes at least one of a first item sample and a second item sample. At least one of the descriptive text and item image of the first item sample contains information that conforms to the prediction criteria. The second item sample includes variant objects of the object of interest in the corresponding seed item sample.

2. The method for generating an article sample according to claim 1, characterized in that, The process of processing the multiple seed item samples and detection standard text to obtain the annotation for each item sample includes: Feature extraction is performed on the descriptive text and the detection standard text in any seed item sample to obtain the first text feature; Feature extraction is performed on the item images in the seed item sample to obtain the first image features; The first text feature and the first image feature are fused based on an attention mechanism to obtain the first fused feature; The first fusion feature is mapped to obtain the label of the seed item sample.

3. The method for generating an article sample according to claim 1, characterized in that, The annotation includes text annotation and image annotation. The text annotation is for descriptive text, and the image annotation is for item images. The process of generating an item sample corresponding to each seed item sample based on each seed item sample, the target detection results of each seed item sample, and the annotation includes: Under the constraints of the text prompt words corresponding to any seed item sample, the target detection results of the seed item sample, and the text annotation, the descriptive text in the seed item sample is subjected to text mutation to obtain mutated descriptive text. The text prompt words include the text mutation method and the process parameters of the text mutation. Under the constraints of the descriptive text in the seed item sample, the target detection result of the seed item sample, the image annotation and visual cue words, visual mutation is performed on the item image in the seed item sample to obtain a mutated item image. The visual cue words include the generation parameters of the visual mutation. The mutation description text is combined with the corresponding mutated item image to obtain the item sample corresponding to the seed item sample.

4. The method for generating an article sample according to claim 3, characterized in that, Under the constraints of the text prompt words corresponding to any seed item sample, the target detection results of the seed item sample, and the text annotation, text mutation is performed on the descriptive text in the seed item sample to obtain mutated descriptive text, including: Text features are obtained by extracting features from the text prompts, the descriptive text in the seed item samples, the target detection results of the seed item samples, and the text annotations. Based on the self-attention mechanism, context aggregation is performed on each word in the text features to obtain text aggregation features; Autoregressive generation is performed based on the text aggregation features to obtain variant descriptive text.

5. The method for generating an article sample according to claim 3, characterized in that, Under the constraints of the descriptive text in the seed item sample, the target detection result of the seed item sample, the image annotation, and the visual cue words, visual mutation is performed on the item images in the seed item sample to obtain mutated item images, including: The visual cue words, the descriptive text in the seed item samples, the target detection results of the seed item samples, and the visual annotations are used to extract features to obtain the second text features; Feature extraction is performed on multiple sub-images of the item image in the seed item sample to obtain an image sequence, wherein the image sequence includes second image features corresponding to the multiple sub-images respectively; The second text features and the image sequence are fused based on an attention mechanism to obtain a fused sequence; Based on the fused sequence, an autoregressive generation is performed to obtain a predicted image sequence; The predicted image sequence is decoded to obtain the image of the mutated item.

6. The method for generating an article sample according to any one of claims 1 to 5, characterized in that, The method includes: Based on the item samples, a risk control model is trained to obtain an updated risk control model.

7. The method for generating an article sample according to claim 6, characterized in that, The method includes: The risk control model obtains multiple samples of missed items in multiple historical periods. These multiple samples of missed items do not conform to the detection standard text but are identified by the risk control model as conforming to the detection standard text. Based on the multiple samples of missing items, the updated risk control model is tested to obtain a test result, which indicates whether the multiple samples of missing items meet the recognition criteria of the test standard text. Based on the detection results, at least a portion of the missing item samples are output as seed item samples.

8. The method for generating an article sample according to claim 7, characterized in that, The updated risk control model is tested based on multiple samples of missed items from multiple historical periods to obtain detection results, including: Obtain the target detection results and the detection standard text for each missing item sample, wherein the missing item sample includes the item description text and the item image; The multiple missing item samples and the detection standard text are processed to obtain the label for each missing item sample; Based on each missed item sample, the target detection results of each missed item sample, and the annotation, a predicted item sample corresponding to each missed item sample is generated. The predicted item sample conforms to the target detection results of the corresponding missed item sample. The predicted item sample includes at least one of a first predicted item sample and a second predicted item sample. At least one of the descriptive text and item image of the first predicted item sample contains information that conforms to the prediction criteria. The second predicted item sample includes the variant object of the object of interest in the corresponding missed item sample. The multiple missed item samples and the predicted item samples are input into the updated risk control model to obtain the detection results.

