Data processing method, device and equipment and computer readable storage medium
By extracting and integrating the game UGC content, calculating the audit pass rate and allocating resources, the problem of low audit efficiency in the existing technology is solved, and more efficient content review and resource utilization is achieved.
Patent Information
- Application Number
- CN202310608874.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-26
- Publication Date
- 2024-11-26
AI Technical Summary
In the prior art, the review efficiency of game UGC content is low and it is difficult to accurately identify non-compliant content, resulting in the need of a large number of manual review resources, which increases the burden and inefficiency.
By performing feature extraction processing on the data to be reviewed, the feature information of different types of data is integrated, the audit pass rate is calculated, and the audit resources are allocated based on the pass rate to improve the audit efficiency.
It improves the accuracy and efficiency of content review, optimizes resource allocation, reduces the burden of manual review, and improves the compliance of platform content.
Smart Images

Figure CN119015714A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of computer technology, and in particular to a data processing method, apparatus, device and computer-readable storage medium. Background Art
[0002] With the rise of social media, online forums, and e-commerce platforms, users can freely share various forms of content, such as text, pictures, videos, etc., which are called user-generated content (UGC). In the field of gaming, the popularity of game UGC content is growing, attracting a wide audience. However, the quality of game UGC content varies, especially under the complexity of game graphics, it is difficult to accurately identify non-compliant content.
[0003] In related technologies, in order to ensure the compliance of the game UGC content on the platform, it is usually necessary to arrange a large number of manual review resources for review. This method not only increases the burden of manual review, but also has low review efficiency.
[0004] Therefore, how to improve the efficiency of content review is an urgent problem to be solved. Summary of the invention
[0005] To solve the above technical problems, the embodiments of the present application provide a data processing method, apparatus, device and computer-readable storage medium, which can improve the efficiency of content review.
[0006] Among them, the technical solution adopted in this application is:
[0007] A data processing method, comprising:
[0008] Perform feature extraction processing on the data to be audited to obtain multiple feature information;
[0009] Based on the weight information corresponding to the type of the data to be reviewed, the plurality of feature information are fused to obtain fused feature information;
[0010] Calculating the review pass rate of the data to be reviewed based on the fused feature information;
[0011] Audit resources are allocated to the data to be audited based on the audit pass rate.
[0012] A data processing device, comprising:
[0013] A processing unit, used for performing feature extraction processing on the data to be audited to obtain multiple feature information;
[0014] A fusion unit, configured to fuse the plurality of feature information based on weight information corresponding to the type of the data to be reviewed to obtain fused feature information;
[0015] A calculation unit, used for calculating the review pass rate of the data to be reviewed based on the fused feature information;
[0016] An allocation unit is used to allocate audit resources to the data to be audited based on the audit pass rate.
[0017] In one embodiment of the present application, based on the aforementioned scheme, the data to be reviewed includes game text content; the processing unit is also used to analyze at least one of the text length, text structure and word frequency information of the game text content to obtain basic text feature information; feature extract the naming text in the game text content to obtain deep text feature information; wherein the naming text includes at least one of the following information of the game text content: game character name information, equipment name information, map location information, game music information, game task information and skill name information; keyword extraction is performed on the game text content to obtain the keyword text information.
[0018] In one embodiment of the present application, based on the aforementioned scheme, the data to be reviewed includes game image content; the processing unit is also used to analyze at least one of the number of images, image size and image type of the game image content to obtain basic image feature information; identify the game static features in the game image content to obtain depth image feature information; wherein the game static features include at least one of the following information of the game image content: equipment information, map location information, character modeling information, game plot information and character status information; perform visual detection calculations on the game image content to obtain image visual feature information.
[0019] In one embodiment of the present application, based on the aforementioned scheme, the data to be reviewed includes game video content; the processing unit is also used to extract basic attributes of the game video content to obtain basic video feature information; wherein the basic attributes include at least one of the following information of the game video content: duration, resolution, frame rate and format; the game dynamic features in the game video content are identified to obtain deep video feature information; wherein the game dynamic features include at least one of the following information of the game video content: character action information, voice information, map scene information, character modeling information, character status information, game music information and game asset information; visual detection calculations are performed on the game video content to obtain video visual feature information.
[0020] In one embodiment of the present application, based on the aforementioned scheme, the calculation unit is further used to perform type calculation based on the multiple feature information to determine the type to which the data to be reviewed belongs; the processing unit is further used to obtain a weight list based on the determined type, and the weight list includes the weight of each feature information; the fusion unit is further used to fuse the multiple feature information based on the weight list to obtain the fused feature information.
[0021] In one embodiment of the present application, based on the aforementioned scheme, the processing unit is also used to filter out feature information used for type judgment from the multiple feature information to obtain at least one type feature information; perform type identification on the at least one type feature information to obtain a type probability distribution of the at least one type feature information; and calculate the type to which the data to be reviewed belongs based on the type probability distribution.
[0022] In one embodiment of the present application, based on the aforementioned scheme, the processing unit is also used for performing type identification on the at least one type feature information to obtain the type probability distribution of the at least one type feature information, including: obtaining a keyword list, the keyword list including a plurality of game keywords; based on the plurality of game keywords, performing game keyword matching on the at least one type feature information to obtain a plurality of game keyword matching rates; and using the plurality of game keyword matching rates as the type probability distribution of the at least one type feature information.
[0023] In one embodiment of the present application, based on the aforementioned scheme, the processing unit is also used to obtain historical audit result information that matches the type of the data to be audited; based on the type of the data to be audited and the historical audit result information, a weight is assigned to each feature information to obtain the weight list.
[0024] In one embodiment of the present application, based on the aforementioned scheme, the processing unit is also used to obtain the weight of each feature information in the weight list; wherein the weight of the feature information of potential abnormal game data in the multiple feature information contained in the weight list is greater than the weight of other feature information in the multiple feature information; the potential abnormal game data includes at least one of the following information of the data to be reviewed: user text information, user voice information, character behavior information and equipment transaction information; based on the weight of each feature information, the multiple feature information are weighted and summed to obtain the fused feature information.
[0025] In one embodiment of the present application, based on the aforementioned scheme, the processing unit is further used to select a target calculation model based on the type of the data to be reviewed; use the target calculation model to predict the fused feature information, and output the review pass rate.
[0026] In one embodiment of the present application, based on the aforementioned scheme, after allocating audit resources to the data to be audited based on the audit pass rate, the processing unit is also used to obtain the audit results corresponding to the data to be audited; and adjust the parameters of the target computing model based on the audit results.
[0027] In one embodiment of the present application, based on the aforementioned scheme, the processing unit is also used to prioritize the data to be reviewed based on the review pass rate; and allocate review resources to the data to be reviewed in the order of priority ranking, wherein the order of allocating review resources is positively correlated with the priority ranking of the data to be reviewed.
[0028] In one embodiment of the present application, based on the aforementioned scheme, the processing unit is also used to obtain an audit pass rate table, in which the audit pass rates of multiple predicted data are recorded in a preset order; the audit pass rate of the data to be audited is compared with the audit pass rates of the multiple predicted data to determine the target sorting position; the audit pass rate of the data to be audited is added to the target sorting position to obtain an updated audit pass rate table.
[0029] A data processing device includes a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the above data processing method is implemented.
[0030] A computer-readable storage medium stores computer-readable instructions, which, when executed by a processor of a computer, enable the computer to execute the above data processing method.
[0031] A computer program product includes computer-readable instructions, and when the computer-readable instructions are executed by a processor, the above data processing method is implemented.
