Risk information identification method and device for target object

Abstract

The invention discloses a risk information identification method, system and device for a target object, electronic equipment and a computer storage medium. The method comprises the steps of obtaining description information for the target object; inputting the description information into a preset completion model to obtain information after completion processing of the description information, and taking the information as first information corresponding to the target object; inputting the description information into a preset information abstract model, and extracting key information in the description information as second information corresponding to the target object; performing information aggregation processing on the information processed in the previous step to obtain aggregation information corresponding to the target object; and providing the aggregated information to a preset risk information identification large model in a predetermined task mode to obtain a risk information identification result for the target object. According to the method, the target object risk information identification efficiency and accuracy are improved, and meanwhile, the cost is reduced.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method, device, electronic device, and computer storage medium for identifying risk information of a target object. Background Art

[0002] With the rapid development of e-commerce and the rise of major e-commerce platforms, the amount of user-generated content (UGC) has exploded. UGC includes not only product information posted by sellers, but also posts, comments, and video sharing by buyers on e-commerce platforms. However, this explosive growth of UGC has also brought some security risks to e-commerce platforms, such as the spread of false information and the prevalence of hate speech on the platforms. These issues not only damage the user experience but may also violate laws and regulations, posing legal risks to the platforms.

[0003] At present, faced with massive amounts of data and complex and ever-changing content formats, the platform mainly relies on manual review of content (such as product information) to identify risk information. However, traditional manual review methods can no longer meet the needs. On the one hand, manual review is not only inefficient but also costly. On the other hand, due to different personal judgment standards during manual review, it is easy for the same product to be reviewed by different reviewers but with inconsistent review results. Moreover, during manual review, faced with complex review information, reviewers are prone to errors such as fatigue, omissions, and misjudgment. Long working hours will also lead to a decline in review quality.

[0004] Therefore, how to improve the efficiency and accuracy of identifying commodity risk information while reducing costs is an urgent problem that needs to be solved. Summary of the Invention

[0005] The embodiment of the present application provides a method for identifying risk information of a target object, so as to improve the efficiency and accuracy of identifying risk information of the target object while reducing costs.

[0006] An embodiment of the present application provides a method for identifying risk information for a target object, including: obtaining descriptive information for the target object; inputting the descriptive information into a preset completion model, obtaining information after completing the descriptive information, as first information corresponding to the target object; inputting the descriptive information into a preset information summary model, extracting key information from the descriptive information, as second information corresponding to the target object; performing information aggregation processing on the information processed in the preceding step, obtaining aggregated information corresponding to the target object; the information processed in the preceding step includes the first information and the second information; providing the aggregated information to a preset risk information identification large model in a predetermined task mode, to obtain a risk information identification result for the target object.

[0007] Optionally, after obtaining the descriptive information for the target object and before executing other subsequent steps, the method also includes: inputting the descriptive information into a preset information distribution model, preliminarily identifying possible risk points existing in the target object, and judging whether the possible risk points are risk points that can be identified with high accuracy by the large model: if so, providing the descriptive information of the target object to subsequent steps; if not, providing the descriptive information of the target object to a manual review link to perform a manual review operation.

[0008] Optionally, the information distribution model further performs the following steps: determining the type of possible risk points existing in the target object, distributing the description information of the target object to corresponding channels corresponding to the type, and performing the other subsequent steps.

[0009] Optionally, the preset completion model includes at least one special model selected according to the components of the descriptive information, and each of the special models corresponds to the components of the processing description information that are suitable for processing by the special model; the information formed after each special model processes the components that belong to the special model constitutes the information after completion processing.

[0010] Optionally, the dedicated model includes one or more of the following types: a graphic object recognition model, a text recognition model, and a face recognition model.

[0011] Optionally, before extracting the key information from the description information, the following steps are performed on the description information: based on the content of the description information, one or more information groups with a preset structure are extracted and constructed; the information group serves as an information unit for executing the step of extracting the key information from the description information.

[0012] Optionally, the extracting of key information from the descriptive information includes performing the following steps on the information group: obtaining a risk probability assessment value for each information group through risk assessment; sorting the risk probability of each information group according to the risk probability assessment value of each information group; and based on the sorting result, using a preset number of information groups with high risk probability as the second information corresponding to the target object.

[0013] Optionally, the extracting of key information from the descriptive information includes performing the following steps on the information group: obtaining a risk probability assessment value for each information group through risk assessment; and taking the information group whose risk probability assessment value is higher than a preset risk threshold as the second information corresponding to the target object.

[0014] Optionally, also include:

[0015] The description information is provided to a preset retrieval enhancement generation knowledge base, and retrieval enhancement information related to the description information is retrieved as the third information corresponding to the target object; the third information is provided as part of the aggregated information provided to the risk information identification model.

[0016] Optionally, the initial version of the risk information identification large model is adjusted in the following manner: a general large model is selected as the large model to be adjusted; a training sample set is constructed using accumulated historical manual review data; and the training sample set is used to supervise and adjust the large model to be adjusted to obtain the risk information identification large model.

[0017] Optionally, for the risk information identification big model, the model update of the first update method is performed in the following manner: the parameters of the risk information identification big model used online are used as initial parameters to form the big model to be updated; a predetermined number of historical training samples are selected from the historical training sample set, and a predetermined number of data from the data newly added during use are selected as new training samples, and the two are combined into a first new training sample set; the first new training sample set is used to supervise and adjust the big model to be updated to obtain the model update of the first update method.

[0018] Optionally, for the risk information identification big model, the model update of the second update method is performed in the following manner: a general big model is selected as the big model to be updated; a predetermined number of historical training samples are selected from the historical training sample set, and a predetermined number of data are selected from the data newly added during use as new training samples, and the two are combined into a second new training sample set; the second new training sample set is used to supervise and adjust the big model to be updated to obtain the model update of the second update method.

[0019] Optionally, for the risk information identification large model, the first updating method and the second updating method are used to update the model:

[0020] The first updating method includes: using the parameters of the risk information identification large model used online as initial parameters to form the large model to be updated;

[0021] Selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and combining the two into a first new training sample set;

[0022] Using the first new training sample set, supervise and adjust the large model to be updated to obtain a model update using a first updating method;

[0023] The second updating method includes: selecting a common large model as the large model to be updated;

[0024] Selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and combining the two to form a second new training sample set;

[0025] Using the second new training sample set, supervise and adjust the large model to be updated to obtain a model update using a second updating method;

[0026] The update period of the first update mode is set to a first period, the update period of the second update mode is set to a second period, and the length of the first period is shorter than the second period.

[0027] Optionally, the risk information identification result of the target object includes one or more of the following information: whether there is a risk, specific risk warning information, and risk type information.

[0028] Optionally, the risk information identification result of the target object includes an identification result that the risk cannot be accurately identified.

[0029] Optionally, when the risk information identification result is that the risk cannot be accurately identified, the following output is provided: extracting elements that help to manually identify the risks of the descriptive information of the target object, and outputting them in a form that is convenient for manual identification.

[0030] Optionally, the elements that help to manually identify risks in the description information of the target object include one of the following information: direct information elements for risk identification, and knowledge related to the identification of potential risks.

[0031] An embodiment of the present application also provides a risk information identification system for a target object, including: a client and a server; the server includes: a completion model, an information summary model and a risk information identification model; the client is used to obtain descriptive information for the target object and send the descriptive information to the server; the server is used to obtain the descriptive information, input the descriptive information into a preset completion model, and obtain information after the descriptive information is completed as first information corresponding to the target object; input the descriptive information into a preset information summary model, extract key information from the descriptive information as second information corresponding to the target object; aggregate the information processed by the previous step to obtain aggregated information corresponding to the target object; the information processed by the previous step includes the first information and the second information; the aggregated information is provided to the preset risk information identification model in a predetermined task mode to obtain a risk information identification result for the target object; the risk information identification result is sent to the client; the client is used to display the risk information identification result.

[0032] Optionally, the server also includes: an information distribution model; after obtaining the descriptive information for the target object and before executing other subsequent steps, the information distribution model is used to: input the descriptive information into the preset information distribution model, preliminarily identify possible risk points existing in the target object, and determine whether the possible risk points are risk points that can be identified with high accuracy by the large model; if so, provide the descriptive information of the target object to subsequent steps; if not, provide the descriptive information of the target object to the manual review link to perform manual review operations.

[0033] Optionally, the information distribution model is further used to: determine the type of possible risk points existing in the target object, distribute the description information of the target object to a corresponding channel corresponding to the type, and execute the other subsequent steps.

[0034] Optionally, the preset completion model includes at least one special model selected according to the components of the descriptive information, and each of the special models corresponds to the components of the processing description information that are suitable for processing by the special model; the information formed after each special model processes the components that belong to the special model constitutes the information after completion processing.

[0035] Optionally, the dedicated model includes one or more of the following types: a graphic object recognition model, a text recognition model, and a face recognition model.

[0036] Optionally, the system further includes an information group construction module, configured to perform the following steps on the description information before extracting key information from the description information:

[0037] Based on the content of the description information, one or more information groups with a preset structure are extracted and constructed; the information groups serve as information units for executing the step of extracting key information from the description information.

[0038] Optionally, extracting key information from the description information includes performing the following steps on the information group:

[0039] Through risk assessment, the risk probability assessment value of each information group is obtained;

[0040] Sort the risk probabilities of each information group according to the risk probability assessment values of each information group;

[0041] According to the sorting result, a preset number of information groups with high risk probability are used as the second information corresponding to the target object.

[0042] Optionally, extracting key information from the description information includes performing the following steps on the information group:

[0043] Through risk assessment, the risk probability assessment value of each information group is obtained;

[0044] The information group whose risk probability evaluation value is higher than the preset risk threshold is used as the second information corresponding to the target object.

[0045] Optionally, the server further includes: retrieving an enhanced generated knowledge base;

[0046] The retrieval enhancement generation knowledge base is used to obtain the description information, and retrieve the retrieval enhancement information related to the description information based on the description information as the third information corresponding to the target object; the third information is provided as part of the aggregated information provided to the risk information identification model.

[0047] Optionally, the initial version of the risk information identification model is adjusted in the following manner:

[0048] Select the general large model as the large model to be adjusted;

[0049] Use the accumulated historical manual review data to construct a training sample set;

[0050] The training sample set is used to perform supervised adjustment on the large model to be adjusted to obtain the risk information identification large model.

