Methods, apparatus, computer equipment and storage media for distributing objects to be reviewed

Abstract

This application relates to a method, apparatus, computer device, storage medium, and computer program product for distributing objects awaiting review. The method includes: responding to an object retrieval request, searching for a historical review index matching the requester; determining, based on the historical review index, a historical object identifier and a historical source identifier for the historical review object corresponding to the requester; searching for at least one target class index including the historical object identifier, and obtaining a first object awaiting review associated with each target class index, wherein the target class indexes are established based on object similarity; searching for a target source index from multiple source indexes based on the historical source identifier, and obtaining a second object awaiting review associated with each target source index; and determining a target object awaiting review matching the requester based on the first and second objects awaiting review, and pushing the target object awaiting review. This method enables accurate delivery of objects awaiting review.

Description

Technical Field

[0001] This application relates to the field of computer information processing technology, and in particular to a method, apparatus, computer equipment, storage medium, and computer program product for distributing objects to be examined. Background Technology

[0002] In this era of rapid internet development, there is an increasing number of things that need to be reviewed, such as content to be published, articles to be pushed out, or information to be promoted, which need to be reviewed to see if they involve sensitive content or whether the content quality meets quality standards.

[0003] In related technologies, the main approach is to randomly push content for review or based on the availability of reviewers. However, with the volume of content review and manual review increasing year by year in the internet sector, and the number of review team members expanding exponentially, this simple push method struggles to take into account the actual situation of reviewers, resulting in inaccurate information distribution. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer device, computer-readable storage medium, and computer program product for distributing objects under review, which can improve the accuracy of distributing objects under review, in response to the above-mentioned technical problems.

[0005] On the one hand, this application provides a method for distributing objects to be examined, the method comprising:

[0006] In response to an object claiming request, search the historical audit index that matches the requester of the object claiming request;

[0007] Based on the historical review index, determine the historical object identifier and historical source identifier of the historical review object corresponding to the requester;

[0008] Find at least one target class index that includes historical object identifiers, and obtain the first object to be examined associated with each target class index. The target class index is built based on the similarity between objects.

[0009] Based on historical source identifiers, the target source index is searched from multiple source indexes, and the second object to be examined associated with each target source index is obtained;

[0010] Based on the first and second pending objects, the target pending objects that match the requester are identified and pushed to the target pending objects.

[0011] On the other hand, this application also provides a device for distributing objects to be examined, the device comprising:

[0012] The lookup module is used to respond to an object claiming request and find the historical audit index that matches the requester of the object claiming request.

[0013] The first determination module is used to determine the historical object identifier and historical source identifier of the historical review object corresponding to the requester based on the historical review index;

[0014] The search module is also used to search for at least one target class index that includes a historical object identifier, and to obtain the first object to be examined associated with each target class index. The target class index is established based on the similarity between objects.

[0015] The search module is also used to search for the target source index from multiple source indexes based on the historical source identifier, and to obtain the second object to be examined associated with each target source index;

[0016] The second determination module is used to determine the target object to be reviewed that matches the requester based on the first object to be reviewed and the second object to be reviewed, and to push the target object to be reviewed.

[0017] On the other hand, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0018] In response to an object claiming request, search the historical audit index that matches the requester of the object claiming request;

[0019] Based on the historical review index, determine the historical object identifier and historical source identifier of the historical review object corresponding to the requester;

[0020] Find at least one target class index that includes historical object identifiers, and obtain the first object to be examined associated with each target class index. The target class index is built based on the similarity between objects.

[0021] Based on historical source identifiers, the target source index is searched from multiple source indexes, and the second object to be examined associated with each target source index is obtained;

[0022] Based on the first and second pending objects, the target pending objects that match the requester are identified and pushed to the target pending objects.

[0023] On the other hand, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:

[0024] In response to an object claiming request, search the historical audit index that matches the requester of the object claiming request;

[0025] Based on the historical review index, determine the historical object identifier and historical source identifier of the historical review object corresponding to the requester;

[0026] Find at least one target class index that includes historical object identifiers, and obtain the first object to be examined associated with each target class index. The target class index is built based on the similarity between objects.

[0027] Based on historical source identifiers, the target source index is searched from multiple source indexes, and the second object to be examined associated with each target source index is obtained;

[0028] Based on the first and second pending objects, the target pending objects that match the requester are identified and pushed to the target pending objects.

[0029] On the other hand, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:

[0030] In response to an object claiming request, search the historical audit index that matches the requester of the object claiming request;

[0031] Based on the historical review index, determine the historical object identifier and historical source identifier of the historical review object corresponding to the requester;

[0032] Find at least one target class index that includes historical object identifiers, and obtain the first object to be examined associated with each target class index. The target class index is built based on the similarity between objects.

[0033] Based on historical source identifiers, the target source index is searched from multiple source indexes, and the second object to be examined associated with each target source index is obtained;

[0034] Based on the first and second pending objects, the target pending objects that match the requester are identified and pushed to the target pending objects.

[0035] The aforementioned method, apparatus, computer equipment, storage medium, and computer program product for distributing objects to be reviewed can directly search for historical review indexes matching the requester upon receiving an object retrieval request. Based on these historical review indexes, the historical object identifier and historical source identifier of the historical review object corresponding to the requester can be determined. This allows for the retrieval of at least one target class index and target source index highly correlated with the review records, based on the requester's previously reviewed historical records. Consequently, the objects to be reviewed associated with these target class and source indexes can be directly obtained. Since the target objects to be reviewed are obtained from the target class index containing the historical object identifier and the target source index containing the historical source identifier, the obtained target objects to be reviewed can match the historical review objects already reviewed by the requester in different dimensions such as object source or object similarity. This allows for the push of objects to be reviewed in a way that aligns with the reviewer's capabilities and work habits, making the distribution of objects to be reviewed more targeted and significantly improving the accuracy of distribution. Furthermore, because the push is more targeted, the reviewer's review experience can be reused promptly and effectively when reviewing target objects to be reviewed, thereby improving review efficiency. Attached Figure Description

[0036] Figure 1 This is a diagram illustrating the application environment of a method for distributing objects to be examined in one embodiment.

[0037] Figure 2 This is a flowchart illustrating a method for distributing objects to be examined in one embodiment;

[0038] Figure 3 This is a schematic diagram of the structure of the historical audit index in one embodiment;

[0039] Figure 4 This is a schematic diagram of the target class index structure in one embodiment;

[0040] Figure 5 This is a schematic diagram of the structure of the target source index in one embodiment;

[0041] Figure 6 This is a schematic diagram of the graph topology structure formed between objects and elements in one embodiment;

[0042] Figure 7 This is a flowchart illustrating the object distribution method in another embodiment;

[0043] Figure 8 This is a schematic diagram of the architecture of the indexing and distribution process of an advertisement to be reviewed in one embodiment;

[0044] Figure 9 This is a schematic diagram of the architecture of the sorting model experiment process in one embodiment;

[0045] Figure 10 This is a structural block diagram of a device for distributing objects to be examined in one embodiment;

[0046] Figure 11 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0048] First, before elaborating on the method for distributing the objects to be examined in this application, let's briefly explain some of the relevant terms used in the embodiments of this application:

[0049] Real-time computing, also known as instantaneous computing, is the study of computer hardware and software systems subject to "instantaneous constraints" in computer science. Instantaneous constraints refer to the maximum time limit between an event occurring and the system's response. Instantaneous programs must guarantee a response within strict time limits. Instantaneous response times are typically measured in milliseconds, but sometimes in microseconds. In contrast, non-instantaneous systems cannot guarantee that their response time will meet real-time constraints under all conditions. It is possible that in most cases, non-instantaneous systems can meet real-time constraints, or even be faster, but they cannot guarantee compliance under all conditions.

[0050] Streaming data refers to a dynamic collection of data that is infinite in terms of time distribution and quantity. The value of the data decreases over time, thus requiring real-time computation to provide a response within seconds.

[0051] Stream computing vs. batch computing: Stream computing refers to real-time computation on a data stream; batch computing refers to a data computation method that collects data uniformly, stores it in a database, and then processes the data in batches. The main differences between the two are as follows:

[0052] (1) Different data timeliness: Stream computing is real-time and has low latency, while batch computing is not real-time and has high latency.

[0053] (2) Different data characteristics: Stream computing data is generally dynamic and without boundaries, while batch computing data is generally static.

[0054] (3) Different application scenarios: Stream computing is used in real-time scenarios, which are also scenarios with high timeliness requirements, such as real-time recommendations or business monitoring. Batch computing is generally used in offline computing scenarios with low real-time requirements, such as data analysis or offline reporting.

[0055] (4) The operation modes are different. Stream computing tasks are usually performed continuously, while batch computing tasks are completed at once.

[0056] Inverted index: also known as reverse index, inverted archive, or reverse archive, is an indexing method primarily used to store the mapping of the storage location of a word in a document or a set of documents under full-text search.

[0057] Collaborative filtering primarily analyzes the similarities between users or things to predict content that a user might be interested in and then recommends that content to the user.

[0058] Graph embedding refers to representing nodes in a graph as low-dimensional dense vectors. Nodes similar in the original graph are expected to have similar low-dimensional representations. These representation vectors can then be used for downstream tasks such as node classification, link prediction, visualization, or reconstructing the original graph.

[0059] The method for distributing objects to be examined provided in the embodiments of this application will be described in detail below:

[0060] In some embodiments, the object distribution method provided in this application can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or integrated into the cloud or other servers. The method for distributing objects to be examined mentioned in this application embodiment can be implemented by terminal 102 or server 104 independently, or by terminal 102 and server 104 collaboratively.

[0061] The following is an example of the method for distributing pending objects according to this application, executed solely by server 104: Terminal 102 sends an object retrieval request to server 104. After receiving the object retrieval request, server 104 searches for a historical review index that matches the requester. Based on the historical review index, server 104 determines the historical object identifier and historical source identifier of the historical review object corresponding to the requester. Server 104 searches for at least one target class index that includes the historical object identifier and obtains the first pending object associated with each target class index. Based on the historical source identifier, server 104 searches for a target source index from multiple source indexes and obtains the second pending object associated with each target source index. Based on the first and second pending objects, server 104 determines the target pending object that matches the requester and pushes the target pending object to terminal 102.

[0062] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. Applications, such as video or audio applications, can run on the terminal to present objects. The server 104 can be a backend server corresponding to software, web pages, or mini-programs, or a server specifically used for distributing objects to be reviewed; this embodiment does not impose specific limitations. The server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers.

[0063] In some embodiments, the above definitions, technical explanations, and implementation environment descriptions are combined, such as... Figure 2 As shown, a method for distributing objects to be reviewed is provided. The specific types of objects to be reviewed can vary depending on the actual application scenario, and this embodiment does not impose specific limitations on this. For example, if the review is of published content, the object to be reviewed can be the published content itself. If the review is of promotional messages (such as advertisements), the object to be reviewed can be the promotional message. This method can be applied to computer devices (specifically, the computer device can be...) Figure 1 Taking a terminal or server as an example, the following steps are included:

[0064] Step 202: In response to the object claim request, look up the historical audit index that matches the requester of the object claim request.