[0032] The computer program product includes a computer program, which is stored in a computer-readable storage medium. The processor of the electronic device reads and executes the computer program from the computer-readable storage medium, so that the electronic device performs the data processing method as described above.
[0033] In the above technical solution, feature extraction processing can be performed on the data to be reviewed to obtain multiple feature information, which can reflect different features of the data to be reviewed from multiple dimensions. Then, based on the type of the data to be reviewed, the multiple feature information is fused, so that the obtained fused feature information can be flexibly adjusted according to the type of the data to be reviewed, has stronger adaptability, can better meet the review scenarios of different types of data, improve the accuracy of the review pass rate, promote the allocation of review resources, and improve the efficiency of content review.
[0034] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] The drawings herein are incorporated into the specification and constitute a part of the specification, showing embodiments consistent with the present application, and together with the specification, are used to explain the principles of the present application. Obviously, the drawings described below are only some embodiments of the present application, and for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:
[0036] Figure 1 It is a schematic diagram of an implementation environment involved in this application;
[0037] Figure 2 It is a schematic diagram of a data processing system designed by the present application;
[0038] Figure 3 It is a service configuration diagram of a data processing system involved in this application when applied in an audit scenario;
[0039] Figure 4 is a flow chart of a data processing method according to an exemplary embodiment;
[0040] Figure 5 is a flow chart of a data processing method according to another exemplary embodiment;
[0041] Figure 6a is a text content feature extraction element diagram according to an exemplary embodiment;
[0042] Figure 6b is an element diagram of image content feature extraction according to an exemplary embodiment;
[0043] Figure 6c is a diagram of video content feature extraction elements according to an exemplary embodiment;
[0044] Figure 7 is a flow chart of a data processing method according to another exemplary embodiment;
[0045] Figure 8 is a flow chart of a data processing method according to another exemplary embodiment;
[0046] Fig. 9 is a flow chart of a data processing method according to another exemplary embodiment;
[0047] Fig.10is a flow chart of a data processing method according to another exemplary embodiment;
[0048] Fig.11 is a flow chart of a data processing method according to another exemplary embodiment;
[0049] Fig.12 It is a schematic diagram of converting priority issues into review pass rates involved in this application;
[0050] Fig.13 is a flow chart of a data processing method according to another exemplary embodiment;
[0051] Fig.14 is a flow chart of a data processing method according to another exemplary embodiment;
[0052] Fig.15 is a block diagram of a data processing device according to an exemplary embodiment;
[0053] Fig.16 The figure is a schematic diagram of the hardware structure of a data processing device according to an exemplary embodiment. DETAILED DESCRIPTION
[0054] Here, exemplary embodiments will be described in detail, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments that are the same as the present application. Instead, they are only examples of devices and methods that are the same as some aspects of the present application as detailed in the attached claims.
[0055] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.
[0056] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.
[0057] It should be noted that the "multiple" mentioned in this application refers to two or more than two. "And / or" describes the association relationship of the associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship.
[0058] It should be noted that in the specific implementation of this application, user-related data is involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.
[0059] Before introducing the technical solutions of the embodiments of the present application, the technical terms involved in the embodiments of the present application are first introduced here.
[0060] XGBoost (eXtreme Gradient Boosting) is a machine learning algorithm that belongs to the gradient boosting tree algorithm in ensemble learning. It performs well in solving classification and regression problems and is widely used in industry and academia. XGBoost combines gradient boosting and regularization techniques to improve prediction accuracy and generalization ability by iteratively training a series of decision tree models.
[0061] Multi-factor fusion refers to the comprehensive processing of multiple feature information to obtain a more comprehensive and accurate feature representation. In data processing and machine learning tasks, multiple features (i.e., features of multiple factors) are usually extracted from the original data. These features can include numerical features, classification features, text features, etc. By fusing these features, the relationship and importance between different features can be comprehensively considered to obtain a more representative and discriminative feature representation.
[0062] A feature vector is a vector used to represent the characteristics of data. A feature vector is a multidimensional vector where each dimension represents a feature or attribute of the data. These features can be numerical, categorical, or other forms of data. In machine learning and pattern recognition tasks, feature vectors are used to describe the key features of the data and are provided as input to the algorithm for model training and prediction. By selecting and extracting appropriate features and representing them in vector form, machine learning algorithms can better understand and process data.
[0063] Feature engineering refers to the processing and transformation of raw data in machine learning and data analysis tasks to extract useful features or create new features to improve model performance. Feature engineering includes steps such as data cleaning, data transformation, feature selection, and feature construction. Through feature engineering, potential patterns and regularities can be discovered from raw data, making the model easier to learn and generalize. In the embodiment of the present application, feature vectors of the data to be reviewed can be extracted through feature engineering processing.
[0064] When reviewing game UGC content, one method is to use machine learning algorithms to review game data. However, the machine learning algorithms proposed in related technologies have low accuracy in reviewing game data, which leads to low reference value of the recognition results of machine learning algorithms. Based on this situation, it is usually necessary to conduct artificial intelligence review after machine algorithm review, which not only increases the burden of manual review, but also has low review efficiency.
[0065] Based on this, the embodiments of the present application respectively propose a data processing method, a data processing device, a data processing equipment, a computer-readable storage medium, and a computer program product. In these embodiments, multiple feature information can be extracted from the data to be reviewed, and the multiple feature information can characterize the features of the data to be reviewed from different dimensions, and then after fusion processing, fused feature information is obtained. The fused feature information can well express the features of the data to be reviewed, so that the review pass rate calculated based on the fused feature information is more accurate, which improves the reference value of the review pass rate for manual review and improves the efficiency of content review.
[0066] It should be noted that machine learning (ML) is a multi-disciplinary interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.
[0067] See also Figure 1 , Figure 1 It is a schematic diagram of an implementation environment involved in this application.
[0068] Figure 1 The illustrated implementation environment includes a feature extraction module 110 , a fusion module 120 , a calculation module 130 , and an audit resource allocation module 140 .
[0069] Among them, the feature extraction module 110 can perform feature extraction on the input data to be reviewed, and the feature extraction can specifically include the process of feature engineering processing and feature extraction processing. After feature extraction, the feature extraction module 110 can output multiple feature information, such as feature information 1, feature information 2... Each feature information can reflect the characteristics of the data to be reviewed from a certain dimension, so a feature information can also be called the feature information of a factor, and the factor is a dimension contained in the data to be reviewed. For example, for a game video, its factors may include clarity, resolution, game scene, game character, game technique, etc. The feature extraction module 110 can extract a corresponding feature information for each factor to reflect the characteristics of the game video under each factor. It should be noted that the feature information in the embodiment of the present application may refer to a feature vector.
[0070] The function of the fusion module 120 is to fuse or combine different feature information from the feature extraction module 110. The purpose of this module is to integrate different feature information to provide a more comprehensive and integrated feature representation. Fusion can include simple feature connection, weighted fusion, feature conversion and other operations to generate a more advanced feature representation. The fusion module 120 can fuse the multiple feature information extracted by the feature extraction module based on the weight information corresponding to the type of data to be reviewed. By adopting appropriate algorithms or methods, these feature information are combined to obtain fused feature information. The fused feature information can comprehensively express the multiple aspects of the data to be reviewed and provide a more comprehensive and accurate information basis.