[0051] Optionally, for the risk information identification large model, the model update of the first updating method is performed in the following manner:

[0052] Use the risk information used online to identify the parameters of the large model as initial parameters to form the large model to be updated;

[0053] Selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and combining the two into a first new training sample set;

[0054] The first new training sample set is used to perform supervised adjustment on the large model to be updated to obtain a model update using a first updating method.

[0055] Optionally, for the risk information identification large model, the second updating method of the model is updated in the following manner:

[0056] Select the general large model as the large model to be updated;

[0057] Selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and combining the two to form a second new training sample set;

[0058] The second new training sample set is used to perform supervised adjustment on the large model to be updated to obtain a model update using a second updating method.

[0059] Optionally, for the risk information identification large model, the first updating method and the second updating method are used to update the model:

[0060] The first updating method includes: using the parameters of the risk information identification large model used online as initial parameters to form the large model to be updated;

[0061] Selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and combining the two into a first new training sample set;

[0062] Using the first new training sample set, supervise and adjust the large model to be updated to obtain a model update using a first updating method;

[0063] The second updating method includes: selecting a common large model as the large model to be updated;

[0064] Selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and combining the two to form a second new training sample set;

[0065] Using the second new training sample set, supervise and adjust the large model to be updated to obtain a model update using a second updating method;

[0066] The update period of the first update mode is set to a first period, the update period of the second update mode is set to a second period, and the length of the first period is shorter than the second period.

[0067] Optionally, the risk information identification result of the target object includes one or more of the following information: whether there is a risk, specific risk warning information, and risk type information.

[0068] Optionally, the risk information identification result of the target object includes an identification result that the risk cannot be accurately identified.

[0069] Optionally, when the risk information identification result is that the risk cannot be accurately identified, the following output is provided: extracting elements that help to manually identify the risks of the descriptive information of the target object, and outputting them in a form that is convenient for manual identification.

[0070] Optionally, the elements that help to manually identify risks in the description information of the target object include one of the following information: direct information elements for risk identification, and knowledge related to the identification of potential risks.

[0071] The present application also provides a device for identifying risk information of a target object, including:

[0072] A description information obtaining unit, configured to obtain description information for a target object;

[0073] a first information obtaining unit, configured to input the description information into a preset completion model, and obtain information after completing the description information as first information corresponding to the target object;

[0074] A second information obtaining unit is configured to input the description information into a preset information summary model and extract key information from the description information as second information corresponding to the target object;

[0075] an aggregate information obtaining unit, configured to aggregate the information processed in the preceding step to obtain aggregate information corresponding to the target object; the information processed in the preceding step includes the first information and the second information;

[0076] The risk information identification result obtaining unit is used to provide the aggregated information to a preset risk information identification model in a predetermined task mode to obtain the risk information identification result for the target object.

[0077] Optionally, it further includes: an information distribution unit;

[0078] After obtaining the descriptive information for the target object and before executing other subsequent steps, the information distribution unit is used to: input the descriptive information into a preset information distribution model, preliminarily identify possible risk points existing in the target object, and determine whether the possible risk points are risk points that can be identified with high accuracy by the large model; if so, provide the descriptive information of the target object to the subsequent steps; if not, provide the descriptive information of the target object to the manual review link to perform a manual review operation.

[0079] Optionally, the information distribution unit is further configured to:

[0080] Determine the type of the possible risk point in the target object, distribute the description information of the target object to the corresponding channel corresponding to the type, and execute the other subsequent steps.

[0081] Optionally, the preset completion model includes at least one special model selected according to the components of the descriptive information, and each of the special models corresponds to the components of the processing description information that are suitable for processing by the special model; the information formed after each special model processes the components that belong to the special model constitutes the information after completion processing.

[0082] Optionally, the dedicated model includes one or more of the following types: a graphic object recognition model, a text recognition model, and a face recognition model.

[0083] Optionally, before extracting the key information from the description information, the following steps are performed on the description information:

[0084] Based on the content of the description information, one or more information groups with a preset structure are extracted and constructed; the information groups serve as information units for executing the step of extracting key information from the description information.

[0085] Optionally, extracting key information from the description information includes performing the following steps on the information group:

[0086] Through risk assessment, the risk probability assessment value of each information group is obtained;

[0087] Sort the risk probabilities of each information group according to the risk probability assessment values of each information group;

[0088] According to the sorting result, a preset number of information groups with high risk probability are used as the second information corresponding to the target object.

[0089] Optionally, extracting key information from the description information includes performing the following steps on the information group:

[0090] Through risk assessment, the risk probability assessment value of each information group is obtained;

[0091] The information group whose risk probability evaluation value is higher than the preset risk threshold is used as the second information corresponding to the target object.

[0092] Optionally, it further includes: a retrieval enhancement generation unit;

[0093] The retrieval enhancement generation unit is used to provide the description information to a preset retrieval enhancement generation knowledge base, retrieve and obtain retrieval enhancement information related to the description information as the third information corresponding to the target object; the third information is provided as part of the aggregated information provided to the risk information identification model.

[0094] Optionally, the initial version of the risk information identification model is adjusted in the following manner:

[0095] Select the general large model as the large model to be adjusted;

[0096] Use the accumulated historical manual review data to construct a training sample set;

[0097] The training sample set is used to perform supervised adjustment on the large model to be adjusted to obtain the risk information identification large model.

[0098] Optionally, for the risk information identification large model, the model update of the first updating method is performed in the following manner:

[0099] Use the risk information used online to identify the parameters of the large model as initial parameters to form the large model to be updated;

[0100] Selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and combining the two into a first new training sample set;

[0101] The first new training sample set is used to perform supervised adjustment on the large model to be updated to obtain a model update using a first updating method.

[0102] Optionally, for the risk information identification large model, the second updating method of the model is updated in the following manner:

[0103] Select the general large model as the large model to be updated;

[0104] Selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and combining the two to form a second new training sample set;

[0105] The second new training sample set is used to perform supervised adjustment on the large model to be updated to obtain a model update using a second updating method.

[0106] Optionally, for the risk information identification large model, the first updating method and the second updating method are used to update the model:

[0107] The first updating method includes: using the parameters of the risk information identification large model used online as initial parameters to form the large model to be updated;

[0108] Selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and combining the two into a first new training sample set;

[0109] Using the first new training sample set, supervise and adjust the large model to be updated to obtain a model update using a first updating method;

[0110] The second updating method includes: selecting a common large model as the large model to be updated;

[0111] Selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and combining the two to form a second new training sample set;

[0112] Using the second new training sample set, supervise and adjust the large model to be updated to obtain a model update using a second updating method;

[0113] The update period of the first update mode is set to a first period, the update period of the second update mode is set to a second period, and the length of the first period is shorter than the second period.

[0114] Optionally, the risk information identification result of the target object includes one or more of the following information: whether there is a risk, specific risk warning information, and risk type information.

[0115] Optionally, the risk information identification result of the target object includes an identification result that the risk cannot be accurately identified.

[0116] Optionally, when the risk information identification result is that the risk cannot be accurately identified, the following output is provided: extracting elements that help to manually identify the risks of the descriptive information of the target object, and outputting them in a form that is convenient for manual identification.

[0117] Optionally, the elements that help to manually identify risks in the description information of the target object include one of the following information: direct information elements for risk identification, and knowledge related to the identification of potential risks.

[0118] The present application also provides an electronic device, which includes a processor and a memory; the memory stores a computer program, and the processor executes the above method after running the computer program.

[0119] The present application also provides a computer storage medium, wherein the computer storage medium stores a computer program, and after the computer program is run by a processor, the above method is executed.

[0120] An embodiment of the present application provides a method for identifying risk information for a target object, including: obtaining descriptive information for the target object; inputting the descriptive information into a preset completion model, obtaining information after completing the descriptive information, as first information corresponding to the target object; inputting the descriptive information into a preset information summary model, extracting key information from the descriptive information, as second information corresponding to the target object; performing information aggregation processing on the information processed in the preceding step, obtaining aggregated information corresponding to the target object; the information processed in the preceding step includes the first information and the second information; providing the aggregated information to a preset risk information identification large model in a predetermined task mode, to obtain a risk information identification result for the target object.

[0121] The risk information identification method for the target object described in the embodiment of the present application, after obtaining the descriptive information for the target object, can obtain the information after the descriptive information is completed based on the preset completion model, and extract the key information in the descriptive information based on the preset information summary model; then, the information processed in the previous step is aggregated to obtain the aggregated information corresponding to the target object; finally, the aggregated information is provided to the preset risk information identification large model in a predetermined task mode, and then the risk information identification result of the target object is obtained. Since this method performs information completion processing and extracts key information, it can combine the advantages of other models, improve the quality of information provided to the large model for identification, avoid the large model from missing risks due to missing information and ignoring key information, effectively improve the efficiency and accuracy of target object risk information identification, and reduce costs at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0122] Figure 1 This is an overall flow chart of a method for identifying risk information of a target object provided by the first embodiment of the present application;

[0123] Figure 2 This is a flowchart of the information distribution process in a risk information identification method for a target object provided by the first embodiment of the present application;

[0124] Figure 3 This is a flow chart of a method for identifying risk information of a target object provided by the first embodiment of the present application;

[0125] Figure 4 This is a schematic diagram of the element information extraction process in a risk information identification method for a target object provided by the first embodiment of the present application;

[0126] Figure 5 This is a schematic diagram of an information summary extraction process in a risk information identification method for a target object provided in the first embodiment of the present application;

[0127] Figure 6 This is a schematic diagram of a model minor version update process in a risk information identification method for a target object provided by the first embodiment of the present application;

[0128] Figure 7 This is a schematic diagram of a major version update process of a model in a risk information identification method for a target object provided in the first embodiment of the present application;

[0129] Figure 8 This is a schematic diagram of a risk information identification system for a target object provided by the second embodiment of the present application;

[0130] Figure 9 This is a schematic diagram of a risk information identification device for a target object provided by the third embodiment of the present application;

[0131] Figure 10 This is a schematic diagram of an electronic device provided in the fourth embodiment of the present application. DETAILED DESCRIPTION

[0132] The following description sets forth many specific details to facilitate a thorough understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of the present application. Therefore, the present application is not limited to the specific implementations disclosed below.

[0133] First, in order to enable those skilled in the art to better understand the solution of the present application, the specific application scenarios of its embodiments are described in detail below based on a risk information identification method for a target object provided in the present application.