[0065] Specifically, an object retrieval request is a request initiated by the requesting party to retrieve an object awaiting review. The types of objects awaiting review can be found in the explanations above. Before this step, the requesting party may have already sent an object retrieval request to the computer device, thus participating in the review process. Therefore, the requesting party may have some historical review records for objects, which can form a historical review index. It should be noted that the reason the computer device searches for historical review indexes matching the requesting party in this step is primarily so that it can subsequently use these historical review indexes to find the target class index and target source index associated with objects awaiting review, and then push the objects awaiting review associated with these two target indexes to achieve the distribution of objects awaiting review.

[0066] The historical review index can include a search key and data records. The search key corresponds to the requester. The search key of the historical review index can be the requester identifier, such as the requester's name (reviewer_name), which serves as the basis for finding historical review indexes that match the requester. Specifically, in actual implementation, the object retrieval request can carry the requester identifier. The computer device can then find the historical review index with that requester identifier as the search key, which will be used as the historical review index that matches the requester.

[0067] Data records can be appended to the search primary key and associated with the historical audit objects of the corresponding requester. Multiple data records can exist, each corresponding to a historical audit object of the requester at different times. Each data record may include a historical object identifier and a historical source identifier. The historical object identifier uniquely corresponds to a historical audit object, and one historical object identifier can correspond to multiple different historical source identifiers. In addition, data records may include other information, such as the file size of the historical audit object; however, this embodiment does not specifically limit this information.

[0068] The above describes one possible structure for a historical review index. In this structure, the search primary key is the requester identifier, and there can be multiple data records. Each data record corresponds to a historical review object, and each data record can include at least one type of data, such as at least the historical object identifier. Each data record can cover all types of data for the historical review object. Another possible structure for the historical review index is where the search primary key is the requester identifier, and there can be multiple data records, but each data record includes only one type of data, such as only the historical object identifier. It should also be noted that there can be multiple historical review indexes, all including the same search primary key, but each historical review index corresponds to a different type of data. Of course, in actual implementation, the structure of the historical review index can be designed based on requirements, and this embodiment does not impose specific limitations on this.

[0069] Step 204: Based on the historical audit index, determine the historical object identifier and historical source identifier of the historical audit object corresponding to the requester.

[0070] As previously discussed, the historical review index can include data records, and these data records can include historical object identifiers. Therefore, for a historically reviewed object that has been reviewed by the requester, the computer device can retrieve the historical object identifier corresponding to the historical review object from the data records of the historical review index, and it can also retrieve the historical source identifier. The historical source identifier can also uniquely correspond to a historical review object. The historical source identifier can be used to point to the source of the historical review object, such as the originator of the historical review object, the field or industry to which the historical review object belongs, etc. It is understandable that there can be more than one historical source identifier for a historical review object.

[0071] The above mainly describes a single historical audit object. Since a requester may have multiple historical audit objects, these multiple historical audit objects can correspond to multiple data records. Each data record can include a historical object identifier and at least one historical source identifier. For the specific structure of the historical audit index, please refer to [reference needed]. Figure 3 .exist Figure 3 In this structure, the first data record corresponds to the first historical review object reviewed by the requester, and the m-th data record corresponds to the m-th historical review object reviewed by the requester. In the first data record, "Historical Object Identifier 1" represents the historical object identifier of the first historical review object, while "Historical Source Identifier 1_1" represents the first historical source identifier of the first historical review object.

[0072] Step 206: Locate at least one target class index that includes a historical object identifier, and obtain the first object to be reviewed associated with each target class index. The target class index is established based on the similarity between objects.

[0073] The target index can include both a search primary key and data records. Computer device lookups include target indexes with historical object identifiers; specifically, the search can be performed on target indexes where the search primary key is the historical object identifier. The data records in the target index can be the object identifiers of the object to be reviewed.

[0074] It should be noted that the target class index is established based on the similarity between different objects. That is, the objects to be reviewed under the same target class index have a certain degree of similarity. Specifically, the similarity between any two objects to be reviewed under the same target class index can reach a preset similarity threshold, or the similarity between all objects to be reviewed under the same target class index and the historical review objects corresponding to the search primary key can reach a preset similarity threshold. This application embodiment does not limit this.

[0075] It is understood that there may be more than one data record in each target class index, that is, there may be more than one first object to be reviewed associated with each target class index. In one embodiment, the computer device may use all objects to be reviewed associated with the target class index as the first object to be reviewed. In other embodiments, considering that there may be many data records attached to the target class index, using all data records as the first object to be reviewed may lead to uneven distribution of the objects to be reviewed. The computer device may use some of the data records as the first object to be reviewed, such as randomly selecting some data records or selecting a specified portion of data records as the first object to be reviewed. This application does not limit this.

[0076] It should be noted that for a specific historical object identifier, such as historical object identifier 1, there may be multiple target class indexes that include historical object identifier 1. This is because when building the target class index, if the historical object identifier is used as the search primary key, various index building methods can be used to build the target class index, thus creating multiple target class indexes corresponding to that historical object identifier. For details, please refer to [reference needed]. Figure 4 .exist Figure 4 In the process, the computer device finds p target class indexes that include "historical object identifier 1".

[0077] Step 208: Based on the historical source identifier, search for the target source index from multiple source indexes and obtain the second object to be reviewed associated with each target source index.

[0078] The target source index may also include a search primary key and data records. The computer device "searches for the target source index," which may be a source index whose search primary key is a historical source identifier, and uses this as the target source index. The data records in the target source index can be objects to be reviewed, i.e., second objects to be reviewed. It is understood that each target source index may contain more than one data record, meaning that each target source index may be associated with more than one second object to be reviewed. In one embodiment, the computer device may use all objects to be reviewed associated with the target source index as second objects to be reviewed. In other embodiments, considering that there may be many data records associated with the target source index, using all data records as second objects to be reviewed may lead to uneven distribution of objects to be reviewed. Therefore, the computer device may use only a portion of the data records as second objects to be reviewed, such as randomly selecting a portion of the data records or selecting a specified portion of the data records as second objects to be reviewed. This application embodiment does not limit this approach.

[0079] The reason for generating "multiple source indexes" in this step is that there may be many types of historical source identifiers, thus multiple source indexes may be built based on different types of historical source identifiers. This step primarily searches among these "multiple source indexes" for the source index whose primary key is the historical source identifier determined in step 204, and uses this as the target source index. It should be noted that since the historical source identifier determined in step 204 may also be of multiple types, there may also be multiple target source indexes found in this step. For details, please refer to [link / reference]. Figure 5 .exist Figure 5 In the text, “Historical Source Identifier 1” to “Historical Source Identifier n” represent different types of historical source identifiers.

[0080] Step 210: Based on the first and second review objects, determine the target review object that matches the requester and push the target review object.

[0081] Specifically, the computer device can combine all the first and second pending objects as target pending objects to match the requester and push the target pending objects to the requester. In other embodiments, the computer device can also filter the set consisting of the first and second pending objects and use the filtered pending objects as target pending objects. Filtering methods include, for example, randomly selecting a preset number of pending objects as target pending objects, or filtering a preset number of pending objects with earlier review times as target pending objects, etc. The specific filtering method can be determined based on the actual situation, and this application embodiment does not limit it.

[0082] In some embodiments, the computer device may also sort the target objects to be reviewed according to a preset sorting rule before pushing them to the review terminal corresponding to the requester. The preset sorting rule may be, for example, the order of review time, the priority order of the source of the objects to be reviewed, etc., and this application embodiment does not limit this.

[0083] It should be noted that the target class index and the target source index may associate overlapping objects under review, meaning there may be overlap between the first and second objects under review. Taking ad A as the object under review and ad B as the corresponding historical review object for the requester as an example: If ad A is stored as a data record in the target class index built with the object identifier of ad B because its content is similar to ad B (in this case, ad A is the first object under review); and simultaneously, ad A is stored as a data record in the target source index built with the corresponding historical source identifier of the automotive industry because it originates from an ad in the automotive industry (in this case, ad A is the second object under review), then ad A will become an overlapping object under review between the first and second objects under review.

[0084] Therefore, in this step, the computer device can determine the deduplicated union of the first and second objects to be reviewed, using it as the target object to be reviewed that matches the requester, and then push the target object to be reviewed to the requester who received the object request. It should be noted that the requester can obtain the distributed objects to be reviewed using the method provided in this application embodiment, and subsequently, the requester can become an examiner to review the objects to be reviewed.

[0085] The aforementioned method for distributing pending review objects targets the historical object identifier and historical source identifier of the historical review objects corresponding to the requester. Since the target pending review object is obtained from the target class index containing the historical object identifier and the target source index containing the historical source identifier, the obtained target pending review object can match the historical review objects already reviewed by the requester in different dimensions such as object source or object content. This allows for targeted distribution that aligns with the reviewer's capabilities, preferences, and work habits, significantly improving the accuracy of the distribution. Furthermore, this more targeted approach enables reviewers to promptly and effectively reuse their review experience when reviewing target pending review objects, thereby improving review efficiency.

[0086] In some embodiments, determining the historical object identifier and historical source identifier of the historical audit object corresponding to the requester based on the historical audit index includes: determining the historical audit object whose audit time occurred within a preset time period from the historical audit index, wherein the preset time period meets the real-time condition; and obtaining the historical object identifier and historical source identifier of the determined historical audit object.

[0087] Depend on Figure 3It is understood that the historical review index may contain multiple data records, each corresponding to a historical review object. Different data records can correspond to historical review objects of the requester at different times. It is comprehensible that the requester typically has many historical review objects already reviewed. In practice, it is unlikely that the historical object identifiers and historical source identifiers of all the requester's historical review objects will be used to obtain the object to be reviewed. For example, historical review objects from too long an interval do not have guiding significance for reusing review experience. Therefore, computer equipment can perform certain filtering on the historical review objects used as the basis for obtaining the object to be reviewed.

[0088] Therefore, in this embodiment of the application, the computer device can determine the review time of each historical review object in the historical review index, and thus perform filtering based on the review time. Now, in conjunction with... Figure 3 To provide an explanation, in Figure 3 In this system, the first data record corresponds to the first historical review object reviewed by the requester, the m-th data record corresponds to the m-th historical review object reviewed by the requester, and the first to the m-th historical review objects represent m historical review objects with sequentially later review times. Therefore, the computer device can filter the m historical review objects based on their respective review times.

[0089] In this embodiment, the real-time condition can be set based on the proximity of the review time to the current time. It is understood that the closer the review time of a historical review object is to the current time, the deeper the review experience learned and the higher its retention rate when the requesting party reviews the historical review object, thus increasing the efficiency of reusing the review experience for subsequent reviews. Therefore, the real-time condition can be set based on the duration of the time gap between the historical review time and the current time. Specifically, the interval between the historical review time and the current time can be less than or equal to a preset time period, such as within one hour. This embodiment does not specifically limit this.

[0090] For example, if the current time is 9:00 AM on July 1, 2022, and the preset time period is from 8:00 AM to 9:00 AM on July 1, 2022, which is within one hour of the current time, the computer device can identify the historical audit objects whose audit time occurred between 8:00 AM and 9:00 AM on July 1, 2022 from the historical audit index, and obtain the historical object identifier and historical source identifier of these historical audit objects from the historical audit index.