[0071] The calculation module 130 is the core part for performing specific tasks or algorithms, and can perform corresponding calculations and analyses based on the obtained fusion feature information. It can calculate the pass rate (PTR) of the audit data to be audited according to a pre-set algorithm or model. This calculation process can rely on machine learning algorithms, statistical analysis methods or other related technologies to obtain an accurate audit pass rate. In an embodiment of the present application, the calculation module 130 may include multiple calculation models, such as an XGBoost model, and for the same calculation model, it can also be a calculation model for different types of data to be audited. For example, the calculation module 130 may include a calculation model for game videos, a calculation model for news graphic content, a calculation model for travel graphic content, etc., which is not limited in the embodiment of the present application.
[0072] Optionally, the fusion module 120 may be included in the calculation module 130 .
[0073] The audit resource allocation module 140 is responsible for allocating audit resources to the audited data according to the audit pass rate obtained by the calculation module 130. According to the audit pass rate, it can be determined to give different levels of audit resources, such as allocating more manpower, time or other audit tools and equipment, to optimize the audit process and improve audit efficiency.
[0074] In addition, the implementation environment may also include a data collection module, a data cleaning module, a data partitioning module, a model training module, a model evaluation module, a model deployment module, a model update module, etc.
[0075] The data collection module is responsible for collecting data to be reviewed, which can be user-uploaded content, web page data, social media data, etc. This module can also collect historical review data including review results, which can include information such as reviewers, review time, and review pass results, and a data set can be constructed.
[0076] The data cleaning module is used to clean and preprocess the collected data, including removing duplicate data, processing missing values, processing outliers, etc., to ensure the quality and integrity of the data.
[0077] The data partitioning module is used to divide the collected historical audit data into training set, validation set and test set according to a certain ratio.
[0078] The model training module is used to train machine learning models for use in the audit process. It can include steps such as model selection, feature selection, and parameter tuning to train the model using labeled training data to give it audit functionality and accuracy.
[0079] The model evaluation module is used to evaluate the performance and effect of the trained model through the test set. The model is evaluated by using a reserved test data set or cross-validation techniques, and the model's accuracy, recall, F1 value and other indicators are calculated to ensure its accuracy and reliability in the audit task. Among them, the F1 value is a commonly used comprehensive evaluation indicator used to measure the performance of a binary classification model. It combines the model's accuracy (Precision) and recall (Recall) to provide a comprehensive evaluation of the model.
[0080] The model deployment module is used to deploy the trained model to the review system, use the model to estimate the review pass rate of the content to be reviewed in the future, and prioritize the content based on the estimated results.
[0081] The model update module is used to regularly update training data and models to adapt to changes in audit content and auditors.
[0082] These modules work together to achieve comprehensive audit processing of audit data through processes such as feature extraction, fusion, calculation and resource allocation, thereby improving audit accuracy and efficiency.
[0083] See also Figure 2 , Figure 2 Schematic diagram of a data processing system involved in this application. The data processing system can be deployed in a computer. Figure 2 As shown, the data processing system may include a front-end system 210 and a prediction system 220. Among them, the front-end system 210 includes a feature extraction module 211, and the prediction system 220 includes a target calculation model 211. And the target calculation model 221 in the prediction system 220 is connected to the content scheduling module 222, and the prediction probability can be output to the content scheduling. The content scheduling module 222 is connected to the resource allocation module 223. The resource allocation module 223 is connected to the lighthouse data channel 224, and the lighthouse data channel 224 can collect the actual result output of the resource allocation module 223. The lighthouse data channel 224 is connected to the offline training module 226, and the offline training module 226 is connected to the model update module 225. In this way, the offline training module 226 can analyze the actual result output collected by the lighthouse data channel 224, and optimize the target calculation model 221 through the model update module 225.
[0084] The front-end system 210 is used to perform feature extraction processing on the data to be predicted. Figure 2 In the process, feature extraction is performed on the data to be audited, and feature information of multiple factors is obtained, namely feature information 1, feature information 2, etc., for example, feature information of factors such as clarity, advertising content, account level, etc. After obtaining these feature information, the front-end system 210 can output them to the prediction system 220.
[0085] The prediction system 220 can input the multiple feature information output by the front-end system 220 into the target calculation model 221, such as the XGBoost model. The target calculation model 221 can fuse the multiple feature information and output the predicted probability of the data to be predicted. After obtaining the predicted probability, it can be output to the content scheduling module 222, so that the content scheduling module 222 prioritizes the predicted data for which the predicted probability has been calculated. After the content scheduling module 222 sorts the predicted data, it can allocate audit resources to the predicted data according to the order of priority through the resource allocation module 223.
[0086] After the audit resources are allocated, the actual result output is obtained. The lighthouse data channel 224 can collect the actual result output and input it into the offline training module 226 for training to train the parameters of the target calculation model 221 and obtain the output related to optimization. Therefore, the target calculation model 221 can be optimized and adjusted by the model update module 225 to make the calculation of the predicted probability more accurate. Among them, the model output included in the offline training module 226 refers to the parameters output by the target calculation model 221, which can include the predicted probability, the parameters related to the multi-factor fusion, etc.
[0087] It should be noted that the data to be predicted may be data that needs to be extracted from multiple factors and fused with multiple factors, and the predicted probability of a certain dimension is calculated through the target calculation model. For example, it may be data to be audited that needs to calculate the audit pass rate, or it may be game background data that needs to calculate the real-time violation probability in the game, or it may be news articles that need to calculate the predicted click rate, etc. Those skilled in the art may make adaptive changes to the data to be predicted according to different application scenarios. Accordingly, the prediction system 220 may be changed according to the different data to be predicted. For example, for the data to be audited, the prediction system 220 may be an audit pass rate calculation system; for the game background data, the prediction system 220 may be a violation detection system; for the news article, the prediction system 220 may be a click rate prediction system, etc. The above-mentioned actual result output refers to the actual performance of the data to be predicted, such as whether it has passed the audit, whether there is a violation, and what the actual click rate is.
[0088] See also Figure 3 , Figure 3 It is a service configuration diagram of a data processing system used in an audit scenario involved in this application.
[0089] like Figure 3 As shown, in this scenario, the data processing system may include a front-end system 310 , an audit system 320 , and a back-end system 330 .
[0090] Among them, the front system 310 and Figure 2 The front-end system 210 shown is the same as that shown in the figure, and can perform multi-factor feature extraction on the data to be audited to obtain feature information of multiple factors, that is, multiple feature information.
[0091] The backend system 330 may include Figure 2 The content scheduling module, resource allocation module, lighthouse data channel and model update module are shown. Among them, in this audit scenario, the resource allocation module can use the method of receiving order review to review the allocation of resources. Receiving order review refers to the evaluation and allocation of tasks or work orders to ensure that they are correctly allocated to the appropriate personnel or departments to achieve efficient workflow and task execution.
[0092] The audit system 320 can integrate multiple feature information and calculate the audit pass rate. The audit system 320 includes an audit front end 321, a priority calculation service 322, an operation front end 323, a rule management service 324 and a configuration center 325.
[0093] The priority calculation service 322 is a key component used to fuse multiple feature information, calculate the audit pass rate and determine the priority of the data to be audited. The service fuses multiple feature information through the rules configured by the operation front end 323, and calculates the audit pass rate based on the fused feature information. Specifically, the rule management service 324 is responsible for configuring the operation rules, including the method of fusing feature information and the method of calculating the audit pass rate. These rule configurations are stored in the configuration center 325, and the priority calculation service 322 perceives the rule configuration in the configuration center 325 in real time, so as to fuse multiple feature information, calculate the audit pass rate and prioritize.