[0134] An embodiment of the present application provides a method for identifying risk information for a target object. After obtaining the descriptive information for the target object, the method can obtain information after completing the descriptive information based on a preset completion model, and extract key information from the descriptive information based on a preset information summary model; then, the information processed in the previous step is aggregated to obtain aggregated information corresponding to the target object; finally, the aggregated information is provided to a preset risk information identification large model in a predetermined task mode, thereby obtaining the risk information identification result of the target object. This method improves the efficiency and accuracy of risk information identification of the target object, while reducing costs.

[0135] Typical application scenarios of the embodiments of the present application, for example, risk information identification of product information of products to be listed on the e-commerce platform; specifically, for a listed product, its product information includes text information such as product title, price, category information, SKU description, SPU attributes, SKU title, seller nickname, store name, and other text information, as well as product main picture, auxiliary picture, SKU picture and other picture information. During the risk review process, it is necessary to combine the product text and picture information and make a judgment on whether the product to be reviewed is in violation of the rules according to the corresponding review standards. If the rules are violated, a corresponding risk classification (or violation classification) is given. In the introduction to the following specific implementation methods, this typical application scenario is used as a specific scenario. At the same time, the method provided by this application can obviously also be applied to other similar risk information identification for different target objects.

[0136] Before describing the specific implementation of the risk information identification method for a target object provided in the embodiment of the present application, the terminology of the embodiment of the present application is first explained.

[0137] Large models are deep learning models with enormous parameters (millions to hundreds of billions) and extensive training data. These models are primarily based on neural network architectures and utilize extensive computing resources for long-term training, enabling them to learn extensive and in-depth data representations and patterns. Typical examples of large models are large language models (such as GPT, BERT, and Qwen), which are capable of handling natural language understanding and generation tasks, including but not limited to text summarization, machine translation, sentiment analysis, and question-answering systems.

[0138] Content review: The process of checking user-generated content (such as text, images, and videos) to ensure that it complies with specific rules, standards, or laws and regulations. This process aims to maintain a healthy, safe, and legal online environment and prevent the spread of harmful information, illegal information, or content that violates platform policies.

[0139] RAG (Retrieval-Augmented Generation) is an AI technology that combines information retrieval and text generation. During the pre-training process of large models, the generated content primarily relies on the knowledge in the pre-training data. However, the information in the pre-training data is inaccurate or not updated in real time, resulting in inaccurate responses generated by the large model. Therefore, maintaining a RAG knowledge base can compensate for the inaccuracy of the content generated by the large model, providing the large model with accurate and up-to-date knowledge, allowing it to provide more accurate responses.

[0140] Information summary: Generally speaking, the descriptive information of the target object targeted in this application that needs to identify risk information is complex and redundant. For example, for the product information that needs to be listed on the e-commerce platform, the descriptive information of a product (i.e., the target object) contains a large amount of text information (seller information, product title, product description) and pictures (main picture, auxiliary picture, SKU picture, wireless picture, three-to-four picture, etc.). Inputting all information into a large-scale language model will cause a waste of computing resources and cover up key information. Therefore, the content is first processed through information summary to extract key information. This process aims to retain the core meaning or key points in the original data while reducing unnecessary redundant information, thereby improving the model processing efficiency and response speed.

[0141] Large model pre-training: The model will be pre-trained on a very large and diverse image and text dataset to learn a wide range of language understanding capabilities.

[0142] Large Model SFT: Usually refers to "Supervised Fine-Tuning" (SFT) or "supervised fine-tuning". This is a training method used to further optimize a model that has been pre-trained with a large amount of data so that it can better adapt to the needs of specific tasks or fields. In this step, the model will learn how to complete specific tasks more accurately based on the information provided in the new dataset, such as sentiment analysis, question-answering systems, etc. Compared to training a brand new model from scratch, using an existing large-scale pre-trained model for supervised fine-tuning can significantly reduce the time and computing resources required, and can improve its performance on specific tasks.

[0143] Token: In large language models (such as GPT and BERT), tokens are the basic units into which text is broken down. These units can be words, characters, or subwords. When the model processes text, it first breaks it down into these small pieces to understand and generate content, just like breaking a sentence into small building blocks to make it easier for computers to learn and use.

[0144] The following combination Figure 1-Figure 2A typical application scenario of the risk information identification method for a target object provided in an embodiment of the present application is described in detail; the application scenario is specifically risk information identification for commodity information.

[0145] See Figure 1 , Figure 1 This is an overall flow chart of a method for identifying risk information for a target object provided by the first embodiment of the present application. In the embodiment of the present application, the target object is a commodity as an example for description. When identifying the risk information of a commodity, step S101 is first executed: obtain the commodity information flow. The commodity information flow is a commodity introduction information composed of specific commodities (generally speaking, the SKU of the commodity is the granularity unit). For each commodity, its information includes the text information and picture information of the commodity, and may also include the video information and sound information of the commodity; for example, the text information of the commodity can be the seller information, the title of the commodity, the description of the commodity, etc., and the picture information of the commodity can be the main picture of the commodity, the auxiliary picture of the commodity, the SKU (Stock Keeping Unit, i.e., the inventory unit) picture, etc. In the process of reviewing the risk information of the commodity, it is necessary to combine different types of information of the commodity, for example, combine text information and picture information.

[0146] After acquiring the product information stream, step S102: Information Distribution is then executed. This information distribution process primarily involves distributing the acquired product information stream through a large information distribution model (Model 1). The large information distribution model performs a preliminary assessment of the potential risk type based on the product information of a specific product. If the product information is considered highly accurate, i.e., the risk is a risk point that the large model can accurately identify, the product information is sent to a subsequent large model for automatic AI assessment. If the product risk is not highly accurate, the product information is directly sent to step S103: Upgrade Review, i.e., manual review. Furthermore, the large information distribution model can further include the following functions: determining the type of possible risk points in the product information and distributing the target object's description information to the corresponding channel for the corresponding type to execute subsequent steps. In other words, the large information distribution model (Model 1) not only determines whether a risk point is suitable for AI review, but also distributes different types of risk points to different large models, allowing for more accurate assessment using the appropriate large model. Of course, the above distribution to different channels can also be omitted, with only a unified large model being used to identify various risk points suitable for large model assessment. Whether this function is available depends on the training and adjustments of the large model. Specific large model channels include, for example, counterfeit and substandard channels and content security channels. Through the information distribution large model, each product's information can be sent to the most appropriate identification channel based on its risk point type, enabling accurate risk assessment and more precise online operations and maintenance, as well as rapid and accurate risk information detection and timely processing.

[0147] More specifically, the information distribution model performs information distribution process, see Figure 2 , Figure 2This is a flowchart of the information distribution process in a risk information identification method for a target object provided in the first embodiment of the present application. First, execute step S201: obtain the review data. The review data is the descriptive information of the target object that needs to be identified for risk information. In the scenario of reviewing product information on the e-commerce platform, it is the product information flow with the product SKU as the target object, wherein the target object is a certain product (generally with SKU as the granularity unit), for example, a certain product that needs to be put on the shelf on the e-commerce platform; then, execute step S202: input the above product information flow into the preset information distribution model for risk point prediction. The information distribution model can determine whether the risk points that may exist in the description information of the target object (a certain product) are risk points that can be identified with high accuracy based on the above product information flow. If the prediction result shows that the risk points contained in the description information of the target object are risk points that can be identified with high accuracy by the risk information identification model (that is, they belong to high-accuracy risks), then execute step S203: use the preset risk information identification channel to perform risk judgment on the risk information of the product, and then determine the risk result of the target object; if the prediction result shows that the target object does not belong to high-accuracy risks (that is, it does not belong to risk points that can be identified with high accuracy), then directly execute step S204: upgrade review (corresponding to Figure 1 Step S103 in the above example) directly performs a manual review operation.

[0148] For the target object entering step S203, the preset risk information identification channel can use the identification mode of the combination of the large model and the special model mentioned later to perform further risk identification. After processing by the risk information identification channel, if there are still risk points that cannot be accurately determined, the task is not completed, and step S204 can be entered again to enter the upgrade review link, that is, to perform manual review operations on the risk points that cannot be accurately determined; at this time, the model in the risk information identification channel can be used to mark the risk points that may be involved and cannot be accurately determined for easy manual identification; if after processing by the risk information identification channel, there are no risk points that cannot be accurately determined, the task is completed.

[0149] It should be noted that although the large model has certain general capabilities, in some specific fields such as OCR (Optical Character Recognition) image character recognition, face recognition, feature extraction, etc., the recognition accuracy of the large model is not as good as that of the small model trained for specific tasks; in addition, directly providing the initial description information to the large model may not be an information form suitable for the large model to identify risk information; therefore, in the specific implementation method provided in this application, the description information of the target object will be preprocessed so that the original description information can be transformed into an information form that is convenient for the large model to make risk judgments.

[0150] An important modification of the initial description information is to provide the description information to the information completion model (model 2) for information completion, such as Figure 1 As shown in S104 in .

[0151] Specifically speaking of the scenario of risk information identification based on product information, generally speaking, merchants will provide product structured information of the products to be put on the shelves according to predetermined requirements, which includes text, pictures, etc. The large model used in the embodiment of the present application is generally a model suitable for natural language processing, and the recognition of picture information may be the weakness of these large models. Therefore, for the pictures in the product structured information, a special model suitable for image recognition (i.e., a small model) can be used to interpret the image information, and the information related to risk judgment can be identified and converted into text information, thereby achieving factor completion. These small models may be a series of special models suitable for different situations. For example, through different preset special models, key information such as text content, faces, and element information (such as LOGO) in each picture can be extracted respectively, and then input into the subsequent review model for comprehensive judgment of risk information identification.

[0152] In short, after obtaining the descriptive information for a certain product, execute step S104: information completion, that is, input the descriptive information into the preset completion model (model 2), and obtain the information after the description information is completed as the first information corresponding to the target object. For example, for the image in the product structured information, element information is extracted, and OCR characters, keywords, and image content can be extracted. Through the completion model, the accuracy of subsequent risk information identification can be further improved.

[0153] Another important modification of the initial description information is to provide an information summary suitable for large-scale model risk assessment, such as Figure 1 As shown in S105 in .