[0091] In the above embodiments, since the determined historical review objects are real-time, the obtained historical object identifiers and historical source identifiers are also real-time, and consequently, the target review object obtained based on the historical object identifiers and historical source identifiers is also real-time. Therefore, the determined target review object can better align with the review experience learned by the requester during the review of historical review objects, thereby enabling the requester's review experience to be reused in a timely and effective manner, thus improving review efficiency.

[0092] In some embodiments, the target class index includes a first dimension index corresponding to the content clustering dimension, a second dimension index corresponding to the modal clustering dimension, and a third dimension index corresponding to the graph clustering dimension; searching for at least one target class index that includes historical object identifiers includes: searching for a first dimension index that includes historical object identifiers from a plurality of candidate indexes corresponding to the content clustering dimension; searching for a second dimension index that includes historical object identifiers from a plurality of candidate indexes corresponding to the modal clustering dimension; and searching for a third dimension index that includes historical object identifiers from a plurality of candidate indexes corresponding to the graph clustering dimension.

[0093] Specifically, the content clustering dimension mainly refers to the relationship between the first dimension index and the content of the object under review. For example, it could be related to the source material. This can be understood as grouping objects under review with the same or similar content together as data records for the first dimension index. The content type of the object under review can be related to the type of the object under review; this embodiment does not specifically limit this. Taking an advertisement as an example, the content type of the object under review can be text, video, or image.

[0094] Modality clustering dimension primarily refers to the correlation between the second-dimensional index and the modality exhibited by the object under review. It can be understood as grouping objects under review with the same or similar modalities together as data records for the second-dimensional index. A modality can be the review performance characteristic exhibited by the object under review during the review process; objects with similar or identical modalities should exhibit similar or identical review performance characteristics when reviewed. For example, whether the content of the object under review has been reviewed can be used as a review performance characteristic, and the type of reason for the content failing review can also be used as a review performance characteristic.

[0095] The graph clustering dimension mainly refers to the third-dimensional index and the graph correlation between the object under review and the content contained within it. This can be understood as clustering objects under review with similar or identical graphs together as data records for the third-dimensional index. The graph similarity between different objects can be represented by the similarity between the various contents contained within each object.

[0096] For example, if object A includes contents a, b, and c, object B includes contents c, d, and e, and object C includes contents b and d, the resulting graph topology can be referenced. Figure 6 The lines connecting object A to a, b, and c respectively can form a graph of object A, and the same applies to objects B and C. The degree of similarity between the graphs of object A and object B can be represented by the degree of similarity between the individual contents of object A and the individual contents of object B.

[0097] It should be noted that, in the embodiments of this application, the process of a computer device searching for at least one target class index that includes historical object identifiers involves a first-dimensional index, a second-dimensional index, and a third-dimensional index. It is understood that in actual implementation, only at least one of the above three dimensions of indexes may be involved, and this embodiment of the application does not specifically limit this. For example, the computer device may pre-construct multiple candidate indexes for the content clustering dimension based solely on historical object identifiers. Therefore, when searching for at least one target class index that includes historical object identifiers, it may only execute the scheme of "searching for the first-dimensional index that includes historical object identifiers from the multiple candidate indexes corresponding to the content clustering dimension," without using the second-dimensional index of the modal clustering dimension or the third-dimensional index of the graph clustering dimension.

[0098] Additionally, the structure of the target class index was mentioned earlier; please refer to [link / reference]. Figure 4 This means that it can include a search key and data records, and in this embodiment, the first-dimensional index, the second-dimensional index, and the third-dimensional index are all target-class indexes. Therefore, the structure of each of these three indexes can include a search key and data records. The search key included in each of these three indexes can be a historical object identifier, and the data records included in each can be constructed based on their respective clustering dimensions.

[0099] It should also be noted that in this embodiment, the first-dimensional index, the second-dimensional index, and the third-dimensional index are all obtained from multiple candidate indexes. The reason for multiple candidate indexes is that there can be many historical object identifiers; that is, they may not be limited to the historical object identifiers of the historical review objects corresponding to the requester of the object retrieval request, but may also include the historical object identifiers of historical review objects corresponding to other requesters. Regardless of the clustering dimension used, these historical object identifiers will generate multiple corresponding candidate indexes, and the target class index, which includes the historical object identifiers determined in step 204, may be among these multiple candidate indexes.

[0100] In the above embodiments, since three-dimensional indexes can be constructed using three clustering dimensions based on the historical object identifiers of the historical review objects corresponding to the requester, it is possible to find the review objects that match the requester's review experience from multiple dimensions, thereby improving the search coverage when searching for target review objects. Furthermore, by combining multiple dimensions and comprehensively considering them to determine the target review objects, the reviewer can maximize the reuse of review experience when reviewing the target review objects, thereby improving review efficiency.

[0101] In some embodiments, the method further includes a step of establishing multiple candidate indices for the content clustering dimension. This step specifically includes the following steps: obtaining objects to be reviewed and determining the content features of each object; obtaining multiple first reference objects; determining objects related to each first reference object based on the content features of each object, wherein the similarity between the content features of the objects related to the first reference objects and the content features of the first reference objects satisfies a first similarity condition; and constructing multiple candidate indices corresponding to the multiple first reference objects under the content clustering dimension based on the object identifiers of the objects related to each first reference object.

[0102] Specifically, as mentioned above, the content type of an object can be related to the object type. Therefore, in this embodiment, the content features of the object to be reviewed can be related to the type of the object to be reviewed. Taking an advertisement as an example, the content features of the object can be text features, video features, or image features. The first reference object can refer to a historically reviewed object that has already been reviewed. Introducing the first reference object mainly requires using the historical object identifier of the first reference object as the search primary key to construct a candidate index. The first reference object can be obtained by randomly selecting multiple historically reviewed objects, or by selecting historically reviewed objects whose review time is closer to the current time. This embodiment does not specifically limit this. In addition, the number of first reference objects can be multiple, so as to establish multiple candidate indexes.

[0103] For any given first reference object, the computer device can calculate the similarity between the content features of the first reference object and the content features of each object to be reviewed. Then, the objects to be reviewed that satisfy a first similarity condition are designated as objects to be reviewed related to the first reference object. The first similarity condition may include a similarity greater than a preset threshold or within a preset range. The similarity between content features can be represented by the similarity between content vectors. Taking an advertisement as an example, advertisements are typically composed of text, images, videos, and audio. Converting these data into vectors and calculating the similarity between the vectors of different advertisements yields the degree of similarity in content features between different advertisements.

[0104] It should be noted that the objects under review typically contain a lot of content and occupy a large amount of space. Therefore, when building a candidate index based on the first reference object, if the data records in the candidate index are directly linked to the object under review itself, it will result in excessive space consumption. Since the object identifier of the object under review is uniquely associated with the object under review, and the data volume of the object identifier is much smaller than that of the object itself, when building a candidate index based on the first reference object, combined with the previous explanation of the candidate index structure, the computer device can use the historical object identifier of the first reference object as the search primary key, and use the object identifiers of the objects under review related to the first reference object as data records to construct the candidate index corresponding to the first reference object. The above process mainly focuses on building a candidate index based on one first reference object. Multiple first reference objects can refer to the above process to build corresponding candidate indexes, thereby obtaining multiple candidate indexes for the content clustering dimension.

[0105] It should also be noted that when constructing the candidate index, different data records in the candidate index can be sorted. For example, the computer device can sort them from high to low according to the similarity between the calculated content features. Therefore, when the computer device subsequently executes step 206 to obtain the first object to be reviewed, it can obtain it according to the sorting generated when constructing the candidate index, such as obtaining the top 100 first objects to be reviewed. This embodiment of the application does not specifically limit this. Since the first objects to be reviewed that are ranked higher naturally align more closely with the historical review objects reviewed by the requester, the subsequent push of target objects to be reviewed can be more targeted, greatly improving the accuracy of distribution. Furthermore, because the push is more targeted, the reviewer's review experience can be reused in a timely and effective manner when reviewing target objects to be reviewed, thereby improving review efficiency.

[0106] In the above embodiments, since it is possible to determine the objects to be reviewed whose content features meet the first similarity condition, and to construct a candidate index based on the determined objects to be reviewed, it can be ensured that the obtained first objects to be reviewed have a high degree of similarity in content features to the historically reviewed objects reviewed by the requesting party. Therefore, thanks to a high degree of familiarity with the content features, the reviewer can reuse review experience to a greater extent, thereby improving review efficiency.

[0107] In some embodiments, the method further includes a step of constructing multiple candidate indices under the modal clustering dimension, which specifically includes the following steps: obtaining the objects to be processed for review and determining the elements included in each object; constructing a review collaboration matrix corresponding to each object based on the review behavior matrix and the elements included in each object, wherein the review behavior matrix represents the review status of different reviewers for different elements; obtaining multiple second reference objects; determining the objects to be reviewed that are related to each second reference object based on the review collaboration matrix of each object, wherein the similarity between the review collaboration matrix of the object and the review collaboration matrix of the second reference object satisfies a second similarity condition; and constructing multiple candidate indices corresponding to the multiple second reference objects under the modal clustering dimension based on the object identifiers of the objects to be reviewed that are related to each second reference object.

[0108] Specifically, the elements mentioned in the embodiments of this application can refer to the content entities of an object. Different elements may correspond to the same content type. For example, taking an advertisement as the object under review, the content type of the advertisement can be text, video, audio, or image. The elements included in the advertisement can be specific entities of text, video, or image, such as text1, text2, video1, video2, image1, and image2, etc. As another example, taking published content as the object under review, such as a book publication, the content type of the book publication can be image or text. The elements included in the book publication can be specific entities of text or image, such as text3, text4, image3, and image4, etc.

[0109] Considering that content entities are typically massive in number, and the computational load of subsequent processing may be related to the data volume of these content entities, in practice, computer equipment can also cluster the content entities, using these clusters as elements. When determining the elements included in the object to be reviewed, the computer equipment can determine the corresponding cluster type based on the content entities included in the object, thus using the cluster type as the elements included in the object. The clustering method can be based on the similarity between content entities. For example, text 1, text 2, and text 3 are all text; if their content is similar, they can be clustered into one category. By using the cluster type as the elements included in the object to be reviewed, the subsequent computational load can be reduced.

[0110] Next, the computer device can construct an audit behavior matrix. In some embodiments, the steps of constructing the audit behavior matrix include: obtaining historical audit records of different auditors, and determining the element range based on the elements included in each historical audit object in the historical audit records; constructing an initial matrix with different auditors as matrix rows and different elements in the element range as matrix columns; for each parameter in the initial matrix, determining the value of the corresponding parameter based on whether the auditor corresponding to the corresponding parameter has audited the historical object to which the element corresponding to the corresponding parameter belongs; and obtaining the audit behavior matrix based on the values ​​of each parameter in the initial matrix.