[0094] In the audit front end 321, the data to be audited is sorted and manually audited according to the priority information provided by the priority calculation service 322. The audit front end 321 obtains the priority of the data to be audited through interaction with the priority calculation service 322, and sorts the data according to the priority, so as to give priority to the data with a high audit pass rate.
[0095] The operation front end 323 is responsible for configuring the operation rules for the rule management service 324. It provides an interface or interface that allows operators to flexibly configure the fusion method of feature information and the calculation method of the audit pass rate. Through the operation front end 323, operators can adjust the weight of feature information, fusion algorithm, and calculation rules of audit pass rate according to specific needs and business scenarios, thereby optimizing the performance and accuracy of the audit system.
[0096] In summary, the priority calculation service 322 is an important component in the audit system responsible for calculating the audit pass rate and determining the priority. It realizes the integration of multiple feature information, the calculation of the audit pass rate and the sorting of priorities through cooperation with the operation front end 323 and the rule management service 324. The configuration center 325 stores the configuration information of the rule management service 324, so that the priority calculation service 322 can perceive the update and change of the rules in real time, thereby maintaining the flexibility and real-time performance of the audit system.
[0097] See also Figure 4 , Figure 4 is a flow chart of a data processing method according to an exemplary embodiment. The method can be applied to Figure 1 The implementation environment shown in the figure is specifically executed by a computer. Of course, the method can also be applied to other implementation environments, and the execution subject of the method is not limited here.
[0098] The data processing method will be described in detail below using a computer as an exemplary execution subject.
[0099] like Figure 4 As shown, in an exemplary embodiment, the method comprises at least the following steps:
[0100] S410, performing feature extraction processing on the data to be audited to obtain multiple feature information.
[0101] The data to be reviewed may include text content, image content and video content. For the data to be reviewed containing different contents, corresponding feature extraction may be performed.
[0102] It should be noted that the data to be reviewed may be one or more of text content, image content and video content. For example, it may be a combination of text content and image content, a combination of text content and video content, a combination of image content and video content, etc.
[0103] For feature extraction of text content, image content, and video content, the following methods can be used:
[0104] Text feature extraction: For text data, natural language processing techniques such as word segmentation, word frequency statistics, word vector representation, etc. can be used to extract dimensional feature information such as keywords, topics, and sentiment trends from the text content of the data to be reviewed. Among them, dimensions can also be called factors. By extracting features from each factor, feature information of multiple factors in the text content can be obtained.
[0105] In the game scenario, natural language processing technology can be used to extract feature information of factors such as game character name information, equipment name information, map location information, game music information, game task information, and skill name information from game-related text data.
[0106] Image feature extraction: For data to be reviewed that contains images, computer vision technologies such as image recognition, object detection, scene analysis, etc. can be used to extract feature information such as objects, locations, scenes, etc. in the image to obtain feature information of multiple factors in the image content.
[0107] In the game scene, computer vision technology can be used to extract feature information of factors such as equipment information, map location information, character modeling information, game plot information, and character status information from game-related image data.
[0108] Video feature extraction: Since a video is composed of a series of continuous image frames, the video can be divided into frames and image features can be extracted for each frame. Similar to image feature extraction, computer vision techniques such as object detection, scene analysis, image recognition and other methods can be used to extract feature information such as objects, positions, scenes, etc. in each video frame to obtain feature information of multiple factors in the video content.
[0109] In the game scene, computer vision technology can be used to divide the game video into continuous image frames, extract character action information, voice information, map scene information, character modeling information, character status information, game music information, game asset information and other visual information feature information.
[0110] In the above feature extraction process, feature engineering can be performed, including feature selection, dimensionality reduction, normalization, etc. By selecting the most relevant features, reducing the dimension of features, or normalizing features, the efficiency and accuracy of the subsequent review process can be improved. When implementing feature engineering, open source tools and libraries such as TensorFlow, PyTorch, OpenCV, etc. can be used to train and predict models. In order to improve the prediction effect of the model, transfer learning technology can be used to fine-tune the pre-trained model on a dataset in a specific field.
[0111] Feature extraction can convert the data to be audited into a feature representation that can be processed by the computer for subsequent data processing and model building. By comprehensively considering the different aspects and features of the data to be audited, more comprehensive and rich information can be obtained, so as to more accurately judge the compliance and audit results of the data to be audited. The computer can normalize the extracted feature data to make it on the same scale. For categorical features, One-Hot Encoding can be used for processing, and the processed feature data can be used as the input data of the XGBoost model.
[0112] In the embodiment of the present application, before performing feature extraction processing on the data to be audited, the following operations may also be performed:
[0113] Data collection: Obtain historical audit records from the content audit system, which include audit results and audit content. In the game scenario, the audit results may include whether the game UGC content has any violations, such as the use of plug-ins and verbal violations.
[0114] Data cleaning: Filter or process invalid or abnormal data, such as duplicate data, missing values, etc. It should be noted that since there may be abnormal data in the game scene, such as when using plug-ins, some parameters will be abnormal. Therefore, you can choose to retain abnormal data according to different scenarios to ensure accurate calculation of the audit pass rate.
[0115] Model training: Select a suitable model to train the historical audit records to obtain a calculation model. For example, the XGBoost model can be selected for training to obtain a calculation model for calculating the audit pass rate. Before training, the historical audit records can be split into a training set and a test set, and then feature extraction is performed and input into the model for training.
[0116] S420: Based on the weight information corresponding to the type of the data to be reviewed, multiple feature information are fused to obtain fused feature information.
[0117] Specifically, feature fusion can be performed based on weight information using weighted average or weighted sum. The feature information of each factor has a corresponding weight according to its importance and influence in the audit process. The weight information corresponding to the type of data to be audited can be used as a guide to adjust the weight values of different factors. Features with higher weights will have a greater impact on the final fused feature information. Multi-factor fusion can be achieved by weighted summing the feature information of each factor and the weight value corresponding to each factor, thereby obtaining fused feature information.
[0118] Among them, the type of the data to be reviewed may refer to the field it involves, such as food, travel, games, animation and other fields. The embodiment of the present application does not limit the type and field.
[0119] For example, if the type of the data to be reviewed is a game, the weight information corresponding to the game type can be obtained, and the multiple feature information can be fused to obtain the final fused feature information.
[0120] Optionally, the type may also include the type of game, such as a shooting game, a simulation game, a competitive game, etc. Each game may have different weight information.
[0121] The weight information may be preset by a person skilled in the art, or may be obtained by a computer assigning weights to multiple feature information in real time according to the type of data to be reviewed. Alternatively, the weight information may be obtained by a computer dynamically modifying the weight information preset by a person skilled in the art according to the type of data to be reviewed, which is not limited in the embodiments of the present application.
[0122] By fusing multiple feature information according to the weight of the data type to be reviewed, the information of each feature can be better comprehensively utilized to fully reflect the characteristics and attributes of the data to be reviewed. The fused feature information obtained in this way can describe the data to be reviewed more comprehensively and accurately.
[0123] S430, calculating the audit pass rate of the data to be audited based on the fused feature information.
[0124] The computer uses fused feature information to evaluate whether the data to be audited meets the audit standards and gives the probability or proportion of passing the audit.
[0125] Specifically, after calculating and obtaining the fusion feature information, the computer can evaluate by obtaining a matching audit model. This model can use a machine learning algorithm, such as an XGBoost model, and train the model using historical audit records to obtain a calculation model.
[0126] Among them, the computer can train a model according to different types of historical audit records, and obtain different types of calculation models. For example, a calculation model for food, a calculation model for tourism, and a calculation model for games, etc. can be obtained. In this way, the final calculated audit pass rate can be more relevant to the type of data to be audited, thereby improving the accuracy of the audit pass rate calculation.