[0154] Taking the risk assessment of product information as an example, considering the limitations of large models in terms of cost and effectiveness, if product descriptions are directly input into the large model, since current product descriptions typically contain a large amount of complex introductory content, inputting too much product information into the large model will not only consume a large amount of computing resources, but will also prevent the large model from grasping the key points of the lengthy information input, resulting in performance degradation. Based on this, after obtaining the product description information, step S105: information summary is performed. Specifically, the description information is input into a preset information summary model to obtain information after key information extraction from the description information as the second information corresponding to the target object. For example, the keywords, key images, and key text fragments most relevant to risk point assessment in the product information can be extracted. Compared to further improving the recognition performance of the large model, using a dedicated model for information summary (a small model) has lower computational consumption, shorter response time, and a more balanced effect and performance for simple tasks. Therefore, by transforming the initial description information through the information summary model, the quality of the information provided to the large risk information recognition model is improved. This can further improve the accuracy of subsequent risk information recognition with more reasonable resource consumption, thereby improving recognition efficiency.

[0155] Another important modification to the initial description information is to add relevant RAG knowledge to the initial description information, such as Figure 1 As shown in S106 in .

[0156] Specifically, in certain specific fields, there are various professional terms that large models may not fully understand. Even very large pre-trained models have their knowledge limitations, especially for recent events or very specialized topics. Manually maintained knowledge databases (which can be RAG knowledge bases, that is, knowledge bases formed through artificial intelligence content generation and knowledge retrieval) allow these models to use real-time updated knowledge databases to fill this gap. By accessing external knowledge databases, RAG knowledge can help models generate more accurate, relevant and up-to-date answers. RAG helps reduce inaccurate answers generated by models based on possible biases or erroneous information in their training data. Based on this, after obtaining the descriptive information for the product, step S106 is executed: obtaining RAG knowledge, specifically providing the descriptive information to a preset knowledge database (for example, a retrieval enhancement generation knowledge base), and retrieving retrieval enhancement information related to the descriptive information as the third information corresponding to the target object. For example, RAG information can be audit knowledge, audit standards, and disposal methods, etc. By retrieval enhancement generation knowledge base, the efficiency and accuracy of subsequent risk information identification can be further improved. The retrieval-enhanced generation knowledge base is a knowledge storage and processing system that integrates retrieval technology and generation models. For the embodiments of the present application, external knowledge, especially real-time updated knowledge, can be introduced through this knowledge base, thereby helping the large model to generate more accurate, relevant and up-to-date answers, thereby improving the relevance and practicality of the responses.

[0157] Then, the information processed by the previous steps is aggregated to obtain aggregated information corresponding to the target object. After that, step S107 is executed: audit judgment, specifically, the aggregated information is provided to the preset risk information identification model (model 4) to obtain the message output of the model 4, that is, the risk information identification result for the target object. Generally, the identification result may include the risk determination result and the risk determination reason. In specific implementation, it may also include a probability score for indicating the degree of confidence of the risk information identification model for the risk information identification result, and further determine whether the probability score is greater than the preset probability score threshold. If so, the audit completion operation is performed on the risk information identification result. If not, step S108 is executed: element information highlighting, that is, the element information related to the risk in the aggregated information is marked, for example, the element information is highlighted, and the highlighted element information is provided to the manual auditor for manual audit judgment. In the embodiment of the present application, the use of the risk information identification model for risk information identification not only improves the efficiency and accuracy of risk information identification, but also greatly reduces the audit cost compared to manual audit. For example, by continuously improving the large model, 80% of risk points can be accurately identified, and the remaining risk points are also highlighted, which can significantly reduce the manual work of risk information identification and improve identification efficiency.

[0158] It should be noted that when using models to identify risk information, over time, after the model is deployed and launched, audit standards may change, and new data formats may continue to emerge. These changes pose a significant challenge to the model's recognition capabilities. Without continuous model updates and optimization, its online performance is likely to gradually decline. Therefore, this application utilizes automated iterative model updates. Automatic updates are divided into minor and major version iterations. Minor version iterations enable the model to complete its capabilities in a very short period of time, while major version iterations are primarily intended to correct for recognition capability drift caused by multiple minor version iterations. This automated iterative model update technology can significantly reduce the maintenance cost of online models and ensure that the model's capabilities remain at a high level. In specific implementation, the training samples required for automatic model iteration are composed of samples randomly sampled from the historical training set of the message log library, and newly added risk samples. Using these samples for model training ensures that the resulting model maintains the recognition capabilities of historical data formats while also having good recognition effects on newly added formats.

[0159] The above is a basic description of the embodiment of the present application in combination with typical application scenarios; it should be said that the basic technical solution provided by the embodiment of the present application can also be applied to different risk information identification scenarios. Generally, in different scenarios, the risk information identification method for the target object described in the embodiment of the present application, after obtaining the descriptive information for the target object, can obtain the information after the descriptive information is completed based on the preset completion model, and extract the key information in the descriptive information based on the preset information summary model; then, the information processed in the previous step is aggregated to obtain the aggregated information corresponding to the target object; finally, the aggregated information is provided to the preset risk information identification large model in a predetermined task mode, thereby obtaining the risk information identification result of the target object. This method improves the efficiency and accuracy of risk information identification of the target object, while reducing costs.

[0160] The specific embodiments of the present application are described in detail below through multiple embodiments and drawings.

[0161] First embodiment

[0162] The first embodiment of the present application provides a method for identifying risk information of a target object. Figure 3 The risk information identification method for the target object is described in detail.

[0163] Step S301: Obtain description information for the target object.

[0164] This step is used to obtain descriptive information for the target object. The target object may be an object of risk information to be identified. The descriptive information of the target object is structured information used to describe relevant information of the target object. It may be information used to describe the attributes, value, category, style and other characteristics of the target object. The descriptive information of the target object includes text information and image information of the target object.

[0165] In the embodiment of the present application, the target object is a commodity. The description information of the commodity, which can also be called the commodity information flow, is structured information that describes the relevant information of the commodity, including the text information and image information of the commodity. The text information of the commodity can be the commodity title, price, category information, SKU description, SPU attributes, SKU title, seller nickname, store name and other information. Among them, the category information of the commodity refers to the category to which the commodity belongs. SKU (Stock Keeping Unit), that is, the inventory unit, is a unique identifier used to identify and manage each product or material in the inventory. SKU is usually composed of a combination of letters and numbers, and contains information such as the brand, style, color, and size of the product. SKU description is mainly for the description of a specific SKU (stock keeping unit), which can distinguish different specifications, such as X model mobile phone, 128GB, starlight color; SPU (Standard Product Unit) attributes refer to the attributes of a group of commodities with common attributes. SPU is the smallest unit of commodity information aggregation and is used to describe the characteristics of a product. SPU attributes define the basic information of a product, including but not limited to brand, model, color, size, etc. These attributes help the system classify and retrieve products, ensuring that users can accurately find the products they need.

[0166] Product image information includes the main image, supplementary images, and SKU images. The main image is the primary display image of the product and is usually the first image consumers see. It is used to highlight the product's core selling points. Supplementary images are mainly used to assist the main image in displaying product details, functions, or usage scenarios, providing more information to help consumers fully understand the product. SKU images are mainly used to display images of different SKUs (stock keeping units) of the product, helping consumers distinguish different colors, sizes, or styles.

[0167] After obtaining the descriptive information for the target object, in order to obtain a preliminary identification result of the possible risk points of the target object, an information distribution process is further performed. The information distribution process mainly inputs the descriptive information into a preset information distribution model, preliminarily identifies the possible risk points existing in the target object, and determines whether the possible risk points belong to the risk points that the large model can identify with high accuracy; if so, the descriptive information of the target object is provided to the subsequent steps; if not, the descriptive information of the target object is provided to the manual review link to perform the manual review operation.

[0168] It should be noted that for the product information flow input (structured information of the product to be reviewed, including text and images), information is first distributed through an information distribution model. The information distribution model makes judgments based on the product information and distributes the product to be reviewed to the corresponding channel. The information distribution model also performs the following steps: determining the type of possible risk points existing in the target object, distributing the description information of the target object to the corresponding channel corresponding to the type, and then performing other subsequent steps. For example, the corresponding channels corresponding to the description information of the target object include counterfeit and shoddy channels, content security channels, etc. Different channels can use corresponding large models that have been adjusted with different supervision. Of course, it is also entirely feasible to use the same large model for different risk types.

[0169] Through the information distribution big model, the description information of the target object can be distributed to the corresponding channel of the corresponding type according to the possible risk points in the information of each product, so that the most appropriate big model can be used for processing, which facilitates online operation and maintenance and quickly discovers risk information and handles it in a timely manner.

[0170] In specific implementation, the commodity information flow is input into the preset information distribution model to obtain the preliminary identification results of the possible risk points of the target object. The information distribution model can determine whether the target object belongs to the risk point that can be identified with high accuracy based on the commodity information flow. If the preliminary identification result shows that the target object belongs to the risk point that can be identified with high accuracy, the preset risk information identification model is used to identify the risk information of the commodity, and then determine the risk result of the target object; if the prediction result shows that the target object does not belong to the risk point that can be identified with high accuracy, the description information of the target object is provided to the manual review link to perform the manual review operation. If multiple channels are set for the subsequent large model to perform risk identification, the description information of the target object can be further distributed to the corresponding channel corresponding to the type of possible risk point existing in the target object determined by the information distribution model, and then other subsequent steps are performed.

[0171] Step S302: inputting the description information into a preset completion model to obtain information after completing the description information as first information corresponding to the target object.

[0172] This step is used to obtain the description information of the target object and further obtain the information after completing the description information based on the description information.

[0173] During specific implementation, after obtaining the descriptive information of the target object, the descriptive information is input into a preset completion model to obtain the information after the description information is completed as the first information corresponding to the target object. It should be noted that although the large model has certain general capabilities, in some specific fields such as OCR (image character recognition), face recognition, and feature extraction, the recognition accuracy of the large model is not as good as that of a dedicated model trained for a specific task, that is, a small model. Therefore, the pictures in the structured information of the product will be completed with features through a series of small models, and the key information such as text content, face, feature information (such as the LOGO on the product) in each picture will be extracted through the preset completion model, and then input into the subsequent review model for comprehensive judgment of risk information identification. Based on this, after obtaining the descriptive information for the product, the element information extraction process is executed. Specifically, the descriptive information is input into the preset completion model to obtain the information after the description information is completed as the first information corresponding to the target object. For example, element information extraction is performed on the pictures in the product structured information, and OCR characters, keywords, and picture content can be extracted. Through the completion model, the accuracy of subsequent risk information identification can be further improved.