9. The method for generating an article sample according to claim 8, characterized in that, The method further includes: According to a preset rule, multiple third item samples are screened out from the multiple missing item samples and the predicted item samples to obtain multiple fourth item samples. The preset rule indicates the screening rules for the multiple missing item samples and the predicted item samples. The multiple third item samples do not meet the preset rule, while the multiple fourth item samples meet the preset rule. The preset rules are adjusted based on the item samples among the plurality of third item samples that do not conform to the detection standard text; The updated risk control model is then tested based on the multiple samples of missed items to obtain the test results, including: The updated risk control model is tested based on the multiple fourth item samples to obtain the test results; The step of outputting at least a portion of the multiple missed item samples as seed item samples based on the detection results includes: Based on the detection results, at least a portion of the fourth item samples are output as seed item samples.

10. The method for generating an article sample according to claim 9, characterized in that, The step of outputting at least a portion of the fourth item samples as seed item samples based on the detection results includes: Based on the detection results, a fifth item sample is obtained, which is a predicted item sample that is identified by the updated risk control model as conforming to the detection standard text. The sixth item sample is output as a seed item sample. The sixth item sample is the item sample in the fifth item sample that does not conform to the detection standard text. The target detection result of the sixth item sample indicates that the item sample to be generated does not conform to the detection standard text.

11. The method for generating an article sample according to claim 1, characterized in that, The method further includes: Obtain a seventh item sample with a scene label, the scene label indicating the standard item that the seventh item sample does not conform to in the detection standard text; Based on the number of seventh item samples corresponding to each scene label, obtain the target scene label, which is the scene label with the fewest corresponding seventh item samples; Obtain multiple eighth item samples related to the seventh item sample corresponding to the target scene label, and output the seventh item sample corresponding to the target scene label and the multiple eighth item samples as seed item samples. The eighth item samples conform to the detection standard text, and the target detection result of the eighth item samples indicates that the item sample to be generated does not conform to the detection standard text.

12. The method for generating an article sample according to claim 7, characterized in that, The method further includes: Obtain a ninth item sample, wherein the ninth item sample conforms to the detection standard text but is identified by the risk control model as not conforming to the detection standard text; Obtain the tenth item sample related to the ninth item sample, and output the ninth item sample and the tenth item sample as seed item samples. The target detection results of the ninth item sample and the tenth item sample indicate that the item sample to be generated conforms to the detection standard text.

13. The method for generating an article sample according to claim 7, characterized in that, The method further includes: Based on text prompt word templates, preset multiple text variation methods, and preset text variation process parameter ranges, the process parameter values ​​are generated through a parameter sampler, and a first method is selected from the multiple text variation methods. The numerical values ​​of the process parameters and the first method are added to the text prompt template to obtain multiple text prompts corresponding to each seed item sample.

14. The method for generating an article sample according to claim 13, characterized in that, The method further includes: Based on the detection results, the leakage rate of the risk control model under the parameter combination in the same text prompt word is obtained. The leakage rate indicates the probability that the item sample obtained under the parameter combination is identified as conforming to the detection standard text. The parameter combination with the highest leakage rate is output as the target parameter combination, thereby increasing the generation probability of the target parameter combination.

15. An apparatus for generating an article sample, characterized in that, The device includes: The first acquisition module is configured to acquire multiple seed item samples, detection standard text, and target detection results for each seed item sample. The seed item samples include item description text and item images. The detection standard text is used to provide standard items for item detection. The target detection results indicate whether the item sample to be generated conforms to the detection standard text. The processing module is configured to process the plurality of seed item samples and detection standard text to obtain a label for each item sample. The label includes at least one of the object of interest and the predicted standard item in the description text and the item image. The object of interest includes the object that hits the standard item in the detection standard text, and the predicted standard item includes the standard item that is not hit in the detection standard text. The generation module is configured to generate an item sample corresponding to each seed item sample based on each seed item sample, the target detection results of each seed item sample, and the annotation. The item sample conforms to the target detection results of the corresponding seed item sample. The item sample includes at least one of a first item sample and a second item sample. At least one of the descriptive text and item image of the first item sample contains information that conforms to the prediction criteria. The second item sample includes a variant of the object of interest in the corresponding seed item sample.

16. An electronic device, characterized in that, The electronic device includes: One or more processors; Memory used to store the executable program code of the processor; The processor is configured to execute the program code to implement the method for generating an article sample as described in any one of claims 1 to 14.

17. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is able to perform the method for generating an article sample as described in any one of claims 1 to 14.

18. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method for generating an article sample as described in any one of claims 1 to 14.

Images