[0127] For example, in a game video review scenario, the fused feature information may include text descriptions of the game video content, image features, and audio features, etc. The corresponding computing model can use this information to determine whether the game video contains illegal content or involves sensitive topics.
[0128] By analyzing and calculating the fused feature information, the audit model can give the probability or proportion of the audited data being audited. The audit pass rate of the audited data can be a binary result (pass or fail) or a continuous value (probability between 0 and 1). The specific value depends on different audit standards and scenario requirements. By accurately calculating and judging the fused feature information, it is possible to more effectively decide whether to pass the audit, thereby improving the accuracy and efficiency of the audit.
[0129] S440, allocating audit resources to the data to be audited based on the audit pass rate.
[0130] Computers can allocate different audit resources to different data based on the audit pass rate of the data to be audited, so as to improve audit efficiency and resource utilization efficiency.
[0131] Specifically, the audit pass rate can be used as an indicator to determine the audit priority of the data to be audited. A higher audit pass rate means that the data to be audited is more likely to pass the audit, so higher priority audit resources can be allocated to these data to complete the audit process faster. On the contrary, a lower audit pass rate may be arranged in a low priority position, the audit resources allocated are more stringent, and the process is longer.
[0132] Resource allocation can include allocating more reviewers, providing more computing resources, or allocating longer review time. Reasonable resource allocation based on the review pass rate of the data to be reviewed can be flexibly adjusted according to the review requirements and risk levels of different data to improve review efficiency. Resource allocation based on the review pass rate can help optimize the order of the review queue and ensure that high-priority data is reviewed and processed faster, thereby reducing review waiting time. This helps improve review efficiency, reduce review costs, and ensure the content compliance and user experience of the platform.
[0133] In the game scenario, according to different types of game content and audit requirements, the computer can dynamically adjust the audit pass rate of data. For example, when the audit pass rate of a certain type of game content is high, fewer audit resources can be deployed so that more resources can be allocated to content types with lower audit pass rates, thereby improving overall audit efficiency and resource utilization efficiency.
[0134] Through this method, the computer performs feature extraction processing on the data to be reviewed and obtains multiple feature information. These feature information can comprehensively reflect the different features of the data to be reviewed from multiple angles. Subsequently, these feature information are fused by combining the type of data to be reviewed, so that the fused feature information can be flexibly adjusted according to the review scenarios of different types of data, showing stronger adaptability. This fusion processing of comprehensive feature information can more accurately evaluate the compliance of the data to be reviewed, quickly and effectively handle the review process, thereby ensuring the quality and compliance of the platform content. In addition, this method significantly improves the accuracy of the review pass rate, while also helping to optimize the allocation of review resources and further improve the efficiency of content review.
[0135] In one embodiment of the present application, another data processing method is provided, which can be executed by a computer. Figure 5 As shown, the data processing method includes at least one execution method of S511 to S513, S521 to S523 and S531 to S533, and S420 to S440. That is, at least two steps of S511 to S513, or at least two steps of S521 to S523, or at least two steps of S531 to S533 are Figure 4The specific implementation method of S410 is shown.
[0136] When the data to be reviewed includes text content, at least two steps from S511 to S513 can be executed; when the data to be reviewed includes image content, at least two steps from S521 to S523 can be executed; when the data to be reviewed includes video content, at least two steps from S531 to S533 can be executed. When the data to be reviewed includes a combination of the above three contents, the corresponding steps can be executed respectively. It should be noted that the execution order of each step in S511 to S513 can be adaptively adjusted according to actual needs. The embodiment of the present application takes the execution order of S511, S512 and S513 as an example for explanation. Similarly, the execution order of S521 to S523 and S531 to S533 can also be adaptively adjusted.
[0137] It should be noted that S511 to S513, S521 to S523, and S531 to S533 in the embodiment of the present application are described by taking the text content included in the audit data as game text content, the image content as game image content, and the video content as game video content as examples. That is, the embodiment of the present application is described by taking the game scene as an example. Of course, the method of the present application can also be applied to other scenes, such as food scenes, travel scenes, etc., without limitation.
[0138] The following game scenario is used as an example to describe S511 to S513:
[0139] S511, analyzing at least one of the text length, text structure and word frequency information of the game text content to obtain basic text feature information.
[0140] The computer will analyze the text content to be reviewed to obtain basic text feature information. This feature information can include the length of the text, the structure of the text, and word frequency statistics. First, the computer will count the number of words or characters in the text to obtain the length information of the text. Next, the computer will analyze the structure of the text, including the number of paragraphs, average paragraph length, number of titles, etc. This information can reveal the organization and structural characteristics of the text. In addition, the computer will also perform word frequency statistics, counting the high-frequency and low-frequency words in the text and calculating their ratio. This word frequency information helps to understand the key words and common words in the text.
[0141] It should be noted that, in the embodiment of the present application, the basic text feature information may include multiple feature information, such as feature information of text length, feature information of text structure, and feature information of word frequency information. In addition, those skilled in the art may add other matching feature information as the basic text feature information to increase the richness of the basic text feature information.
[0142] S512, extracting features of the named text in the game text content to obtain deep text feature information.
[0143] The naming text includes at least one of the following information of the game text content: game character name information, equipment name information, map location information, game music information, game task information and skill name information.
[0144] The computer can extract features of the named texts included in the text content by means of named entity recognition (NER), and these named texts may include game-related vocabulary, and include but are not limited to game character name information, equipment name information, map location information, game music information, game task information, and skill name information.
[0145] For different naming texts, feature extraction can be performed separately, as follows:
[0146] Game character name information: By analyzing the character names that appear in the game text, the word frequency, word vector representation, or other semantic features of the character names can be extracted.
[0147] Equipment name information: For the equipment names in the game text, we can extract features such as word frequency, word vector, equipment attributes, etc. These features can be used for equipment classification, evaluation or recommendation.
[0148] Map location information: For map location descriptions in game text, features such as geographic location vocabulary, terrain features, or location relationships can be extracted, which can be used for map visualization, navigation, or location query.
[0149] Game music information: By analyzing the music names, descriptions or comments in the game text, we can extract the music style, emotional tendencies or features associated with the game scenes, which can be used for music matching or evaluation.
[0150] Game task information: For the task description, goal or reward in the game text, we can extract features such as task type, difficulty, reward value, etc., which can be used for task classification and player behavior analysis.
[0151] Skill name information: By analyzing the skill names, effects or descriptions in the game text, we can extract features such as skill type, energy consumption, and range of action. These features can be used for skill classification, balance adjustment, or player character strategy analysis.
[0152] For example, a game text includes the following sentence: When Xiao Qiao uses the knock-up skill, pink light will be emitted. In this sentence, the character name information may be Xiao Qiao, and the knock-up skill and the pink light are the skill name information, and the computer can extract the information of the two named texts.
[0153] It should be noted that those skilled in the art can add or delete the information types included in the naming text according to actual conditions.
[0154] Optionally, the computer can also extract semantic features of the game text content to be reviewed and perform sentiment analysis to obtain deep text feature information. Specifically, by using pre-trained natural language processing (NLP) models, such as the BERT (Bidirectional Encoder Representations from Transformers) model and the generative pre-trained model (GPT), the computer can extract the semantic feature vectors of the text, which can express the semantic information of the text. At the same time, the computer will use the sentiment analysis model to predict the emotional tendency of the text and determine whether the sentiment of the text content is positive, negative or neutral.