[0174] The preset completion model includes at least one special model selected according to the components of the description information, and each of the special models processes the components of the description information that are suitable for processing by the special model; the information formed after each special model processes the components that belong to the special model constitutes the information after completion processing. The special model includes one or more of the following types: graphic object recognition model, text recognition model, and human face recognition model. Among them, the image object recognition model can recognize graphic objects in the description information, the text recognition model can recognize text in the description information, and the human face recognition model can recognize faces in graphic objects in the description information. The description information can be completed through the above-mentioned special models, thereby realizing element completion.

[0175] For specific implementation, see Figure 4 , Figure 4 This is a schematic diagram of the element information extraction process in a risk information identification method for a target object provided in the first embodiment of the present application. Figure 4The product description, including text and images, is displayed. After performing optical character recognition (OCR), object recognition, person recognition, and logo recognition on the product images, corresponding element recognition results are obtained, namely, the OCR, object recognition, person recognition, and logo recognition results. For example, the recognition result for Image 1 is "OCR recognition result: LED bedside lighting; object recognition result: lamp; person recognition result: none; logo recognition result: none." The recognition result for Image 2 is "OCR recognition result: lighting, aluminum alloy,...; object recognition result: lamp; person recognition result: none; logo recognition result: none." The recognition result for Image 3 is "LED bedside lighting, adjustable...; object recognition result: lamp; person recognition result: none; logo recognition result: none."

[0176] Through the above-mentioned element completion process, the information in the target object's description information that is not easy for the risk information identification model to process is transformed into text information that is easy to process. This is equivalent to extracting the originally hidden information content, realizing the completion of information elements, and avoiding the omission of risk points during the identification process.

[0177] The above is a process of inputting the description information into a preset completion model, obtaining the information after the description information is completed, and using it as the first information corresponding to the target object. The following describes a process of inputting the description information into a preset information summary model, extracting the key information in the description information, and using it as the second information corresponding to the target object.

[0178] Step S303: inputting the description information into a preset information summary model, extracting key information from the description information as second information corresponding to the target object.

[0179] This step is used to obtain the description information of the target object and further obtain the information after key information is extracted from the description information based on the description information.

[0180] In a specific implementation, after obtaining the descriptive information of the target object, the descriptive information is input into a preset information summarization model to extract key information from the descriptive information as the second information corresponding to the target object. Prior to extracting the key information from the descriptive information, the following steps are performed on the descriptive information: Based on the content of the descriptive information, one or more information groups having a preset structure are extracted and constructed; the information groups serve as information units for performing the step of extracting the key information from the descriptive information. For example, a single product often has more than 50 images, so it is necessary to construct one or more information groups in the form of "product title + product image + image OCR result". In a specific implementation, a small information structuring model can be used to construct the information groups. Extracting the key information from the descriptive information includes performing the following steps on the information groups: obtaining a risk probability assessment value for each information group through risk assessment; ranking the risk probabilities of the information groups based on the risk probability assessment values of the information groups; and selecting a preset number of information groups with high risk probabilities as the second information corresponding to the target object based on the ranking results. Among them, the above-mentioned process of obtaining the risk probability assessment value of each information group and sorting the risk probability of each information group can be implemented through a factor sorting small model. For example, the above-constructed information group "product title + single product picture + picture OCR result" can be used as the input of the factor sorting small model. The small model outputs the corresponding risk score. By sorting the scores of all pictures, the TOP-N pictures with the highest scores are taken as the second information corresponding to the product, and given to the subsequent review model for comprehensive judgment.

[0181] As another embodiment, the extracting of key information from the descriptive information includes performing the following steps on the information group: obtaining a risk probability assessment value for each information group through risk assessment; and using the information group with a risk probability assessment value higher than a preset risk threshold as the second information corresponding to the target object. In specific implementation, after obtaining the risk probability assessment value for each information group, a pre-set risk threshold can also be used to use the information group with a risk probability assessment value higher than the preset risk threshold as the second information corresponding to the target object. It should be noted that the reason why the information summary model is set in the embodiment of the present application is mainly because after obtaining the descriptive information of the target object, the large model is limited by cost and effect considerations. If the descriptive information of the target object is directly input into the large model, the input of too much product information into the large model will not only cause a large amount of computing resources to be consumed, but also make the large model unable to grasp the key points of the lengthy information input, resulting in performance degradation. Compared with the large model, the small model has lower computing consumption and shorter response time, and is more balanced in effect and performance for simple tasks. Based on this, the embodiments of the present application provide an information summary model that can extract key information about the target object, such as keywords, key images, and key text fragments, from the target object's description information. By inputting the key information of the target object into the large model, the consumption of computing resources can be effectively reduced, while also greatly improving the performance of the large model.

[0182] The key information obtained through the information summary information can be processed in different ways.

[0183] First, these key information can be provided to the risk information recognition model together with other information, and by emphasizing them in the task prompts, the risk information recognition model can focus on this information, thereby increasing the probability of the risk information recognition model identifying risk points.

[0184] The second method is to provide only the above-mentioned key information that has been screened to the risk information identification model without providing other information, thereby simplifying the workload of the large model.

[0185] The above two treatment methods each have their own advantages, and one specific method can be adopted according to the situation.

[0186] In addition, the information completed by the completion model in the previous step can also be provided to the information summary model together with the description information, that is, the completed information is also used as the input of the information summary model for extraction to avoid missing information related to risk points.

[0187] For specific implementation, see Figure 5 , Figure 5This is a schematic diagram of the information summary extraction process in a risk information identification method for a target object provided in the first embodiment of the present application. Figure 5 The description information of the product is displayed, including text information and picture information. For example, for multiple pictures, Picture 1 to Picture 6, after the above pictures are input into the information summary model, the information summary model obtains the risk probability assessment of each picture, and the output risk probability assessment result is "the risk value of Picture 1 is 0.98, the risk value of Picture 2 is 0.99, the risk value of Picture 3 is 0.32, the risk value of Picture 4 is 0.31, the risk value of Picture 5 is 0.78, and the risk value of Picture 6 is 0.92". Then, the scores of the above pictures are sorted from high to low through the factor sorting model. The sorted results are: Picture 2, Picture 1, Picture 6, Picture 5, Picture 3, Picture 4. According to the sorting results, a preset number of pictures with high risk probability can be taken as the second information. For example, Picture 2, Picture 1, Picture 6, and Picture 5 can be used as the second information and given to the subsequent review model for comprehensive judgment.

[0188] In an embodiment of the present application, the descriptive information can also be provided to a preset retrieval-augmented generation knowledge base, and retrieval-augmented information related to the descriptive information can be retrieved as third information corresponding to the target object; the third information is provided as part of the aggregated information provided to the risk information identification model. Retrieval-Augmented Generation (RAG) is an artificial intelligence technology that combines information retrieval and text generation. By integrating an external knowledge base into the generation process, RAG helps reduce inaccurate answers generated by the model based on possible biases or erroneous information in its training data, thereby helping the model generate more accurate, relevant and up-to-date answers.

[0189] It should be noted that the reason why a retrieval-enhanced generation knowledge base is set up in the embodiment of the present application is mainly because in the process of pre-training the large model, the generated content mainly depends on the knowledge of the pre-training data. In certain specific fields, there are various professional terms that the large model may not fully understand. Direct access to the latest information in RAG can allow the model to provide answers that are closer to the needs, thereby improving the relevance and practicality of the response. In addition, the information in the pre-training data is not accurate enough or is not updated in real time, resulting in the responses generated by the large model being inaccurate. Even a very large pre-training model has its knowledge limitations, especially for recent events or very specialized topics. Therefore, by maintaining the RAG retrieval library, accurate and up-to-date knowledge can be given to the large model, allowing the large model to make more accurate responses.

[0190] In practice, the process of extracting reinforcement information from the RAG knowledge base is jointly accomplished by units such as the retriever, embedding model, and vector database. Through steps such as input processing, information retrieval, document processing, prompt generation, answer generation, and post-processing, the input information of the large model is reinforced. In specific implementation, the input processing step is first performed. Specifically, the input text (i.e., the information to be reinforced) is processed, including word segmentation, stop word removal, and stemming, to facilitate subsequent retrieval and generation. Then, the input text is converted into a vector representation using an embedding model. The embedding model converts text into a vector representation for efficient similarity calculation. Common models such as BERT, RoBERTa, and GPT use deep learning algorithms to map text into a high-dimensional vector space. Afterwards, the information retrieval process is performed. Retrieval is performed by the retriever, which is responsible for retrieving information related to the input query or question from the external knowledge base. The search method can use keyword search, such as the BM25 algorithm, to search based on the degree of keyword matching; or vector search, such as using embedding models such as BERT and DPR (Dense Passage Retrieval), to convert text into vectors and calculate similarity to find the most relevant documents or fragments. After the input text is converted into a vector representation through the embedding model, the vector representation can be stored in a vector database (Vector Database), which is used to store the vector representation of the text and support efficient vector search. In specific implementation, various efficient index structures (such as inverted index, HNSW graph, KD tree, etc.) can be used to accelerate vector search.

[0191] After the input text is converted into a vector representation using the embedding model, the query vectorization process is performed. Specifically, similarity is calculated between the vector representation of the input text and the vectors in the vector database. Based on the similarity calculation results, the documents or fragments most relevant to the input text are retrieved from the vector database. In actual processing, a document processing step may also be performed, namely, filtering and sorting the retrieved documents or fragments to ensure the most relevant and high-quality content returned. Of course, further document preprocessing may also be required, such as removing redundant information and extracting key information. The prompt generation step is then performed, combining the input text and the retrieved relevant documents or fragments into a context to construct the prompt required by the generative model. Finally, the answer generation step is performed, which inputs the prompt into a generative model (such as a large pre-trained language model like GPT-3 or T5). The generative model combines the context information and produces the final answer or text through a generative algorithm. After the text is generated, it can also be processed to ensure that it is correctly formatted, the language is natural, and there are no errors. For example, this processing process may include text formatting, grammar checking, and information correction.