[0111] Specifically, determining the element range is primarily for defining the columns of the audit behavior matrix. In practice, all elements within the element range can be used as matrix columns. The reason for defining the element range is to ensure that each incoming audit target can be used to construct a corresponding audit collaboration matrix based on the audit behavior matrix. This requires that the elements covered in the audit behavior matrix encompass all elements included in the incoming audit target. To achieve this, computer equipment can exhaustively search for all possible elements included in the audit target by obtaining historical audit records from different auditing parties. Combining this with the above explanation regarding elements being cluster types, even if there are many elements, exhaustive searching could lead to a massive data volume; element clustering can reduce the data volume. Furthermore, even if the elements included in the incoming audit target have not been previously obtained from historical audit records, these unobtained elements can be converted into cluster types, and the cluster types can be used as elements, ensuring that the unobtained elements are covered by the audit behavior matrix.

[0112] The computer equipment constructs an initial matrix with different reviewers as rows and different elements within the element range as columns. At this point, the parameters at different positions in the initial matrix are not assigned values. The parameter in the a-th column and b-th row of the initial matrix represents the review status of the b-th reviewer for the a-th element. There are two review statuses: reviewed and not reviewed. The review status can be quantified numerically; for example, a value of 1 indicates reviewed, and a value of 0 indicates not reviewed. Specifically, taking a total of 4 reviewers and 4 elements as an example, the review behavior matrix can be shown in Table 1 below:

[0113] Table 1

[0114]

[0115]

[0116] In the above embodiments, since the review behavior matrix can be used to construct the review collaboration matrix corresponding to the object to be reviewed, and the review collaboration matrix can serve as the basic candidate index for modality clustering, subsequently, for historically reviewed objects reviewed by the requester, objects whose review collaboration matrices are similar to those of the historically reviewed objects can be obtained from the second-dimensional index and used as target objects to be reviewed. Therefore, thanks to a high degree of familiarity with the object content, the reviewer can reuse review experience to a large extent, thereby improving review efficiency.

[0117] After constructing the review behavior matrix according to the above process, the computer device can construct a review collaboration matrix corresponding to each review object based on the review behavior matrix and the elements included in each review object. Specifically, for any review object, the computer device can retain the parameter values ​​of the corresponding element column vectors in the review behavior matrix based on the elements included in the review object, while setting the parameter values ​​of the corresponding element column vectors of elements not included in the review behavior matrix to invalid values. For example, if a parameter value of 1 represents a reviewed object and 0 represents a not reviewed object, then an invalid value can also be 0. Referring to Table 1 above, taking the review object including elements a and b as an example, the review collaboration matrix corresponding to the review object can be as follows:

[0118]

[0119] Through the above process, each object to be reviewed can obtain a corresponding review collaboration matrix. Next, the computer device can acquire multiple second reference objects. The process of acquiring multiple second reference objects can refer to the previous process of acquiring multiple first reference objects. These multiple second reference objects can be the same as or different from the previous multiple first reference objects; this embodiment does not specifically limit this.

[0120] For any second reference object, the computer device can calculate the similarity between the review collaboration matrix of the second reference object and the review collaboration matrix of each object to be reviewed. Then, objects to be reviewed whose similarity satisfies the second similarity condition are considered as objects to be reviewed related to the second reference object. The similarity between review collaboration matrices can be represented by matrix similarity. The calculation process for matrix similarity can involve first converting the matrix into a vector, such as by concatenating the column vectors. Then, cosine similarity or Pearson similarity methods can be used to calculate the similarity between the vectors, which is then used as the matrix similarity. If the second reference object and each object to be reviewed are considered as a batch of objects, the pairwise similarity between objects in this batch can be referenced in Table 2 below.

[0121] Table 2

[0122] object c1 object c2 object c3 object c4 object c1 Cos<c1,c1> Cos<c1,c2> Cos<c1,c3> Cos<c1,c4> object c2 Cos<c2,c1> Cos<c2,c2> Cos<c2,c3> Cos<c2,c4> object c3 Cos<c3,c1> Cos<c3,c2> Cos<c3,c3> Cos<c3,c4> object c4 Cos<c4,c1> Cos<c4,c2> Cos<c4,c3> Cos<c4,c4>

[0123] It should be noted that, similarly, the objects under review typically contain a large amount of content and occupy a significant amount of space. Therefore, when constructing a candidate index based on this second reference object, the historical object identifier of the second reference object can be used as the search primary key, while the object identifiers of the objects under review related to the second reference object are used as data records to construct the candidate index corresponding to the second reference object. The above process mainly involves constructing a candidate index based on one of the first reference objects. Multiple first reference objects can refer to the above process to construct corresponding candidate indexes, thereby obtaining multiple candidate indexes in the modality clustering dimension.

[0124] It should also be noted that when constructing the candidate index, different data records within the candidate index can also be sorted. For example, the computer device can sort the data from highest to lowest based on the similarity between the calculated review collaboration matrices. Therefore, when the computer device subsequently executes step 206 to obtain the first object to be reviewed, it can obtain the object according to the sorting generated during the construction of the candidate index, such as obtaining the top 100 first objects to be reviewed. This embodiment of the application does not specifically limit this. Similarly, since the first objects to be reviewed that are ranked higher naturally align more closely with the historical review objects reviewed by the requester, the subsequent push of target objects to be reviewed can be more targeted, greatly improving the accuracy of distribution. Furthermore, because the push is more targeted, the reviewer's review experience can be reused in a timely and effective manner when reviewing target objects to be reviewed, thereby improving review efficiency.

[0125] In the above embodiments, since the review collaboration matrix can serve as a candidate index for modal clustering, subsequent review objects that have been reviewed by the requester can be identified in the second-dimensional index as target review objects with similar review collaboration matrices to those historical review objects. Therefore, thanks to a high degree of familiarity with the object content, the reviewer can reuse review experience to a greater extent, thereby improving review efficiency.

[0126] In some embodiments, the method further includes a step of establishing multiple candidate indices in the graph clustering dimension, which specifically includes the following steps: obtaining the objects to be processed and obtaining multiple third reference objects; performing graph clustering based on the elements included in the third reference objects and the elements included in each object to be processed to obtain objects to be processed that are associated with each third reference object; and constructing multiple candidate indices in the graph clustering dimension that correspond to the multiple third reference objects based on the object identifiers of the objects to be processed associated with each third reference object.

[0127] Specifically, the process of obtaining multiple third reference objects can refer to the previous process of obtaining multiple first reference objects. In actual implementation, any two of the first, second, and third reference objects can be the same or different, and this application embodiment does not specifically limit this. In some embodiments, the process of graph clustering processing by the computer device may include: for any third reference object and any object to be examined, performing pairwise similarity comparisons between the element features of the elements included in the object to be examined and the element features of the elements included in the third reference object, and obtaining the graph similarity between the object to be examined and the third reference object based on the results of the pairwise similarity comparisons. If the graph similarity is greater than a preset threshold, the object to be examined is regarded as an object to be examined related to the third reference object.

[0128] Elements can be transformed into element features. For example, text, images, or audio in an advertisement can be converted into feature vectors, which can then serve as element features. The similarity comparison between element features can be used to represent the similarity of feature vectors, such as cosine similarity. The graph similarity between the object under review and the third reference object can be calculated using the following formula:

[0129]

[0130] In the above formula, A can represent the object to be reviewed, and B can represent a third reference object. S(A,B) represents the graph similarity between A and B. I(A) represents the number of elements included in A, and I(B) represents the number of elements included in B. i (A) represents the i-th element included in A, I j (B) represents the j-th element included in B. c represents the constant decay factor, which can take values ​​from 0 to 1.

[0131] It should be noted that when constructing candidate indexes based on the third reference object, the historical object identifier of the third reference object can be used as the search primary key, while the object identifiers of the objects to be reviewed related to the third reference object are used as data records to construct the candidate indexes corresponding to the third reference object. The above process mainly focuses on constructing candidate indexes based on one third reference object. Multiple third reference objects can be constructed using the above process, thus obtaining multiple candidate indexes in the graph clustering dimension. Furthermore, when constructing candidate indexes, different data records within the candidate indexes can also be sorted. For example, computer devices can be sorted from high to low according to the calculated graph similarity.

[0132] In the above embodiments, since the reviewers have reviewed historical review objects that have been reviewed by the requesting party, subsequent review objects with graph similarity to those historical review objects can be obtained from the third-dimensional index and used as target review objects. Therefore, thanks to a high degree of familiarity with the object content, the reviewers can reuse their review experience to a large extent, thereby improving review efficiency.

[0133] In some embodiments, the index position of each object identifier in the target class index is determined based on the similarity between the object to be reviewed pointed to by each object identifier and the historical review object pointed to by the historical object identifier; obtaining the first object to be reviewed associated with each target class index includes: obtaining the object identifier at the index position in the target class index that meets the first preset position condition, and finding the corresponding first object to be reviewed based on the obtained object identifier.

[0134] Specifically, the object identifiers in the target class index can be sorted from high to low according to a certain criterion: the similarity between the object to be reviewed pointed to by the object identifier and the historical review object pointed to by the historical object identifier. Objects to be reviewed with higher similarity can be ranked higher in the target class index, while those with higher similarity can be ranked lower. The first preset position condition can refer to the object identifiers ranked higher, such as the object identifiers ranked in the top 100 of the index. Since each object identifier corresponds to an object to be reviewed, the computer device can find the object to be reviewed corresponding to the obtained object identifier and use it as the first object to be reviewed.

[0135] In the above embodiments, since the first candidate for review, ranked higher, is naturally more similar to previously reviewed candidates, subsequent push notifications of target candidates for review can be more targeted, greatly improving the accuracy of distribution. Furthermore, because the push notifications are more targeted, the reviewer's experience in reviewing target candidates can be reused promptly and effectively, thereby improving review efficiency.

[0136] In some embodiments, the historical source identifier includes a historical channel identifier and a historical source identifier. Searching for a target source index from multiple source indexes based on the historical source identifier includes: searching for a target source index corresponding to the historical channel identifier from multiple channel candidate indexes based on the historical channel identifier; and searching for a target source index corresponding to the source identifier from multiple source candidate indexes based on the historical source identifier.

[0137] Specifically, the historical channel identifier can be used to indicate which historical channel the historically reviewed object originated from. The channel can include the industry to which the object belongs and the creator of the object. Taking an advertisement as an example, the channel can include the industry to which the advertisement belongs and the advertising company. The historical source identifier can be used to indicate the source of the historically reviewed object, specifically referring to the object's owner, such as the advertiser.

[0138] Since historical source identifiers include historical channel identifiers and historical source identifiers, the source index can be divided into channel candidate indexes and source candidate indexes. It is understood that since there may be multiple historical channel identifiers and multiple historical source identifiers, there will also be multiple channel candidate indexes and multiple source candidate indexes. It should be noted that this embodiment uses both historical channel and historical source dimensions; in actual implementation, at least one dimension can be used to find the target source index, and this embodiment does not specifically limit this.

[0139] In the above embodiments, because the target objects to be reviewed pushed can match the historical review objects already reviewed by the requester in terms of both the object's channel and the object's source, the push can be tailored to the reviewer's capabilities and work habits, making it more targeted and greatly improving the accuracy of distribution. Furthermore, because the push is more targeted, the reviewer's review experience can be reused promptly and effectively when reviewing target objects, thereby improving review efficiency.