[0155] In this way, semantic feature information and sentiment feature information can be used as deep text feature information, which can more comprehensively understand and analyze the meaning and sentiment of the text.
[0156] In addition, those skilled in the art may add other matching feature information as deep text feature information to increase the richness of the deep text feature information.
[0157] S513, extract keywords from the game text content to obtain keyword text information.
[0158] The computer will perform keyword extraction on the text content to be reviewed to obtain keyword text information. For example, the computer can use algorithms such as Term Frequency-Inverse Document Frequency (TF-IDF) or TextRank to extract the most representative keywords in the text based on the importance and relevance of the words in the text. These keywords can summarize and generalize the theme and key points of the text, providing important reference for subsequent review and analysis.
[0159] Optionally, the computer can also perform topic identification on the text content, and use a topic model, such as a probabilistic topic model (Latent Dirichlet Allocation, LDA), to assign topic tags to the text content to obtain tag information of the text content, which is also a kind of feature information.
[0160] like Figure 6a The figure shows an exemplary text content feature extraction element diagram. The computer can extract features of the text content such as text length, text structure, word frequency, semantics, sentiment, keywords and themes. Those skilled in the art can adaptively add or delete objects for feature extraction to obtain richer feature extraction settings that better meet actual needs.
[0161] In text recognition in the field of games, natural language processing technology can be used to extract key information. By analyzing game-related text, key elements such as characters, game scenes, plots, and game screens can be identified. Through technologies such as word segmentation and entity recognition, words and phrases in the text can be matched and classified with predefined game-related vocabulary. For example, by identifying keywords such as specific character names, locations, and task names, the characters and game scenes involved in the text can be determined. In addition, by analyzing the context and description of the text, relevant information about the game plot, task objectives, and game screens can be obtained. Combining these extracted text features, the game content can be better understood and analyzed, and used in game-related review scenarios.
[0162] Through S511 to S513, the computer analyzes and processes the text content in different aspects. By extracting basic text features, deep text features and keyword text information, the computer can more comprehensively and accurately understand and evaluate the text data to be reviewed. This provides a basis and reference for subsequent review and resource allocation, and improves the efficiency and accuracy of content review.
[0163] The following describes S521 to S523 using a game scenario as an example:
[0164] S521, analyzing at least one of the number of images, image size, and image type of the game image content to obtain basic image feature information.
[0165] The computer analyzes the game's graphic content and can focus on at least the following characteristic information: the number of images, image size, and image type. First, the computer counts the number of images in the graphic content, which can help the computer understand the number of images contained in the article and their relative importance. Second, the computer calculates the average size, maximum size, and minimum size of the image to gain an understanding of the image size distribution. Finally, the computer counts the image type, such as JPEG, PNG, etc., to understand the image format used.
[0166] S522, identifying the static features of the game in the game image content to obtain depth image feature information.
[0167] Among them, the game static features include at least one of the following information of the game image content: equipment information, map location information, character modeling information, game plot information and character status information.
[0168] The computer uses a deep learning model to identify the static features of the game in the image content, thereby obtaining deep image feature information. For example, the computer can use an object detection model, such as the You Only Look Once (YOLO) model, the Faster R-CNN model, etc., to identify the character modeling information in the image and determine its map location information. Or use a scene recognition model to identify the type of scene to which the image belongs. These models can obtain an understanding and representation of the image content through pre-training to generate deep image feature information.
[0169] For different game static features, feature extraction can be performed separately, as follows:
[0170] Equipment information recognition: Through object detection and image segmentation technologies, equipment items in game images, such as tools and armor, are identified, and their feature information, such as name and attributes, is extracted.
[0171] Map location information recognition: Through scene analysis and image recognition technology, map locations in game images, such as cities, forests, mountains, etc., are identified, and their location feature information, such as coordinates, geographic features, etc., are extracted.
[0172] Character modeling information recognition: Through technologies such as character recognition and posture estimation, the character model in the game image is identified, including the character's appearance, clothing, posture and other characteristic information.
[0173] Game plot information recognition: Through text recognition and natural language processing technology, combined with image content, extract plot information from game images, such as dialogues, task prompts, etc., and further analyze its emotional tendencies, keywords and other features.
[0174] Character status information recognition: Through technologies such as facial expression recognition and motion analysis, the status information of the characters in the game images, such as anger, happiness, walking, running, etc., is identified, and relevant feature information is extracted.
[0175] Through the above method, different static features in the game image content can be identified and extracted to obtain deep image feature information for tasks such as image understanding, sentiment analysis, and behavior judgment in game scenarios.
[0176] It should be noted that those skilled in the art can add or delete the information types included in the static features of the game according to actual conditions.
[0177] S523, performing visual detection calculation on the game image content to obtain image visual feature information.
[0178] Computers can use pre-trained computer vision models, such as ResNet and VGG, to extract visual features of images. These models are trained on large-scale image data and can effectively capture visual information in images. By inputting images into these models, computers can obtain visual feature information associated with the image content for subsequent analysis and processing.
[0179] like Figure 6b The figure shows an exemplary image content feature extraction element diagram. The computer can extract features of the image content such as the number of images, image size, image type, vision, label, object, scene, etc. Those skilled in the art can adaptively add or delete objects for feature extraction to obtain richer feature extraction settings that better meet actual needs.
[0180] In the image feature extraction in the field of games, computer vision technology can be used to identify and analyze key information in the image. Through image recognition and object detection technology, game characters, props, environmental elements and other objects in the image can be automatically identified, and their position, size, shape and other feature information can be further extracted. At the same time, using scene analysis technology, game scenes in the image can be identified and classified, such as battle scenes, urban landscapes, forests, etc., so as to obtain scene features in the image content. In addition, image processing technology, such as color analysis, texture extraction and other methods, can also be used to obtain the visual features of the image, such as color matching, light and shadow effects, etc., to further enrich the feature information of the image.
[0181] Through S521 to S523, the computer can extract and analyze the features of the image part. These steps can provide image feature information on multiple aspects such as image quantity, size, type, object information, location information, scene information, and visual feature vectors, providing rich visual content for subsequent review and analysis. At the same time, the computer can also apply the feature engineering solution of text content mentioned above to process the text part of the graphic content to obtain more comprehensive feature information.
[0182] The following describes S531 to S533 using a game scenario as an example:
[0183] S531, extracting basic attributes of the game video content to obtain basic video feature information; wherein the basic attributes include at least one of the following information of the game video content: duration, resolution, frame rate and format.
[0184] The computer performs feature extraction on short video content, mainly including the basic attributes of the video. First, the computer extracts the duration of the video to obtain the total duration information of the video. Secondly, the computer obtains the resolution of the video, including the width and height of the video, which is used to describe the picture clarity of the video. In addition, the computer records the frame rate of the video, which indicates the number of image frames displayed per second. It is also necessary to obtain the video encoding format, such as H.264, VP9, etc., and the audio encoding format, such as AAC, MP3, etc., to describe the compression and encoding methods of the video and audio.
[0185] It should be noted that those skilled in the art can adaptively add basic attributes of a video according to actual application scenarios to obtain basic video feature information corresponding to each basic attribute.
[0186] S532, identifying the game dynamic features in the game video content to obtain deep video feature information.
[0187] Among them, the game dynamic features include at least one of the following information of the game video content: character action information, voice information, map scene information, character modeling information, character status information, game music information and game asset information.