[0192] In specific implementation, for example, taking the product title "Men's Perfume Long-Lasting Light Fragrance Refreshing Tobacco Natural Flavor 50ML" as an example, the RAG result is "Knowledge algorithm result: mention word: [tobacco], entity name: tobacco product entity type: tobacco, risk label: tobacco product management regulations, review standard: "Related laws and regulations, the state implements strict control over the production, sale and advertising of tobacco products; Internet platforms are not allowed to sell tobacco products (including cigarettes, cigars, tobacco, etc.); any promotional activities involving tobacco products (such as pictures, text, links, etc.) are illegal. For example, taking the product title "LED bedside display cabinet fish tank table lamp, universal clamp lighting work light hose USB energy-saving spotlight," the RAG result is "Knowledge algorithm result: mention word: ['led'], entity name: LED lamp, entity type: lamp category, risk label: LED lamp management regulations, review standard: "LED lamps must have CCC certification (China Compulsory Product Certification), products must be marked with the 'China Energy Efficiency Label' and indicate the energy efficiency level, and should be accompanied by a test report issued by a third-party testing agency."

[0193] By introducing RAG, on the one hand, the accuracy of answers can be improved. By retrieving relevant information from external knowledge bases, the generative model can combine contextual information to generate more accurate answers. On the other hand, it can reduce the phenomenon of hallucinations. When generating text, large models may produce "hallucination" content that lacks factual basis. By introducing external knowledge, RAG can effectively reduce the occurrence of this phenomenon. On the other hand, it can enhance the flexibility of the model. RAG allows large models to dynamically obtain the latest and relevant knowledge during runtime, thereby improving their ability to adapt to new information.

[0194] Step S304: Aggregate the information processed in the previous step to obtain aggregated information corresponding to the target object; the information processed in the previous step includes the first information and the second information.

[0195] This step is used to aggregate the information processed in the previous steps to obtain aggregated information corresponding to the target object.

[0196] In an embodiment of the present application, the information processed in the preceding step includes not only the first information and the second information, but also includes information other than the first information and the second information, for example, the retrieval enhancement information obtained by searching the retrieval enhancement generation knowledge base, and the above information is subjected to information aggregation processing. The information aggregation processing specifically includes completing the description information of the target object through the completion model, extracting key information from the description information of the target object through the information summary model, and providing the description information of the target object to a preset retrieval enhancement generation knowledge base, and aggregating the retrieval enhancement information related to the description information to form aggregated information of the target object.

[0197] Step S305: providing the aggregated information to a preset risk information identification model in a predetermined task mode to obtain a risk information identification result for the target object.

[0198] This step is used to obtain the risk information identification result for the target object. Specifically, the aggregated information is provided to a preset risk information identification model in a predetermined task mode, and the risk information identification model outputs the risk information identification result of the target object.

[0199] In the embodiment of the present application, the initial version of the risk information identification large model is adjusted and obtained in the following manner: a general large model is selected as the large model to be adjusted; a training sample set is constructed using accumulated historical manual review data; and the training sample set is used to supervise and adjust the large model to be adjusted to obtain the risk information identification large model. By performing SFT training (supervisory adjustment or supervised fine-tuning) on the large model for audit tasks, the large model can perform audit judgments more professionally on audit tasks.

[0200] It should be noted that, as time goes by, after the model is deployed online, the audit standards may change, and new data forms will continue to emerge. These changes pose a major challenge to the recognition ability of the model. If the model is not continuously updated and optimized, its online performance is likely to gradually decline. Therefore, in the embodiment of the present application, a technology for automatic iterative updating of the model is proposed, which can greatly reduce the maintenance cost of the online model and ensure that the model's capabilities continue to maintain a high level.

[0201] The automatic update iteration of the model is divided into small version iteration and large version iteration. As an embodiment, for the risk information identification large model, the following method is used to perform the model update of the first update method: the parameters of the risk information identification large model used online are used as initial parameters to form the large model to be updated; a predetermined number of historical training samples are selected from the historical training sample set, and a predetermined number of data are selected from the data newly added during use as new training samples, and the two are combined into a first new training sample set; the first new training sample set is used to supervise and adjust the large model to be updated to obtain the model update of the first update method.

[0202] In specific implementation, the model update of the first update method is a small version iteration. Figure 6 , Figure 6 This is a schematic diagram of the model minor version update process in a method for identifying risk information for a target object, provided in the first embodiment of this application. Minor version iteration is designed to address the problem of insufficient existing model recognition capabilities and facilitate rapid iterative updates. When new product risk manifestations emerge and the model's insufficient recognition capabilities lead to misreviews, minor version iteration can be used to quickly replenish the model's capabilities and prevent the risk. Minor version iteration uses an online model for parameter initialization, with training samples limited to approximately 100,000. A predetermined number of historical training samples are selected from the historical training sample set, and a predetermined number of newly added data during use are selected as new training samples, for example, 50%. This means that 50% of the historical training set is randomly sampled, and 50% of the newly added risk samples are used. This ensures that the resulting model maintains its recognition capabilities for the historical data while also providing good recognition for the newly added data. In practical applications, training on 100,000 samples can generally be completed within one day, and online deployment can be completed within three days.

[0203] As an embodiment, for the risk information identification big model, the following method is used to perform the model update of the second update method: select a general big model as the big model to be updated; select a predetermined number of historical training samples from the historical training sample set, and select a predetermined number of data from the data newly added during use as new training samples, and combine the two into a second new training sample set; use the second new training sample set to supervise and adjust the big model to be updated to obtain the model update of the second update method.

[0204] In specific implementation, the model update of the second update method is a major version iteration. Please refer to Figure 7 , Figure 7 This is a schematic diagram of the process of updating a large version of the model in a risk information identification method for a target object provided by the first embodiment of the present application. The iteration of the large model is mainly to correct the problem of recognition capability offset caused by multiple iterations of the small version. After multiple iterations of the small version, the recognition capability of the model will gradually shift to the data of the small version, and shift from global optimization to local optimization (the large model focuses on the data recognition of the small version iteration, and gradually forgets the full amount of data recognition). Therefore, through the form of large version iteration, historical training data and all recent new samples are absorbed to perform a large-scale upgrade, which can pull the large model recognition back to the global optimization solution and solve the offset problem. The disadvantage is that the training cycle is long and the resource consumption is large.

[0205] As an embodiment, for the risk information identification big model, a first update method and a second update method are used to update the model: the first update method includes: using the parameters of the risk information identification big model used online as initial parameters to form the big model to be updated; selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and merging the two into a first new training sample set; using the first new training sample set, supervising and adjusting the big model to be updated to obtain the model update of the first update method; the second update method includes: selecting a general big model as the big model to be updated; selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and merging the two into a second new training sample set; using the second new training sample set, supervising and adjusting the big model to be updated to obtain the model update of the second update method; wherein, the update period of the first update method is set to the first period, and the update period of the second update method is set to the second period, and the length of the first period is less than the second period. In the embodiment of the present application, the model is updated using a first update method and a second update method, i.e., the first update method is a minor version iteration method and the second update method is a major version iteration method. The update cycle of the minor version iteration method is the first cycle, for example, a minor version iteration can be updated once a week, and the update cycle of the major version iteration method is the second cycle, for example, a major version iteration can be updated once a month. The automatic iterative update technology of the model can solve the problem of model attenuation and keep the model at a high level of recognition accuracy at all times.

[0206] The following are examples of large model training samples, for example, large model input: Picture 1: Picture 1\nPicture 2: Image 2\nDetermine whether the content violates the rules? Service identity: XX, app_key: XX, Review channel: , Number of images: 2, Product title: 5-pack of socks, Price: 20.00, Product category: Household daily necessities > Disposable products > Disposable socks, Hit keywords: , Historical penalty information: 0, SKU description: , SPU description: {"material":"cotton"}. Large model output: No. For another example, large model input: Picture 1: Picture 1\nPicture 2: Picture 2\nPicture 3: Picture 3\nPicture 4: Image 4: Determine if the content violates regulations. Service identity: XX, app_key: XX, Review channel: Default, Number of images: 2, Product title: Authentic and sharp household stainless steel meat slicing kitchen knife set, Price: 30.00, Product category: Kitchen / Cooking Utensils > Cooking Utensils > Knives > Kitchen Knives, Keyword: Kitchen Knives, Review criteria: Knife products must indicate attributes such as blade length, tip angle, handle length, and overall blade length. Historical penalty information: Controlled instruments: 7, SKU description: [null][Specification 1][Specification 1], SPU description: {"Whether sharpened": "Yes","Brand":"other / Other","Item number":"00000000","Maximum blade length":"30.1cm","Applicable scenario":"Daily gift giving","Designed service life":"20 years","Target group":"General public","Blade material":"Stainless steel"}. Model output: Yes, the knife attributes are incomplete.

[0207] As can be seen from the above example, when setting up an audit task, after inputting product information, the large model will make a judgment based on the output and provide a judgment result of "yes" or "no". If it is judged to be a violation, the audit label (i.e., the reason for the violation) will be given in addition to the "yes" answer. In actual application, the probability score of the first token being "yes" or "no" is calculated. When the score is greater than the preset probability score, such as 0.9, it means that the model is very confident in the judgment result, and the task is automatically completed.

[0208] In an embodiment of the present application, the predetermined task mode can be to specify a large model to discover a certain type of risk based on aggregated information, and the risk information identification result of the target object includes one or more of the following information: whether there is a risk, specific risk warning information, and risk type information. For example, the predetermined task mode is: You are a content security auditor, and you need to judge whether the content has a content security risk based on the product information and the review standards, and briefly explain the reason (within 200 words). For example, the product information is as follows: Service identity: XX, app_key: XX, Review channel: Hourly inflow, Number of pictures: 6, Product title: Vitamin D calcium soft capsule, 105g / bottle, Price: 138.00, Product category: Health food / dietary nutritional supplement food>Health food>Vitamin / mineral supplement health food, Hit keywords:, Historical penalty information: XX, Face algorithm result:, Violence and terrorism algorithm result: unhit_0.0, Flag algorithm result:, OCR result: Health food, this product is made with calcium carbonate, vitamin D, and soybean oil as the main raw materials, food, this product has the health function of calcium supplement. [Main ingredients] calcium sulfate, vitamin D, soybean oil, [Effective ingredients and content] Every 100g contains: calcium, 24800mg-26100mg, vitamin D, 0.42mg-0.54mg, [Health function] calcium supplement, [Suitable people] Those who need calcium supplement, [Not suitable people] None, [Use method and consumption amount] Once a day, 1-2 tablets each time, [Specification] 1.05g / tablet, [Shelf life] 24 months, [Storage method] Store in a cool and dry place, [Precautions] This product cannot replace medicine, [Production date] See bottle label (year / month / day), Product description:, SKU description:, SPU description: {\"Not suitable for people\":\"None\",\"Main ingredients\":\"calcium carbonate, vitamin D, soybean oil\"}".