[0140] In some embodiments, the index position of each object identifier in the target source index is determined based on the review time of the object to be reviewed pointed to by each object identifier. Obtaining the second object to be reviewed associated with each target source index includes: obtaining the object identifier at the index position in the target source index that meets the second preset position condition, and finding the corresponding second object to be reviewed based on the obtained object identifier.

[0141] Specifically, the object identifiers in the target source index can be sorted from latest to earliest according to their review time. Objects with later review times can be ranked before those in the target class index, while objects with earlier review times can be ranked after. The second preset position condition can refer to the object identifiers ranked earlier, such as those in the first 100 positions of the index. Each object identifier corresponds to a review object, allowing the computer device to locate the review object corresponding to the obtained object identifier and use it as the second review object.

[0142] In the above embodiments, since the reviewer has a higher natural memory retention rate for the audited objects that were reviewed later, the reviewer can reuse the review experience with a higher retention rate when reviewing the target audited objects, thereby improving the review efficiency.

[0143] In some embodiments, determining and pushing target objects for review that match the requester based on a first object to be reviewed and a second object to be reviewed includes: determining a set of objects to be reviewed consisting of a first object to be reviewed and a second object to be reviewed, and determining the object characteristics of each object to be reviewed in the set of objects to be reviewed; obtaining the review behavior characteristics corresponding to the requester; sorting each object to be reviewed in the set of objects to be reviewed based on the review behavior characteristics and object characteristics to obtain target objects to be reviewed with a sorting order; and pushing the target objects to be reviewed with the sorting order.

[0144] Specifically, considering that the first and second objects to be reviewed may overlap, the computer equipment can perform deduplication when determining the set of objects to be reviewed, which consists of both. The object characteristics can be related to the object type. Taking an advertisement as an example, advertisement characteristics may include the advertisement bid or the advertisement duration. The requester's review behavior characteristics may include the requester's industry, the requester's area of ​​expertise, or the requester's average review time.

[0145] In some embodiments, the computer device sorts the objects in the set of objects to be reviewed based on review behavior characteristics and object characteristics. This may include: for any review object, inputting the review behavior characteristics and the object characteristics of the object to be reviewed into a sorting model, and outputting the matching degree between the review object and the requester; and sorting the objects in the set of objects to be reviewed according to the matching degree corresponding to each object. After the sorting process is completed, the computer device can push the target objects to be reviewed in the sorted order to the requester, and the requester can review the target objects to be reviewed in the sorted order. The sorting model may adopt a neural network structure model, and this embodiment does not specifically limit it.

[0146] In the above embodiments, since the objects to be reviewed in the set of objects to be reviewed can be sorted based on the characteristics of the requester's review behavior and the object characteristics of the review objects, the requester can prioritize reviewing the objects to be reviewed that have a higher degree of matching with itself. Prioritizing the review of objects to be reviewed that have a higher degree of matching can improve the efficiency of the initial review. At the same time, the efficient review process in the early stage can accumulate review experience for the subsequent objects to be reviewed that have a lower degree of matching, thereby improving the overall review efficiency.

[0147] In some embodiments, before pushing target objects to be reviewed in a sorted order, the method further adjusts the target objects to be reviewed in a sorted order by at least one of the following methods: adjusting the order of the target objects to be reviewed based on a preset priority rule; obtaining fingerprint information of each target object to be reviewed, and adjusting the order of the target objects to be reviewed based on the fingerprint information; pushing the target objects to be reviewed in a sorted order, including: pushing the adjusted target objects to be reviewed.

[0148] Specifically, the pre-defined priority rules are mainly aimed at certain target sources, such as advertisers with priority review rights. By using pre-defined priority rules, the review priority of the target objects pending review from these target sources can be increased, which means that the ranking of these target objects pending review can be advanced.

[0149] Fingerprint information can be used to represent the content similarity between various target review objects, allowing for deduplication of target review objects with identical fingerprint information. For example, if target review objects A and B, which have the same fingerprint information and are in a sorted order, are removed by the computer, such as removing A, A can subsequently reuse B's review result. Target review objects with high fingerprint similarity can be placed in adjacent sorting orders. Specifically, fingerprint information can be the binary code of the target review object stored in memory.

[0150] In the above embodiments, since the order of target review objects can be adjusted based on preset priority rules, specific priority review requirements can be met, expanding the applicable scenarios of the review process. Furthermore, since deduplication can be performed based on the fingerprint information of the target review objects, duplicate reviews can be avoided, improving review efficiency. Additionally, since the order of target review objects can be adjusted based on fingerprint information, target review objects with similar content can be placed in a proximity order, allowing the requester to review target review objects with similar content consecutively, thereby improving review efficiency.

[0151] In some embodiments, the step of sorting each object in the set of objects to be reviewed based on review behavior characteristics and object characteristics is performed by a target sorting model. The step of determining the target sorting model includes: obtaining multiple candidate sorting models, wherein each candidate sorting model is different from the others; obtaining the test object characteristics of each test object to be reviewed generated in the test phase, as well as the review behavior characteristics of the test party; performing sorting tests on the multiple candidate sorting models based on the test object characteristics and review behavior characteristics of each test object to be reviewed, and obtaining test results; and determining the target sorting model from the candidate sorting models based on the test results corresponding to each candidate sorting model.

[0152] Specifically, the multiple candidate ranking models can all be models using a neural network structure, but the structures of the multiple candidate ranking models can be different from each other. The test object characteristics of the test subjects and the review behavior characteristics of the test party can be referred to the explanation of the object characteristics of the test subjects and the review behavior characteristics of the requesting party in the preceding content. In some embodiments, the computer device performs ranking tests on the multiple candidate ranking models based on the test object characteristics and review behavior characteristics of each test subject. The test results may include: determining the designated test subjects corresponding to each candidate ranking model from the test subjects; combining the object characteristics of the designated test subjects corresponding to each candidate ranking model with the review behavior characteristics to form the input data for each candidate ranking model; processing the input data of each candidate ranking model and outputting prediction results; ranking the designated test subjects corresponding to each candidate ranking model according to the prediction results output by each candidate ranking model and submitting them for review, and obtaining the review time.

[0153] It is understood that there can be multiple test objects in this embodiment. The computer device determines the designated test object corresponding to each candidate ranking model from the test objects, mainly by distributing the multiple test objects based on the number of candidate ranking models. For example, taking two candidate ranking models, A and B, 50% of the multiple test objects can be distributed to model A, and the remaining 50% to model B. The distributed test objects are the designated test objects corresponding to the candidate ranking models.

[0154] For any candidate ranking model and its corresponding designated object to be reviewed, the computer device can combine the object features and review behavior features of the designated object to obtain input data for that object. Both object features and review behavior features can be feature vectors, and the combination of object features and review behavior features can be vector concatenation. Using the input data for the designated object to be reviewed through the candidate ranking model, the computer device can output the corresponding prediction result for that object. The prediction result can be the matching degree between the designated object and the test subject. For each candidate ranking model and its corresponding designated object to be reviewed, the computer device processes them using the above process to obtain the prediction result for each candidate ranking model for each designated object to be reviewed.

[0155] Understandably, the matching degree for different specified objects to be reviewed will vary. Therefore, the computer device can sort the specified objects to be reviewed for each candidate ranking model according to the matching degree, resulting in experimental objects to be reviewed with a ranking order for each candidate ranking model. The computer device can obtain the corresponding review time by sending the experimental objects to the testing party for review.

[0156] After obtaining the review time, the computer equipment can determine the target ranking model from the candidate ranking models based on the test results corresponding to each candidate ranking model. This process can specifically include: determining the candidate ranking model with the shortest review time, and using the candidate ranking model with the shortest review time as the target ranking model. For any given candidate ranking model, since there may be multiple specified objects to be reviewed for that candidate ranking model, there will be multiple review times for multiple specified objects to be reviewed. In practice, the computer equipment can take the average of the review times corresponding to multiple specified objects to be reviewed, and use this average as the review time for the candidate ranking model.

[0157] In addition, the process can specifically include: adjusting the flow of multiple test subjects to each candidate ranking model in multiple stages to obtain multiple flow adjustment results; determining the candidate ranking model with the shortest review time among all candidate ranking models under each flow adjustment result; and determining the target ranking model based on the candidate ranking model with the shortest review time under each flow adjustment result. The process of the computer equipment determining the target ranking model based on the candidate ranking model with the shortest review time under each flow adjustment result can include: counting the number of times each candidate ranking model appears as the candidate ranking model with the shortest review time under multiple flow adjustment results, and selecting the candidate ranking model with the most occurrences as the target ranking model.

[0158] For example, taking a candidate ranking model of 2, A and B, as an example: 50% of the multiple test subjects are assigned to A for processing, and the remaining 50% are assigned to B for processing, which is the first result of the allocation adjustment; 60% of the multiple test subjects are assigned to A for processing, and the remaining 40% are assigned to B for processing, which is the second result of the allocation adjustment; 80% of the multiple test subjects are assigned to A for processing, and the remaining 20% ​​are assigned to B for processing, which is the third result of the allocation adjustment.

[0159] Assume that under the first traffic splitting adjustment result, the candidate ranking model with the shortest review time is A. Under the second traffic splitting adjustment result, the candidate ranking model with the shortest review time is A. Under the third traffic splitting adjustment result, the candidate ranking model with the shortest review time is B. By counting the number of times A and B each appear as candidate ranking models with the shortest review time, we can determine that A will be the target ranking model.

[0160] It should be noted that the above process mainly involves selecting the target ranking model from various candidate ranking models using experimental subjects under review. Since the ranking model can be a neural network structure, it involves a training process. Specifically, the object features of the sample subjects under review and the review behavior features of the requester can be used as training samples, while the actual review time ranking of the sample subjects under review by the requester can be used as training labels to train the ranking model. Alternatively, in the actual application of the target ranking model to rank the subjects under review, the object features of the subjects under review and the review behavior features of the requester can also be used as training samples, while the actual review time ranking of the subjects under review by the requester can be used as training labels to iteratively train the target ranking model. This embodiment of the application does not specifically limit this approach.

[0161] In the above embodiments, since the sorting result of the sorting model for the objects to be reviewed may affect the review time, by sorting the candidate sorting models, a target sorting model that is beneficial to shortening the review time can be selected. Then, the target sorting model can be applied to sort the objects to be reviewed, which can shorten the review time and improve the review efficiency.

[0162] In a specific embodiment, such as Figure 7 As shown, a method for distributing objects to be reviewed is provided. Taking the object to be reviewed as an advertisement and the requester as a reviewer as an example, this method is applied to a computer device (the computer device can specifically be...). Figure 1 Taking a terminal or server as an example, the following steps are included:

[0163] Step 702: In response to the ad claiming request, locate the historical audit index that matches the auditor of the ad claiming request.