[0188] Computers use deep learning models to analyze video content to extract deep video feature information. Computers can use action recognition models, such as 3D convolutional neural networks (3D-CNN), inflated 3D networks (I3D), etc., to identify the types of actions in the video. By analyzing the video at the frame level and inferring the model, the computer can determine the actions that appear in the video and their characteristics. In addition, the computer can also use speech recognition models, such as DeepSpeech, to convert the speech in the video into text, thereby obtaining the speech information in the video.
[0189] For different game dynamic features, feature extraction can be performed separately, as follows:
[0190] Character action information recognition: Through technologies such as action recognition and posture estimation, the action sequences of characters in game videos, such as attacking, jumping, running, etc., are analyzed, and relevant action feature information, such as action type, duration, frequency, etc., is extracted.
[0191] Voice information recognition: Through technologies such as voice recognition and semantic analysis, the voice content in the game video is converted into text, and semantic information is further extracted, such as character dialogues, game instructions, etc. Alternatively, for games that allow voice communication, the voice information of the game players can also be extracted, and semantic information can be extracted.
[0192] Map scene information recognition: Through scene analysis and image recognition technology, map scene changes in game videos, such as different regions, different weather, etc., are identified, and relevant scene feature information, such as place names, climate conditions, and terrain changes, are extracted.
[0193] Character modeling information recognition: Through technologies such as character recognition and posture estimation, the appearance, clothing, posture and other features of the characters in the game video are analyzed, and relevant character modeling information is extracted.
[0194] Character status information recognition: Through facial expression recognition and motion analysis, the emotional state of the character in the game video, such as happiness, anger, fatigue, etc., is analyzed, and relevant status feature information is extracted. For fighting games, the current status of the character can be analyzed, such as health, skill cooldown, whether it is controlled, etc.
[0195] Game music information recognition: Through audio analysis and music recognition technologies, background music and sound effects in game videos can be identified, and their characteristic information, such as melody, rhythm, sound effects, etc. can be extracted.
[0196] Game asset information recognition: Through technologies such as object detection and image segmentation, game assets in game videos, such as gold coins, treasure chests, props, special equipment, etc., are identified, and relevant asset feature information, such as name, effect, etc., is extracted.
[0197] It should be noted that those skilled in the art can add or delete the information types included in the game dynamic features according to actual conditions.
[0198] S533, performing visual detection calculation on the video content to obtain video visual feature information.
[0199] Computers can use computer vision algorithms and techniques to analyze videos and extract visual features from them. These features can include object detection, scene recognition, key frame extraction, etc. By performing visual detection calculations on each frame in the video, the computer can obtain feature information about objects, scenes, and other visual content in the video.
[0200] like Figure 6c The figure shows an exemplary image content feature extraction element diagram. The computer can extract features of the video content such as video duration, video resolution, video frame rate, video format, audio format, vision, motion and voice. Those skilled in the art can adaptively add or delete objects for feature extraction to obtain richer feature extraction settings that better meet actual needs.
[0201] In the video feature extraction in the field of games, the video can be decomposed into a series of continuous image frames, and image features can be extracted for each frame. With the help of computer vision technology, object detection, scene analysis, image recognition and other methods can be applied to extract key information from video frames. Through object detection technology, game characters, props, specific objects, etc. in the video can be identified and detected, and their position, action, appearance and other feature information can be extracted. At the same time, through scene analysis technology, game scenes in the video, such as battle scenes, urban landscapes, indoor environments, etc., can be identified and classified, so as to obtain scene features in the video content. In addition, image recognition technology can also be used to identify and classify video frames, such as character recognition and expression recognition, to obtain characteristics such as character behavior and emotions in the video.
[0202] Through S531 to S533, the computer can extract and analyze the basic properties of the video, such as duration, resolution, frame rate and format. The computer can also use the deep learning model to identify the action information and voice information in the video and extract deep video feature information. In addition, through visual detection calculations, the computer can obtain more comprehensive video visual feature information, including objects, scenes, and key frames. It should be noted that the feature engineering scheme for text content mentioned earlier can also be applied to the text obtained from speech recognition to obtain more comprehensive feature information.
[0203] Optionally, when performing feature extraction, for at least one of the game text, game image, and game video to be reviewed, the computer may extract features of at least the following factors:
[0204] Game operation: including the player's key input, mouse movement, joystick operation, etc., which can extract features such as operation frequency, operation duration, key combination, etc.
[0205] Game behavior: including the player's movement path, attack behavior, skill release, item use, etc., which can extract characteristics such as behavior frequency, behavior sequence, and behavior timing.
[0206] Game interaction: including interactions, transactions, team formation, etc. between players, and can extract features such as interaction frequency, interaction objects, and interaction modes. Game achievements: including players' task completion, level upgrades, rewards, etc., and can extract features such as achievement completion, achievement type, and achievement difficulty.
[0207] Game economy: includes the player’s gold coins, props, equipment and other property status, and can extract features such as property quantity, property source, property transactions, etc.
[0208] Social interaction: includes players’ social relationships, friend lists, guild relationships, etc. in the game, and can extract features such as social relationship strength, social influence, and social behavior.
[0209] Game environment: including game maps, scene features, non-player character (NPC) distribution, etc., can extract features such as map location, scene type, NPC interaction, etc.
[0210] Game time: includes players’ game time, login frequency, game time period, etc., and can extract features such as game time, login time period, and game activity.
[0211] Those skilled in the art may add or delete specific factors based on the game scenarios to adapt to different game scenarios.
[0212] Through the above method, the computer can extract features from at least one of the text content, image content and video content included in the audit data, and obtain multiple feature information, which can describe and characterize the audit data from different dimensions. The fusion processing of multiple feature information can improve the accuracy of the audit pass rate. Different feature information can complement and verify each other, further improving the accuracy and reliability of the audit.
[0213] In one embodiment of the present application, another data processing method is provided, which can be executed by a computer. Figure 7 As shown, the data processing method may include S410 to S710 to S730, S430 to S440. That is, S710 to S730 are Figure 4 The specific implementation method of S420 is shown.
[0214] S710 to S730 are described below:
[0215] S710, performing type calculation based on multiple feature information to determine the type of the data to be reviewed.
[0216] The type may refer to the field involved in the data to be reviewed. The computer may use the extracted multiple feature information to perform field or subject calculations on the data to be reviewed. Data in different fields or subjects may differ in features. By analyzing these feature information, the type of the data to be reviewed may be determined.
[0217] S720: Obtain a weight list based on the determined type, where the weight list includes a weight for each feature information.
[0218] Once the computer determines the type of data to be reviewed, it can obtain a weight list related to that type. The weight list includes the importance weight of each feature information in the review process. These weights can be adjusted according to the type of data to be reviewed to reflect the importance of different features to that type, so as to more accurately integrate the feature information.
[0219] S730, fusing the multiple feature information based on the weight list to obtain fused feature information.
[0220] The computer can weight and fuse each feature information into a comprehensive feature vector or feature representation according to its weight. The purpose of fusing feature information is to comprehensively utilize the advantages of different features, reduce redundant information, and better express the characteristics of the type of data to be audited. Fusing feature information can more accurately describe the type of data to be audited and provide more valuable information for the subsequent audit process.
[0221] In the game scene, the types of games may include at least action games, adventure games, shooting games, role-playing games, strategy games, simulation games, sports games, music games, multiplayer online tactical competitive games, etc. The computer can use machine learning and data analysis methods to perform feature selection on the extracted game features. By training and analyzing a large amount of labeled game data, the computer can determine which features have high discrimination and importance for different types of games. For example, the appearance features of certain specific characters may have a high discrimination for judging the type of game, and the elements of certain key scenes may have obvious differences between different types of games. Based on these discriminations and differences, the computer can identify the game type indicated by the game content when the data to be reviewed is game content, and then obtain the weight information corresponding to the game type, perform fusion processing, and obtain fused feature information.