[0209] Based on the model's input, the model output is: This product presents a risk. The product information mentions "[Main Ingredient] Calcium Sulfate," but there's an obvious error here: "Calcium Sulfate" should be "Calcium Carbonate."

[0210] In an embodiment of the present application, the risk information identification result of the target object includes an identification result that cannot accurately identify the risk. When the risk information identification result is that the risk cannot be accurately identified, the following output is provided: extracting elements that help to manually identify the risk of the target object's descriptive information, and outputting and displaying them in a form that is convenient for manual identification. The elements that help to manually identify the risk of the target object's descriptive information include one of the following information: direct information elements for risk identification, and knowledge related to the identification of potential risks. Specifically, the direct information elements for risk identification in the descriptive information can be highlighted and displayed, and the knowledge related to the identification of potential risks can be provided to a display for manual identification using red font.

[0211] It should be noted that for some tasks, since the large model cannot guarantee their completion, these tasks will be transferred to manual review and completed by human reviewers. Based on the output of the large model and aggregated information, factors, knowledge, and other results can be output and displayed in a form that is easy for human recognition. For example, they can be displayed to reviewers in a highlighted form to achieve the purpose of improving efficiency and minimize review errors caused by reviewers missing or misreading lengthy product information.

[0212] The risk information identification method for a target object described in an embodiment of the present application includes: obtaining descriptive information for the target object; inputting the descriptive information into a preset completion model to obtain information after the descriptive information is completed as first information corresponding to the target object; inputting the descriptive information into a preset information summary model to extract key information in the descriptive information as second information corresponding to the target object; performing information aggregation processing on the information processed in the previous step to obtain aggregated information corresponding to the target object; the information processed in the previous step includes the first information and the second information; and providing the aggregated information to a preset risk information identification large model in a predetermined task mode to obtain a risk information identification result for the target object.

[0213] The risk information identification method for a target object described in an embodiment of the present application can, after obtaining the descriptive information for the target object, obtain the information after completing the descriptive information based on a preset completion model, and extract the key information in the descriptive information based on a preset information summary model; then, aggregate the information processed in the previous step to obtain the aggregated information corresponding to the target object; finally, provide the aggregated information to the preset risk information identification large model in a predetermined task mode, and then obtain the risk information identification result of the target object. This method improves the efficiency and accuracy of risk information identification of the target object, while reducing costs.

[0214] Second embodiment

[0215] In the first embodiment described above, a method for identifying risk information for a target object is provided. Correspondingly, the second embodiment of the present application provides a system for identifying risk information for a target object. Since the system embodiment is substantially similar to the first method embodiment, the description is relatively brief. For relevant details, please refer to the description of the method embodiment.

[0216] Please refer to Figure 8 , is a schematic diagram of a risk information identification system for a target object provided in the second embodiment of the present application.

[0217] The risk information identification system 800 for a target object includes: a client 801 and a server 802; the server 802 includes: a completion model, an information summary model and a risk information identification model; the client 801 is used to obtain descriptive information for the target object and send the descriptive information to the server 802; the server 802 is used to obtain the descriptive information, input the descriptive information into a preset completion model, and obtain information after the descriptive information is completed as first information corresponding to the target object; input the descriptive information into a preset information summary model, extract key information from the descriptive information as second information corresponding to the target object; aggregate the information processed by the previous step to obtain aggregated information corresponding to the target object; the information processed by the previous step includes the first information and the second information; provide the aggregated information to the preset risk information identification model in a predetermined task mode to obtain a risk information identification result for the target object; send the risk information identification result to the client; the client 801 is used to display the risk information identification result.

[0218] Optionally, the server 802 also includes: an information distribution big model; after obtaining the descriptive information for the target object and before executing other subsequent steps, the information distribution big model is used to: input the descriptive information into the preset information distribution big model, preliminarily identify possible risk points existing in the target object, and determine whether the possible risk points are risk points that the big model can identify with high accuracy; if so, provide the descriptive information of the target object to the subsequent steps; if not, provide the descriptive information of the target object to the manual review link to perform the manual review operation.

[0219] Optionally, the information distribution model is further used to: determine the type of possible risk points existing in the target object, distribute the description information of the target object to a corresponding channel corresponding to the type, and execute the other subsequent steps.

[0220] Optionally, the preset completion model includes at least one special model selected according to the components of the descriptive information, and each of the special models corresponds to the components of the processing description information that are suitable for processing by the special model; the information formed after each special model processes the components that belong to the special model constitutes the information after completion processing.

[0221] Optionally, the dedicated model includes one or more of the following types: a graphic object recognition model, a text recognition model, and a face recognition model.

[0222] Optionally, the system further includes an information group construction module, configured to perform the following steps on the description information before extracting key information from the description information:

[0223] Based on the content of the description information, one or more information groups with a preset structure are extracted and constructed; the information groups serve as information units for executing the step of extracting key information from the description information.

[0224] Optionally, the extracting of key information from the descriptive information includes performing the following steps on the information group: obtaining a risk probability assessment value for each information group through risk assessment; sorting the risk probability of each information group according to the risk probability assessment value of each information group; and based on the sorting result, using a preset number of information groups with high risk probability as the second information corresponding to the target object.

[0225] Optionally, the extracting of key information from the descriptive information includes performing the following steps on the information group: obtaining a risk probability assessment value for each information group through risk assessment; and taking the information group whose risk probability assessment value is higher than a preset risk threshold as the second information corresponding to the target object.

[0226] Optionally, the server 802 also includes: a retrieval enhancement generation knowledge base; the retrieval enhancement generation knowledge base is used to obtain the description information, and retrieve the retrieval enhancement information related to the description information for the description information as the third information corresponding to the target object; the third information is provided as part of the aggregated information to the risk information identification model.

[0227] Optionally, the initial version of the risk information identification large model is adjusted in the following manner: a general large model is selected as the large model to be adjusted; a training sample set is constructed using accumulated historical manual review data; and the training sample set is used to supervise and adjust the large model to be adjusted to obtain the risk information identification large model.

[0228] Optionally, for the risk information identification big model, the model update of the first update method is performed in the following manner: the parameters of the risk information identification big model used online are used as initial parameters to form the big model to be updated; a predetermined number of historical training samples are selected from the historical training sample set, and a predetermined number of data from the data newly added during use are selected as new training samples, and the two are combined into a first new training sample set; the first new training sample set is used to supervise and adjust the big model to be updated to obtain the model update of the first update method.

[0229] Optionally, for the risk information identification big model, the model update of the second update method is performed in the following manner: a general big model is selected as the big model to be updated; a predetermined number of historical training samples are selected from the historical training sample set, and a predetermined number of data are selected from the data newly added during use as new training samples, and the two are combined into a second new training sample set; the second new training sample set is used to supervise and adjust the big model to be updated to obtain the model update of the second update method.

[0230] Optionally, for the risk information identification big model, a first update method and a second update method are used to update the model: the first update method includes: using the parameters of the risk information identification big model used online as initial parameters to form the big model to be updated; selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and merging the two into a first new training sample set; using the first new training sample set, supervising and adjusting the big model to be updated to obtain the model update of the first update method; the second update method includes: selecting a general big model as the big model to be updated; selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and merging the two into a second new training sample set; using the second new training sample set, supervising and adjusting the big model to be updated to obtain the model update of the second update method; wherein, the update period of the first update method is set to the first period, and the update period of the second update method is set to the second period, and the length of the first period is less than the second period.

[0231] Optionally, the risk information identification result of the target object includes one or more of the following information: whether there is a risk, specific risk warning information, and risk type information.

[0232] Optionally, the risk information identification result of the target object includes an identification result that the risk cannot be accurately identified.

[0233] Optionally, when the risk information identification result is that the risk cannot be accurately identified, the following output is provided: extracting elements that help to manually identify the risks of the descriptive information of the target object, and outputting them in a form that is convenient for manual identification.

[0234] Optionally, the elements that help to manually identify risks in the description information of the target object include one of the following information: direct information elements for risk identification, and knowledge related to the identification of potential risks.

[0235] Third embodiment

[0236] In the first embodiment described above, a method for identifying risk information for a target object is provided. Correspondingly, the third embodiment of the present application provides an apparatus for identifying risk information for a target object. Since the apparatus embodiment is substantially similar to the first method embodiment, the description is relatively brief. For relevant details, please refer to the description of the method embodiment. The apparatus embodiment described below is merely illustrative.

[0237] Please refer to Figure 9 , is a schematic diagram of a risk information identification device for a target object provided in the third embodiment of the present application.

[0238] The risk information identification device 900 for a target object includes: a description information obtaining unit 901, used to obtain description information for the target object; a first information obtaining unit 902, used to input the description information into a preset completion model, and obtain information after the description information is completed as the first information corresponding to the target object; a second information obtaining unit 903, used to input the description information into a preset information summary model, and extract key information in the description information as the second information corresponding to the target object; an aggregated information obtaining unit 904, used to perform information aggregation processing on the information processed by the previous step, and obtain aggregated information corresponding to the target object; the information processed by the previous step includes the first information and the second information; a risk information identification result obtaining unit 905, used to provide the aggregated information to a preset risk information identification large model in a predetermined task mode, and obtain a risk information identification result for the target object.

[0239] Optionally, it also includes: an information distribution unit; after obtaining the descriptive information for the target object and before executing other subsequent steps, the information distribution unit is used to: input the descriptive information into a preset information distribution model, preliminarily identify possible risk points existing in the target object, and determine whether the possible risk points are risk points that can be identified with high accuracy by the large model; if so, provide the descriptive information of the target object to subsequent steps; if not, provide the descriptive information of the target object to the manual review link to perform a manual review operation.

[0240] Optionally, the information distribution unit is further used to: determine the type of the possible risk point existing in the target object, distribute the description information of the target object to a corresponding channel corresponding to the type, and execute the other subsequent steps.

[0241] Optionally, the preset completion model includes at least one special model selected according to the components of the descriptive information, and each of the special models corresponds to the components of the processing description information that are suitable for processing by the special model; the information formed after each special model processes the components that belong to the special model constitutes the information after completion processing.

[0242] Optionally, the dedicated model includes one or more of the following types: a graphic object recognition model, a text recognition model, and a face recognition model.