[0164] The historical review index can include a search key and data records. The search key of the historical review index uniquely corresponds to the reviewer, and can be a reviewer identifier, such as the reviewer's name (reviewer_name). Data records can be appended to the search key of the historical review index. Data records can be associated with historically reviewed advertisements reviewed by the corresponding reviewer. There can be multiple data records, and different data records can correspond to historically reviewed advertisements reviewed by the same reviewer at different times. Each data record can include a historical advertisement identifier and a historical source identifier. The historical source identifier can refer to the source of the advertisement and can include a historical channel identifier and a historical source party identifier. The historical channel identifier can include an advertising industry identifier and may also include the identifier of the advertising company that produced the advertisement, while the historical source party identifier can include the advertiser's entity identifier; this embodiment does not specifically limit this.

[0165] Step 704: Based on the historical review index, determine the historical ad identifier and historical source identifier of the historical reviewed ads corresponding to the reviewer.

[0166] Since the historical review index can include historical ad identifiers and historical source identifiers, in actual implementation, the computer equipment can directly obtain the historical ad identifiers and historical source identifiers of historically reviewed ads from the historical review index. Specifically, the historical ad identifiers and historical source identifiers determined in this step can be the historical ad identifiers and historical source identifiers of ads reviewed by the reviewer within the past hour.

[0167] Step 706: Locate at least one target class index that includes historical ad identifiers, and obtain the first ad to be reviewed associated with each target class index. The target class index is built based on the ad content.

[0168] Specifically, the target class index may include a first-dimensional index corresponding to the content clustering dimension, a second-dimensional index corresponding to the modal clustering dimension, and a third-dimensional index corresponding to the graph clustering dimension. The computer device can search for the first-dimensional index, second-dimensional index, and third-dimensional index that include historical ad identifiers to obtain the first ad to be reviewed associated with each index. The above three types of indexes can be obtained through the following process: The computer device can search for the first-dimensional index that includes historical ad identifiers from multiple candidate indexes corresponding to the content clustering dimension. It can also search for the second-dimensional index that includes historical ad identifiers from multiple candidate indexes corresponding to the modal clustering dimension. Finally, it can search for the third-dimensional index that includes historical ad identifiers from multiple candidate indexes corresponding to the graph clustering dimension.

[0169] The process of constructing multiple candidate indexes corresponding to the content clustering dimension may include: obtaining multiple approved first reference advertisements and obtaining advertisements to be processed; for each first reference advertisement, calculating the content similarity between the first reference advertisement and each advertisement to be processed; using the advertisement identifiers of advertisements to be processed with content similarity greater than a preset threshold as data records, and constructing candidate indexes with the advertisement identifiers of the first reference advertisements as the search primary key. Each first reference advertisement corresponds to one candidate index, and multiple first reference advertisements correspond to multiple candidate indexes.

[0170] The process of constructing multiple candidate indexes corresponding to the modality clustering dimension may include: obtaining multiple approved second reference advertisements; obtaining advertisements to be processed and determining the elements included in each advertisement; constructing an approval collaboration matrix corresponding to each advertisement based on the approval behavior matrix and the elements included in each advertisement, where the approval behavior matrix represents the approval status of different reviewers for different elements; calculating the similarity between the approval collaboration matrix of the first reference advertisement and each advertisement to be processed for each second reference advertisement; and using the advertisement identifiers of advertisements to be processed with similarity greater than a preset threshold as data records to construct candidate indexes with the advertisement identifiers of the second reference advertisements as the search primary key. Each second reference advertisement corresponds to one candidate index, and multiple second reference advertisements correspond to multiple candidate indexes.

[0171] The process of constructing multiple candidate indices corresponding to the graph clustering dimension may include: obtaining multiple approved third reference advertisements and obtaining advertisements to be processed; for each third reference advertisement, comparing the element features of the elements included in the third reference advertisement with the element features of the elements included in each advertisement to be processed pairwise, and obtaining the graph similarity between the third reference advertisement and each advertisement to be processed based on the results of the pairwise similarity comparison, and taking the advertisements to be processed with a graph similarity greater than a preset threshold as advertisements to be processed related to the third reference advertisement; and constructing multiple candidate indices corresponding to the multiple third reference advertisements under the graph clustering dimension based on the advertisement identifiers of the advertisements to be processed related to each third reference advertisement.

[0172] Step 708: Based on the historical source identifier, search for the target source index from multiple source indexes and obtain the second pending advertisement associated with each target source index.

[0173] Specifically, the historical source identifier can be an advertising industry identifier, an advertising company identifier, or an advertiser's entity identifier. Taking the advertising industry identifier as an example, computer equipment can search for a target source index that includes the advertising industry identifier among multiple source indexes and obtain the second pending advertisement associated with the target source index.

[0174] Step 710: Based on the first and second pending advertisements, determine the target pending advertisements that match the reviewers and push the target pending advertisements.

[0175] Considering the potential overlap between the first and second pending advertisements, the computer device can determine a set of pending advertisements consisting of the first and second pending advertisements. The computer device can then select the pending advertisements in this set as target pending advertisements matched with the reviewers and push them to the reviewers. During the push process, the computer device can sort the pending advertisements in the set according to their advertising characteristics (such as advertisement duration or bid price) and the reviewers' review behavior characteristics (such as industry affiliation), and push the target pending advertisements to the reviewers based on the sorting results.

[0176] In the above embodiments, regarding the historical object identifier and historical source identifier of the historical review object corresponding to the requester, since the target review object is obtained from the target class index including the historical object identifier and the target source index including the historical source identifier, the obtained target review object can match the historical review objects already reviewed by the requester in different dimensions such as object source or object content. This allows for the push of information that aligns with the reviewer's capabilities and work habits, making it more targeted and significantly improving the accuracy of distribution. Furthermore, because the push is more targeted, the reviewer's review experience can be reused promptly and effectively when reviewing the target review object, thereby improving review efficiency.

[0177] This application also provides an application scenario that utilizes the aforementioned object distribution method, primarily focusing on the application process of the objects to be reviewed. See below for reference. Figure 8 Taking the advertisement as the target, the reviewer as the requester, and the server as the execution entity as an example, this paper provides a detailed explanation of the object distribution method for this application:

[0178] like Figure 8 As shown, the method for distributing pending advertisements in this application can be considered to include two main modules: an offline link and an online link. The offline link, located on the right, primarily utilizes big data and artificial intelligence technology to calculate the ranking index of the pending advertisements based on the characteristics of advertisements already reviewed by auditors and the characteristics of the advertisements to be reviewed. The online link, located on the left, primarily retrieves the pending advertisements from the ranking index calculated by the offline link based on the characteristics of advertisements already reviewed by auditors who need to receive the advertisements, and then distributes them to auditors for review, thereby optimizing the efficiency of auditors' work. Both the offline and online links can be implemented using servers; in practice, they can be different servers, the same server, or a server cluster.

[0179] The offline process is explained in detail below. "Order entry service" refers to advertisements awaiting review as task orders, which can be generated by the reviewer through their review terminal. "Stream computing" mainly refers to a real-time stream computing engine, which can determine the corresponding sorting index for advertisements arriving for review within 30 seconds. "Batch computing," on the other hand, refers to a non-real-time batch computing engine, with a latency of less than 1 hour. The reason for using two types of computing is primarily due to the source index in the constructed sorting index, for example... Figure 8 The same advertiser index and vertical channel index, where the vertical channel and source are used as the corresponding search primary keys, can be predetermined and usually do not change. Once an ad awaiting review arrives, its ad identifier (corresponding to...) can be directly applied based on its vertical channel or source. Figure 8 The tid in the index is attached to the corresponding index.

[0180] Instead of real-time batch calculation, it requires selecting the ad identifiers of ads already reviewed by the reviewers as the search primary key in real time, and building a sorted index based on the similarity of ad content between the reviewed ads and the subsequent batch of ads awaiting review. Figure 8 The system includes similar ad indexes, creative collaborative indexes, and graph similarity indexes. When a new batch of ads awaiting review arrives, the server can select the ad identifiers of ads already reviewed by reviewers as the search primary key in real time and rebuild the sorting index. Therefore, after each determination of the search primary key, one hour of real-time pending review data can be accumulated for building the sorting index.

[0181] Since similarity indexing involves the similarity of content features between ads, creative collaborative indexing involves the similarity of review collaboration matrices between ads, and graph similarity indexing involves the similarity of elements included in ads, and these similarities involve vector calculations, all of these calculation processes can be performed using... Figure 8 The AI ​​service's computing engine is implemented within this framework. The results of stream computing can be termed real-time features, which indicate which index—either the vertical channel index or the same advertiser index—the ad identifier for the pending advertisement can be placed in.

[0182] The results of batch computation can be called offline features. Offline features can indicate which index among the similar ad index, creative collaborative index, and graph similarity index the ad identifier of the ad to be reviewed can be placed in. Real-time features and offline features can be stored in a database, and the storage format can be referenced. Figure 8 The index structure in the database, and the index building module can be used to build... Figure 8 The index structure in the database. Figure 8The reviewer history index in the middle corresponds to the historical review index mentioned earlier, used to record the advertising identifiers and source identifiers that the reviewer has reviewed in the past. Furthermore, the algorithm strategies involved in the offline process can all be implemented by algorithm strategy personnel.

[0183] To facilitate understanding of the present... Figure 8 The various index structures are explained below. The vertical channel index can be considered as the channel-related source index mentioned above. The same advertiser index can be considered as the source index related to the source mentioned above. The similar ad index can be considered as the candidate index corresponding to the content clustering dimension mentioned above. The creative collaboration index can be considered as the candidate index corresponding to the modal clustering dimension mentioned above. The graph similarity index can be considered as the candidate index corresponding to the graph clustering dimension mentioned above.

[0184] The vertical channel index uses the vertical channel identifier (id) as its search primary key. The index is organized as an inverted index, allowing the server to retrieve pending advertisements under that vertical channel based on the vertical channel ID. A vertical channel can refer to a specific industry, a specific advertiser entity, or a specific advertiser group, etc. For the same advertiser index, the advertiser identifier (e.g., an advertiser account identifier can be represented by uid) is used as the search primary key. The index is also organized as an inverted index, allowing retrieval of all pending advertisements under that uid.

[0185] The similar ad index uses the ad identifier (tid) as its search key and is organized as an inverted index. The server can use historically approved ad identifiers to find pending ads that are similar in content to those already approved. The creative collaborative index also uses the ad identifier (tid) as its search key and is organized as an inverted index. Pending ads with the same modality are clustered within the same index. When a tid is input, the server can retrieve all tids within the current cluster. Pending ads within the same cluster exhibit similar behavior from reviewers. For example, ads within the same cluster may be rejected for similar reasons or all may have been approved.

[0186] The graph similarity index uses the ad identifier (tid) as its search key and is organized as an inverted index. Through a pre-defined graph clustering algorithm, the server can identify ads that are similar to other ads in the second-order graph based on the reviewer's behavior patterns. Therefore, the server can find the corresponding graph-similar ads based on their tid.

[0187] In addition to the index used for clustering pending advertisements, the index building module can also build a historical review index for reviewers. The primary search key is the reviewer's name (reviewer_name), and the index is organized as an inverted index. The server can retrieve the advertisement identifiers and source identifiers of some advertisements reviewed by the current reviewer based on their name (reviewer_name). Furthermore, the advertisement identifiers of pending advertisements within the index can also be sorted, such as by the advertisement's creation time.

[0188] After building the sorting index through the offline link on the right, the review order request (that is, the object retrieval request mentioned above) can be processed through the online link on the left. In other words, the materials to be reviewed can be efficiently distributed according to the characteristics of the reviewers and the characteristics of the advertisements to be reviewed in the current review queue.

[0189] Specifically, after a reviewer issues an order request, the server, based on the reviewer's name (reviewer_name), uses the reviewer vector recall module to determine the ad identifier, vertical channel identifier, and advertiser identifier of the ads reviewed by that reviewer within the current hour from the reviewer's historical review index in the database. After obtaining the ad identifier, vertical channel identifier, and advertiser identifier, the server can use the reviewer vector recall module to perform multi-path recall on several types of indexes previously generated in the offline process. Multi-path recall is mainly divided into the following categories: UID recall, similarity recall (including content similarity, material collaboration similarity, and graph similarity), and channel recall, as detailed below:

[0190] The server can use historical channel identifiers to query pending advertisements in the same vertical channel and distribute them to the reviewer. Since the reviewer is familiar with the industry review rules and conditions for the same vertical channel, they can quickly find the key content that needs to be reviewed. Therefore, by distributing pending advertisements in the same vertical channel as those previously reviewed by the reviewer, the review efficiency can be greatly improved.

[0191] The server can also query the same advertiser index with the same UID as the advertiser identifier of the ads previously reviewed by the reviewer, thereby retrieving the ads to be reviewed corresponding to the advertiser identifier in the same advertiser index and distributing the retrieved ads to the reviewer. Since ad creatives with the same advertiser identifier share commonalities and have the same memory characteristic for the same reviewer, distributing ads with the same advertiser identifier as ads previously reviewed by the reviewer to that reviewer can improve review efficiency.

[0192] Furthermore, the server can use the ad identifier (tid) of an ad reviewed by the reviewer to query the index of similar ads with that ad identifier as the primary key, thereby distributing the ads to be reviewed corresponding to that ad identifier in the found similar ad index to the reviewer. Since ads with similar content have similar styles and have a memory property of being reviewed by the same reviewer, distributing ads with similar content to the same reviewer can improve review efficiency.

[0193] Furthermore, the server can use the ad identifier (tid) of an ad reviewed by the reviewer to query the creative collaborative index, thereby distributing the ad to be reviewed corresponding to the ad identifier in the searched creative collaborative index to that reviewer. Since a reviewer's review behavior can be used as an implicit feature, this implicit feature can be effectively extracted through collaborative processing. This allows ad to be reviewed that is similar in implicit feature to ads previously reviewed by that reviewer to be distributed to that reviewer, thus improving the reviewer's review efficiency.

[0194] Furthermore, the server can use the ad identifier (tid) of an ad reviewed by the same reviewer to query the graph similarity index, thereby distributing the ads to be reviewed corresponding to the ad identifiers in the graph similarity index to that reviewer. Since graph vectors have second-order memory properties and can express the second-degree relationship between two ads, distributing ads to be reviewed that have a second-degree relationship with ads reviewed by the same reviewer to the same reviewer can improve the review efficiency.

[0195] After the server obtains the ads to be reviewed through the aforementioned multi-path recall, considering the possibility of overlapping ads in different indexes, it can deduplicate the recalled ads. Next, the server can sort the recalled ads using a target ranking model, distributing them to the respective reviewers' terminals according to the ranking, allowing reviewers to review them sequentially. Furthermore, before distributing to reviewers, the server can be adjusted through operational mechanisms, such as prioritizing and fingerprinting the ads. Prioritizing can mean advancing the ranking of ads from advertisers with review priority. Fingerprint aggregation can deduplicate ads with identical or similar content, allowing the review results to be reused; it can also group ads with similar content together, enabling reviewers to review multiple ads with similar content consecutively, thus improving review efficiency.

[0196] After the order receipt results are output, the auditors can begin their review process. The review time can be used as training data for the target ranking model, while the advertising characteristics of the reviewed advertisements and the reviewer's review behavior can be used as training samples to further train the target ranking model, ensuring that the ranking results generated by the model are more conducive to improving the reviewer's efficiency.

[0197] In addition, before determining which ranking model to use for ranking the pending ads in multi-channel recall, the target ranking model can be determined from multiple candidate ranking models through an experimental process. For details, please refer to [link / reference needed]. Figure 9 .exist Figure 9 In the offline process, the processing remains the same, primarily providing the data foundation for ad recall. In the online process, after recalling multiple trial ads awaiting review, these ads can be ranked using different candidate ranking models (e.g., different structures) to obtain the ranking effect of each model (which could be the review time of the trial reviewers after ranking). By executing multiple trials, each with a different proportion of ad traffic allocated to each candidate ranking model, and comparing the overall review time of each model across multiple trials, the candidate ranking model with the shortest review time can be selected as the target model. Finally, the computer system can determine the target ranking model through the trials and generate trial reports for product personnel to reference, allowing for redesign of the ranking model or other optimizations to the entire distribution process.

[0198] It should be noted that the method for distributing pending publications in this application can also be applied to other scenarios, such as the review of published content. Reviewers can initiate a request to retrieve published content through a review terminal. The server responds to the request by searching for historical review indexes that match the requester. Based on these historical review indexes, the server determines the historical publication content identifier and historical source identifier corresponding to the requester. The server searches for at least one target class index containing the historical publication content identifier and retrieves the first pending publication content associated with each target class index. Based on the historical source identifier, the server searches for target source indexes from multiple source indexes and retrieves the second pending publication content associated with each target source index. Based on the first and second pending publication contents, the server determines the target pending publication content matching the reviewer and pushes it to the review terminal. Accordingly, the target class index can still include a first-dimensional index corresponding to the content clustering dimension, a second-dimensional index corresponding to the modality clustering dimension, and a third-dimensional index corresponding to the graph clustering dimension. The target source index can still include a channel index and a source index. The channel index for published content can be related to the industry or region of the published content, without limitation. The source index can be based on the creator of the published content.

[0199] Of course, the method for distributing objects under review provided in this application is not limited to the application scenarios described above; it can also be used for reviewing videos, audio files, or documents. The above scenarios are only used to illustrate the method for distributing objects under review in this application and are not intended to limit the application scenarios of this application.

[0200] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0201] Based on the same inventive concept, this application also provides a device for distributing objects under examination to implement the above-described method for distributing objects under examination. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the object distribution device provided below can be found in the limitations of the object distribution method described above, and will not be repeated here.

[0202] In one embodiment, such as Figure 10 As shown, a device 1000 for distributing objects to be reviewed is provided, including: a search module 1002, a first determination module 1002, and a second determination module 1004, wherein:

[0203] The lookup module 1002 is used to look up historical audit indexes that match the requester of the object retrieval request in response to an object retrieval request.

[0204] The first determination module 1004 is used to determine the historical object identifier and historical source identifier of the historical audit object corresponding to the requester based on the historical audit index.

[0205] The lookup module 1002 is also used to look up at least one target class index that includes a historical object identifier, and to obtain the first object to be examined associated with each target class index. The target class index is built based on the object content.

[0206] The lookup module 1002 is also used to look up the target source index from multiple source indexes based on the historical source identifier, and obtain the second object to be examined associated with each target source index.

[0207] The second determination module 1004 is used to determine the target object to be reviewed that matches the requester based on the first object to be reviewed and the second object to be reviewed, and to push the target object to be reviewed.

[0208] In one embodiment, the first determining module 1004 is specifically used to determine historical review objects whose review time occurred within a preset time period from the historical review index, wherein the preset time period meets the real-time condition; and to obtain the historical object identifier and historical source identifier of the determined historical review object.

[0209] In one embodiment, the target class index includes a first dimension index corresponding to the content clustering dimension, a second dimension index corresponding to the modal clustering dimension, and a third dimension index corresponding to the graph clustering dimension; the search module 1002 is specifically used to search for the first dimension index containing historical object identifiers from the multiple candidate indexes corresponding to the content clustering dimension; to search for the second dimension index containing historical object identifiers from the multiple candidate indexes corresponding to the modal clustering dimension; and to search for the third dimension index containing historical object identifiers from the multiple candidate indexes corresponding to the graph clustering dimension.

[0210] In one embodiment, the device further includes an index building module for acquiring objects to be reviewed and determining the content features of each object to be reviewed; acquiring multiple first reference objects; determining objects to be reviewed that are related to each first reference object based on the content features of each object to be reviewed, wherein the similarity between the content features of the objects to be reviewed and the content features of the first reference objects satisfies a first similarity condition; and constructing multiple candidate indices corresponding to the multiple first reference objects under the content clustering dimension based on the object identifiers of the objects to be reviewed that are related to each first reference object.

[0211] In one embodiment, the index building module is further configured to: acquire objects to be processed for review and determine the elements included in each object; construct a review collaboration matrix corresponding to each object based on the review behavior matrix and the elements included in each object, wherein the review behavior matrix represents the review status of different reviewers for different elements; acquire multiple second reference objects; determine objects to be reviewed that are related to each second reference object based on the review collaboration matrix of each object, wherein the similarity between the review collaboration matrix of the object and the review collaboration matrix of the second reference object satisfies a second similarity condition; and construct multiple candidate indices corresponding to the multiple second reference objects under the modality clustering dimension based on the object identifiers of the objects to be reviewed that are related to each second reference object.

[0212] In one embodiment, the index building module is further configured to obtain historical audit records of different auditors, and determine the element range based on the elements included in each historical audit object in the historical audit records; construct an initial matrix with different auditors as matrix rows and different elements in the element range as matrix columns; for each parameter in the initial matrix, determine the value of the corresponding parameter based on whether the auditor corresponding to the corresponding parameter has audited the historical object to which the element corresponding to the corresponding parameter belongs; and obtain the audit behavior matrix based on the values ​​of each parameter in the initial matrix.

[0213] In one embodiment, the index building module is further configured to obtain the objects to be processed and obtain multiple third reference objects; perform graph clustering based on the elements included in the third reference objects and the elements included in each object to be processed to obtain objects to be processed that are associated with each third reference object; and construct multiple candidate indexes corresponding to the multiple third reference objects in the graph clustering dimension based on the object identifiers of the objects to be processed associated with each third reference object.

[0214] In one embodiment, the index position of each object identifier in the target class index is determined based on the similarity between the object to be reviewed pointed to by each object identifier and the historical review object pointed to by the historical object identifier; the search module 1002 is also used to obtain the object identifier at the index position in the target class index that meets the first preset position condition, and find the corresponding first object to be reviewed based on the obtained object identifier.

[0215] In one embodiment, the historical source identifier includes a historical channel identifier and a historical source identifier; the lookup module 1002 is further configured to, based on the historical channel identifier, search for a target source index corresponding to the historical channel identifier from multiple channel candidate indexes; and based on the historical source identifier, search for a target source index corresponding to the source identifier from multiple source candidate indexes.

[0216] In one embodiment, the index position of each object identifier in the target source index is determined based on the review time of the object to be reviewed pointed to by each object identifier; the search module 1002 is also used to obtain the object identifier at the index position in the target source index that meets the second preset position condition, and find the corresponding second object to be reviewed based on the obtained object identifier.