[0243] Optionally, before extracting the key information from the description information, the following steps are performed on the description information: based on the content of the description information, one or more information groups with a preset structure are extracted and constructed; the information group serves as an information unit for executing the step of extracting the key information from the description information.

[0244] Optionally, the extracting of key information from the descriptive information includes performing the following steps on the information group: obtaining a risk probability assessment value for each information group through risk assessment; sorting the risk probability of each information group according to the risk probability assessment value of each information group; and based on the sorting result, using a preset number of information groups with high risk probability as the second information corresponding to the target object.

[0245] Optionally, the extracting of key information from the descriptive information includes performing the following steps on the information group: obtaining a risk probability assessment value for each information group through risk assessment; and taking the information group whose risk probability assessment value is higher than a preset risk threshold as the second information corresponding to the target object.

[0246] Optionally, it also includes: a retrieval enhancement generation unit; the retrieval enhancement generation unit is used to provide the description information to a preset retrieval enhancement generation knowledge base, retrieve and obtain retrieval enhancement information related to the description information as the third information corresponding to the target object; the third information is provided as part of the aggregated information provided to the risk information identification model.

[0247] Optionally, the initial version of the risk information identification large model is adjusted in the following manner: a general large model is selected as the large model to be adjusted; a training sample set is constructed using accumulated historical manual review data; and the training sample set is used to supervise and adjust the large model to be adjusted to obtain the risk information identification large model.

[0248] Optionally, for the risk information identification big model, the model update of the first update method is performed in the following manner: the parameters of the risk information identification big model used online are used as initial parameters to form the big model to be updated; a predetermined number of historical training samples are selected from the historical training sample set, and a predetermined number of data from the data newly added during use are selected as new training samples, and the two are combined into a first new training sample set; the first new training sample set is used to supervise and adjust the big model to be updated to obtain the model update of the first update method.

[0249] Optionally, for the risk information identification big model, the model update of the second update method is performed in the following manner: a general big model is selected as the big model to be updated; a predetermined number of historical training samples are selected from the historical training sample set, and a predetermined number of data are selected from the data newly added during use as new training samples, and the two are combined into a second new training sample set; the second new training sample set is used to supervise and adjust the big model to be updated to obtain the model update of the second update method.

[0250] Optionally, for the risk information identification big model, a first update method and a second update method are used to update the model: the first update method includes: using the parameters of the risk information identification big model used online as initial parameters to form the big model to be updated; selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and merging the two into a first new training sample set; using the first new training sample set, supervising and adjusting the big model to be updated to obtain the model update of the first update method; the second update method includes: selecting a general big model as the big model to be updated; selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and merging the two into a second new training sample set; using the second new training sample set, supervising and adjusting the big model to be updated to obtain the model update of the second update method; wherein, the update period of the first update method is set to the first period, and the update period of the second update method is set to the second period, and the length of the first period is less than the second period.

[0251] Optionally, the risk information identification result of the target object includes one or more of the following information: whether there is a risk, specific risk warning information, and risk type information.

[0252] Optionally, the risk information identification result of the target object includes an identification result that the risk cannot be accurately identified.

[0253] Optionally, when the risk information identification result is that the risk cannot be accurately identified, the following output is provided: extracting elements that help to manually identify the risks of the descriptive information of the target object, and outputting them in a form that is convenient for manual identification.

[0254] Optionally, the elements that help to manually identify risks in the description information of the target object include one of the following information: direct information elements for risk identification, and knowledge related to the identification of potential risks.

[0255] Fourth embodiment

[0256] Corresponding to the above method embodiments of the present application, the fourth embodiment of the present application further provides an electronic device. Figure 10 As shown, Figure 10This is a schematic diagram of an electronic device provided in the fourth embodiment of the present application. The electronic device includes: at least one processor 1001, at least one communication interface 1002, at least one memory 1003 and at least one communication bus 1004; optionally, the communication interface 702 can be an interface of a communication module, such as an interface of a GSM module; the processor 701 may be a processor CPU, or a specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present invention. The memory 1003 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk storage. The memory 1003 stores a program, and the processor 1001 calls the program stored in the memory 1003 to execute the method provided in the above embodiment of the present application.

[0257] Fifth embodiment

[0258] Corresponding to the above method of the present application, the fifth embodiment of the present application further provides a computer storage medium. The computer storage medium stores a computer program, which is executed by a processor to execute the method provided in the above embodiment of the present application.

[0259] Although the present application is disclosed as above with the preferred embodiments, it is not intended to limit the present application. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

[0260] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0261] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0262] 1. Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include non-transitory media such as modulated data signals and carrier waves.

[0263] 2. Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0264] It should be noted that the embodiments of the present application may involve the use of user data. In actual applications, user-specific personal data can be used in the scheme described herein within the scope permitted by applicable laws and regulations, subject to the requirements of applicable laws and regulations of the country where the user is located (for example, with the user's explicit consent, effective notification to the user, etc.).

[0265] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

Claims

1. A method for identifying risk information of a target object, characterized in that: include: Get descriptive information about the target object; Inputting the description information into a preset completion model to obtain information after completing the description information as first information corresponding to the target object; Inputting the description information into a preset information summary model, extracting key information from the description information as second information corresponding to the target object; Aggregate the information processed in the previous step to obtain aggregated information corresponding to the target object; the information processed in the previous step includes the first information and the second information; The aggregated information is provided to a preset risk information identification model in a predetermined task mode to obtain a risk information identification result for the target object.

2. The method for identifying risk information of a target object according to claim 1, characterized in that: After obtaining the description information of the target object and before executing other subsequent steps, the method further includes: Input the description information into a preset information distribution model to preliminarily identify possible risk points in the target object and determine whether the possible risk points are risk points that can be identified with high accuracy by the model; If yes, the description information of the target object is provided to the subsequent steps; If not, the description information of the target object is provided to the manual review link to perform a manual review operation.

3. The method for identifying risk information of a target object according to claim 2, characterized in that: The information distribution model also performs the following steps: Determine the type of the possible risk point in the target object, distribute the description information of the target object to the corresponding channel corresponding to the type, and execute the other subsequent steps.

4. The method for identifying risk information of a target object according to claim 1, characterized in that: The preset completion model includes at least one special model selected according to the components of the description information, and each of the special models corresponds to the components of the description information that are suitable for processing by the special model; the information formed after each special model processes the components that belong to the special model to be processed constitutes the information after the completion processing.

5. The method for identifying risk information of a target object according to claim 4, characterized in that: The dedicated model includes one or more of the following types: a graphic object recognition model, a text recognition model, and a human face recognition model.

6. The method for identifying risk information of a target object according to claim 1, characterized in that: Before extracting the key information from the description information, the following steps are performed on the description information: Based on the content of the description information, one or more information groups with a preset structure are extracted and constructed; the information groups serve as information units for executing the step of extracting key information from the description information.

7. The method for identifying risk information of a target object according to claim 6, characterized in that: The extracting key information from the description information includes performing the following steps on the information group: Through risk assessment, the risk probability assessment value of each information group is obtained; Sort the risk probabilities of each information group according to the risk probability assessment values of each information group; According to the sorting result, a preset number of information groups with high risk probability are used as the second information corresponding to the target object.

8. The method for identifying risk information of a target object according to claim 1, characterized in that: Also includes: The description information is provided to a preset retrieval enhancement generation knowledge base, and retrieval enhancement information related to the description information is retrieved as the third information corresponding to the target object; the third information is provided as part of the aggregated information provided to the risk information identification model.

9. The method for identifying risk information of a target object according to claim 1, characterized in that: The initial version of the risk information identification model is obtained by adjusting the following method: Select the general large model as the large model to be adjusted; Use the accumulated historical manual review data to construct a training sample set; The training sample set is used to perform supervised adjustment on the large model to be adjusted to obtain the risk information identification large model.

10. The method for identifying risk information of a target object according to claim 1, characterized in that: For the risk information identification large model, the first update method and the second update method are used to update the model: The first updating method includes: using the parameters of the risk information identification large model used online as initial parameters to form the large model to be updated; Selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and combining the two into a first new training sample set; Using the first new training sample set, supervise and adjust the large model to be updated to obtain a model update using a first updating method; The second updating method includes: selecting a common large model as the large model to be updated; Selecting a predetermined number of historical training samples from the historical training sample set, and selecting a predetermined number of data from the data newly added during use as new training samples, and combining the two to form a second new training sample set; Using the second new training sample set, supervise and adjust the large model to be updated to obtain a model update using a second updating method; The update period of the first update mode is set to a first period, the update period of the second update mode is set to a second period, and the length of the first period is shorter than the second period.

11. The method for identifying risk information of a target object according to claim 1, characterized in that: The risk information identification result of the target object includes one or more of the following information: whether there is a risk, specific risk warning information, and risk type information.

12. A risk information identification system for a target object, characterized in that: include: Client and server; the server includes: completion model, information summary model and risk information identification model; The client is configured to obtain description information of a target object and send the description information to the server; The server is configured to obtain the description information, input the description information into a preset completion model, obtain information after completing the description information, and use it as first information corresponding to the target object; input the description information into a preset information summary model, extract key information from the description information, and use it as second information corresponding to the target object; perform information aggregation on the information processed in the preceding steps to obtain aggregated information corresponding to the target object; the information processed in the preceding steps includes the first information and the second information; provide the aggregated information to a preset risk information identification model in a predetermined task mode to obtain a risk information identification result for the target object; and send the risk information identification result to the client; The client is used to display the risk information identification result.

13. A risk information identification device for a target object, characterized in that: include: A description information obtaining unit, configured to obtain description information for a target object; a first information obtaining unit, configured to input the description information into a preset completion model, and obtain information after completing the description information as first information corresponding to the target object; A second information obtaining unit is configured to input the description information into a preset information summary model and extract key information from the description information as second information corresponding to the target object; an aggregate information obtaining unit, configured to aggregate the information processed in the preceding step to obtain aggregate information corresponding to the target object; the information processed in the preceding step includes the first information and the second information; The risk information identification result obtaining unit is used to provide the aggregated information to a preset risk information identification model in a predetermined task mode to obtain the risk information identification result for the target object.

14. An electronic device, characterized in that: The electronic device includes a processor and a memory; The memory is configured to store a program of a risk information identification method for a target object; the processor is capable of executing the method according to any one of claims 1 to 11 according to the program stored in the memory.

15. A computer storage medium, characterized in that The computer storage medium stores a computer program, and after the computer program is run by the processor, the method according to any one of claims 1 to 11 is executed.

Images