[0217] In one embodiment, the second determining module 1004 is further configured to determine a set of objects to be reviewed consisting of a first object to be reviewed and a second object to be reviewed, and to determine the object characteristics of each object to be reviewed in the set of objects to be reviewed; to obtain the review behavior characteristics corresponding to the requester; to sort each object to be reviewed in the set of objects to be reviewed based on the review behavior characteristics and the object characteristics, to obtain target objects to be reviewed with a sorting order; and to push the target objects to be reviewed with a sorting order.

[0218] In one embodiment, the second determining module 1004 is further configured to adjust the order of the target objects to be reviewed based on a preset priority rule before pushing the target objects to be reviewed with a sorting order; obtain the fingerprint information of each target object to be reviewed, and adjust the order of the target objects to be reviewed by deduplication or sorting based on the fingerprint information; and push the adjusted target objects to be reviewed.

[0219] In one embodiment, the step of sorting each object in the set of objects to be reviewed based on review behavior characteristics and object characteristics is performed by a target sorting model. The device also includes a model determination module for acquiring multiple candidate sorting models, wherein each candidate sorting model is different from the others; acquiring the test object characteristics of each test object to be reviewed generated in the test phase, as well as the review behavior characteristics of the test party; performing sorting tests on the multiple candidate sorting models based on the test object characteristics and review behavior characteristics of each test object to be reviewed, and obtaining test results; and determining the target sorting model from the candidate sorting models based on the test results corresponding to each candidate sorting model.

[0220] In one embodiment, the test results include review time; the model determination module is further configured to determine the designated review objects corresponding to each candidate ranking model from the test review objects; combine the object features of the designated review objects corresponding to each candidate ranking model with the review behavior features to form the input data of each candidate ranking model; process the input data of each candidate ranking model and output the prediction results; and rank the designated review objects corresponding to each candidate ranking model according to the prediction results output by each candidate ranking model and submit them for review, and obtain the review time.

[0221] The aforementioned object distribution device, targeting the historical object identifier and historical source identifier of the historical review object corresponding to the requester, obtains the target object for review from the target class index containing the historical object identifier and the target source index containing the historical source identifier. This allows the obtained target object for review to align with the requester's previously reviewed historical review objects in different dimensions, such as object source or object content. Consequently, the device pushes objects that match the reviewer's skill preferences and work habits, making it more targeted and significantly improving the accuracy of distribution. Furthermore, this more targeted approach allows reviewers to promptly and effectively reuse their review experience when reviewing target objects for review, thereby improving review efficiency.

[0222] Each module in the aforementioned object distribution device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0223] In one embodiment, a computer device is provided, which may be a terminal or a server, and its internal structure diagram may be as follows: Figure 11As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores object data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a method for distributing pending objects.

[0224] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0225] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0226] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0227] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0228] 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, data stored, data displayed, 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 the relevant data shall comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0229] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0230] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0231] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for distributing objects to be reviewed, characterized in that, The method includes: In response to an object claiming request, search the historical audit index that matches the requester of the object claiming request; Based on the historical review index, determine the historical object identifier and historical source identifier of the historical review object corresponding to the requester; Locate at least one target class index that includes the identifier of the historical object, and obtain the first object to be reviewed associated with each target class index, wherein the target class index is established based on the similarity between objects; Based on the historical source identifier, the target source index is searched from multiple source indexes, and the second object to be reviewed associated with each target source index is obtained; Based on the first and second objects to be reviewed, a target object to be reviewed that matches the requester is determined and the target object to be reviewed is pushed to the requester.

2. The method according to claim 1, characterized in that, The step of determining the historical object identifier and historical source identifier of the historical review object corresponding to the requester based on the historical review index includes: From the historical review index, determine the historical review objects whose review time occurred within a preset time period, where the preset time period meets the real-time requirement; Obtain the historical object identifier and historical source identifier of the identified historical audit object.

3. The method according to claim 1, characterized in that, The target class index includes a first-dimensional index corresponding to the content clustering dimension, a second-dimensional index corresponding to the modal clustering dimension, and a third-dimensional index corresponding to the graph clustering dimension; The search includes at least one target class index with the identifier of the historical object, including: From multiple candidate indexes corresponding to the content clustering dimension, find the first dimension index that includes the historical object identifier; From multiple candidate indices corresponding to the modality clustering dimension, find the second dimension index that includes the historical object identifier; From multiple candidate indices corresponding to the graph clustering dimension, find the third-dimensional index that includes the identifier of the historical object.

4. The method according to claim 3, characterized in that, The method further includes: Obtain the objects to be reviewed and determine the content characteristics of each object. Obtain multiple first reference objects; Based on the content characteristics of each object to be reviewed, objects to be reviewed that are related to each first reference object are determined, wherein the content characteristics of the object to be reviewed that are related to the first reference object represent the degree of similarity between the content characteristics of the object to be reviewed and the content characteristics of the first reference object, which satisfies the first similarity condition. Based on the object identifier of the object to be reviewed associated with each first reference object, multiple candidate indexes are constructed under the content clustering dimension, corresponding to the multiple first reference objects respectively.

5. The method according to claim 3, characterized in that, The method further includes: Obtain the objects to be reviewed and determine the elements included in each object; Based on the audit behavior matrix and the elements included in each audit object, an audit collaboration matrix corresponding to each audit object is constructed. The audit behavior matrix represents the audit status of different audit parties for different elements. Obtain multiple second reference objects; Based on the review collaboration matrix of each subject to review, the subject to review associated with each second reference subject is determined, wherein the similarity between the review collaboration matrix of the subject to review associated with the second reference subject and the review collaboration matrix of the second reference subject satisfies the second similarity condition. Based on the object identifier of the object to be examined associated with each second reference object, multiple candidate indices corresponding to the multiple second reference objects are constructed under the modality clustering dimension.

6. The method according to claim 5, characterized in that, The steps for constructing the audit behavior matrix include: Obtain historical audit records from different auditing parties, and determine the element range based on the elements included in each historical audit object in the historical audit records; Construct an initial matrix by using different reviewers as matrix rows and different elements within the specified element range as matrix columns; For each parameter in the initial matrix, the value of the corresponding parameter is determined based on whether the reviewer corresponding to the corresponding parameter has reviewed the historical object to which the element corresponding to the corresponding parameter belongs; The audit behavior matrix is ​​obtained based on the values ​​of each parameter in the initial matrix.

7. The method according to claim 3, characterized in that, The method further includes: Retrieve pending review objects and multiple third-party reference objects; Based on the elements included in the third reference object and the elements included in each object to be reviewed, graph clustering is performed to obtain the objects to be reviewed that are respectively related to each third reference object. Based on the object identifier of the object to be reviewed associated with each third reference object, multiple candidate indices are constructed under the graph clustering dimension, corresponding to the multiple third reference objects respectively.

8. The method according to claim 1, characterized in that, The index position of each object identifier in the target class index is determined based on the degree of similarity between the object to be reviewed pointed to by each object identifier and the historical review object pointed to by the historical object identifier. The step of obtaining the first object to be reviewed associated with each of the target class indices includes: Obtain the object identifier at the index position that satisfies the first preset position condition in the target class index, and find the corresponding first object to be reviewed based on the obtained object identifier.

9. The method according to claim 1, characterized in that, The historical source identifier includes a historical channel identifier and a historical source identifier. The step of searching for the target source index from multiple source indexes based on the historical source identifier includes: Based on the historical channel identifier, the target source index corresponding to the historical channel identifier is searched from multiple channel candidate indexes; Based on the historical source identifier, the target source index corresponding to the source identifier is searched from multiple source candidate indexes.

10. The method according to claim 1, characterized in that, The index position of each object identifier in the target source index is determined based on the review time of the object to be reviewed pointed to by each object identifier. Obtaining the second object to be reviewed associated with each of the target source indices includes: Obtain the object identifier at the index position that satisfies the second preset position condition in the target source index, and find the corresponding second object to be reviewed based on the obtained object identifier.

11. The method according to any one of claims 1 to 10, characterized in that, The step of determining a target candidate for review that matches the requester based on the first candidate for review and the second candidate for review, and then pushing the target candidate for review, includes: Determine a set of objects to be reviewed consisting of the first object to be reviewed and the second object to be reviewed, and determine the object characteristics of each object to be reviewed in the set of objects to be reviewed; Obtain the review behavior characteristics corresponding to the requester; Based on the audit behavior characteristics and the object characteristics, each object to be audited in the set of objects to be audited is sorted to obtain target objects to be audited with a sorting order. The target objects to be reviewed, which are sorted in order, are pushed to the system.

12. The method according to claim 11, characterized in that, Before pushing the target objects to be reviewed with the sorted order, the method further adjusts the target objects to be reviewed with the sorted order by at least one of the following methods: The order of the target objects to be reviewed, which have a sorting order, is adjusted based on a preset priority rule. Obtain fingerprint information of each target candidate for review, and perform deduplication or sorting adjustment on the target candidates for review with sorting order based on the fingerprint information; The process of pushing the target review objects with a sorted order includes: The revised target candidates for review are pushed out.

13. The method according to claim 11, characterized in that, The step of sorting the objects in the set of objects to be reviewed based on the review behavior characteristics and the object characteristics is performed by a target sorting model. The step of determining the target sorting model includes: Obtain multiple candidate ranking models, where each candidate ranking model is different from the others; To obtain the characteristics of each test subject to review generated during the trial phase, as well as the review behavior characteristics of the trial party; Based on the characteristics of the test subjects and the characteristics of the review behavior of each test subject, the multiple candidate ranking models are ranked and tested to obtain the test results. Based on the test results corresponding to each candidate ranking model, the target ranking model is determined from the candidate ranking models.

14. The method according to claim 13, characterized in that, The test results include review time. Based on the characteristics of the test subjects and the characteristics of the review behavior of each test subject, the multiple candidate ranking models are ranked and tested respectively to obtain the test results, including: From the experimental candidates, determine the designated candidates for review that correspond to each candidate ranking model; The object features of the specified object to be reviewed corresponding to each candidate ranking model are combined with the review behavior features to form the input data of each candidate ranking model. Each candidate ranking model processes its own input data and outputs prediction results. Based on the prediction results output by each candidate ranking model, the specified objects to be reviewed corresponding to each candidate ranking model are ranked and submitted for review, and the review time is obtained.

15. A device for distributing objects to be reviewed, characterized in that, The device includes: The lookup module is used to respond to an object claiming request and find the historical audit index that matches the requester of the object claiming request. The first determining module is used to determine the historical object identifier and historical source identifier of the historical review object corresponding to the requester based on the historical review index; The search module is also used to search for at least one target class index that includes the identifier of the historical object, and to obtain the first object to be reviewed associated with each target class index, wherein the target class index is established based on the object content; The search module is also used to search for a target source index from multiple source indexes based on the historical source identifier, and to obtain the second object to be reviewed associated with each target source index; The second determining module is used to determine a target object for review that matches the requester based on the first object for review and the second object for review, and to push the target object for review.

16. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 14.

17. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 14.

18. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 14.

Images