[0222] Through this method, the computer fuses multiple feature information based on the weight list to obtain fused feature information, which can improve the description accuracy and comprehensive expression ability of the data to be audited.
[0223] In one embodiment of the present application, another data processing method is provided, which can be executed by a computer. Figure 8 As shown, the data processing method may include S810 to S830, wherein S810 to S830 are Figure 7 The specific implementation method of S710 is shown.
[0224] S810 to S830 are described below:
[0225] S810, filtering out feature information used for type judgment from multiple feature information to obtain at least one type of feature information.
[0226] The computer analyzes multiple feature information and selects the feature information that is more important and discriminative for type judgment. The goal of screening is to find features that can effectively distinguish different types for subsequent type recognition process.
[0227] In the game scene, the feature information corresponding to some types of factors can strongly express the type of game, such as map feature information, character feature information, sound effect feature information, operation interface feature information and game mechanism feature information, etc. These feature information can be extracted as the basis for type judgment.
[0228] S820: Perform type identification on at least one type of feature information to obtain a type probability distribution of the at least one type of feature information.
[0229] The computer determines the feature information used for type judgment, and can use the corresponding classification model or algorithm to identify the type of these features. By inputting the feature information of the data to be reviewed, the model will output a type probability distribution, indicating the possibility that the data to be reviewed belongs to each type.
[0230] Optionally, the computer can obtain a keyword list, which includes multiple game keywords, and based on the multiple game keywords, perform game keyword matching on at least one type feature information to obtain multiple game keyword matching rates, and use the multiple game keyword matching rates as the type probability distribution of at least one type feature information.
[0231] Specifically, the computer constructs a game keyword library (i.e., a keyword list) containing various keywords related to the game, such as game names, characters, items, etc., wherein the keywords in the keyword library may exist in the form of feature information. Then, the computer matches at least one type feature information with the game keyword library, and calculates the matching degree or similarity between each keyword and at least one type feature information, so as to obtain the type probability distribution of at least one type feature information.
[0232] Next, the computer determines the probability that at least one type feature information belongs to different game types based on the matching evaluation results. For example, it can determine whether the data belongs to a certain game type by setting a threshold or weight. Finally, the type is identified based on the probability to determine which game type the data to be reviewed belongs to.
[0233] S830, calculating the type of the data to be reviewed based on the type probability distribution.
[0234] The computer can calculate the most likely type of the data to be reviewed based on the type probability distribution. Usually, the type with the highest probability value can be selected as the final type judgment result. This final type information can be used for subsequent review processing and decision-making.
[0235] For example, for a scenario where the data to be reviewed is game content, it may be identified that the probability of the game being distributed in shooting games is 70%, the probability of being distributed in simulation games is 20%, and the probability of being distributed in role-playing games is 5%. In this way, the computer can determine that the game type is a shooting game, and therefore it can be determined that the type of the data to be reviewed is a shooting game type.
[0236] Through this method, the computer determines the type of data to be reviewed by calculating the type probability distribution, improves the accuracy of type judgment, provides type probability distribution, and comprehensively considers multiple type information, thereby enhancing the review system's ability to judge the type of data to be reviewed, which helps to improve the accuracy and efficiency of content review.
[0237] In one embodiment of the present application, another data processing method is provided, which can be executed by a computer. Fig. 9 As shown, the data processing method may include S910 to S920, wherein S910 to S920 are Figure 7 The specific implementation method of S720 is shown.
[0238] S910 to S920 are described below:
[0239] S910, obtaining historical audit result information that matches the type of the data to be audited.
[0240] The computer can retrieve the historical audit result database or audit records stored in the audit system to find the historical audit results that match the type of data to be audited. These historical audit results may be audit results of similar or related types of data that have been audited before. By obtaining historical audit result information that matches the type of data to be audited, references to audit results for that type of data can be obtained.
[0241] S920: Based on the type of the data to be reviewed and the historical review result information, weights are assigned to each feature information to obtain a weight list.
[0242] The computer can assign weights to each feature information based on the type of data to be audited and the historical audit results information that has been obtained. The weight assignment can be based on statistical analysis, machine learning models, or manually set rules. By analyzing the correlation between historical audit results and the type of data to be audited, the system can assign different degrees of weight to the feature information based on its importance.
[0243] In this way, for different types of data, the weight of feature information can be adjusted according to its relative importance in the audit process, so that the fused feature information can more accurately reflect the characteristics of the data to be audited, thereby improving the accuracy and adaptability of the audit.
[0244] For example, in a game scenario, the computer analyzes different game types and their corresponding historical audit result information, including audit requirements, audit standards, and statistical data of historical audit results for different types of games. Through these analyses, we can understand the important features of different types of games in the audit and the correlation of audit results. Therefore, we can assign weights to each feature information based on these correlations, and a higher weight means that the feature has a greater influence on the audit decision. In this way, we can get a weight list for the game.
[0245] Through this method, the computer can adjust the weight of feature information according to different types of data and audit history, thereby improving the accuracy and adaptability of fused feature information.
[0246] In one embodiment of the present application, another data processing method is provided, which can be executed by a computer. Fig.10 As shown, the data processing method may include S1010 to S1020, wherein S1010 to S1020 are Figure 7 The specific implementation method of S730 is shown.
[0247] S1010 to S1020 are described below:
[0248] S1010, obtaining the weight of each feature information in the weight list.
[0249] Among them, the weight of the feature information of potential abnormal game data among the multiple feature information included in the weight list is greater than the weight of other feature information among the multiple feature information; the potential abnormal game data includes at least one of the following information of the data to be reviewed: user text information, user voice information, character behavior information and equipment transaction information.
[0250] Based on previous processing and analysis, the computer will extract the weights assigned to each feature information from the weight list. These weights represent the relative importance or contribution of each feature information in the fusion process.
[0251] S1020, performing weighted sum calculation on multiple feature information based on the weight of each feature information to obtain fused feature information.
[0252] The computer will multiply each feature information by its corresponding weight and sum them up to obtain the fused feature information. By weighted summing the feature information, the importance of each feature can be considered and the fused feature information can be adjusted according to its weight to better reflect the multiple features of the data to be reviewed.
[0253] Through this method, the computer can more accurately integrate information of different features by adjusting the weights, thereby improving the ability to express and judge the features of the data to be audited.
[0254] In one embodiment of the present application, another data processing method is provided, which can be executed by a computer. Fig.11 As shown, the data processing method may include S410 to S420, S1110 to S1120 and S440. That is, S1110 to S1120 are Figure 4 The specific implementation method of S430 is shown.
[0255] S1110 to S1120 are described below:
[0256] S1110, selecting a target computing model based on the type of data to be reviewed.
[0257] The computer can select a target computing model suitable for the type of data to be reviewed according to the type of data to be reviewed. The target computing model is a trained machine learning model, deep learning model or other type of model, such as an XGBoost model, which is used to calculate the review pass rate of the data to be reviewed.
[0258] It should be noted that different types of computing models can be stored in a computer, such as a computing model for food, a computing model for travel, a computing model for games, and so on.
[0259] Optionally, the computing models for different food, travel, and games are actually computing models for different fields. In this application, different fields can also be subdivided to divide into more detailed types. For example, for the game field, a computing model of shooting game type, a computing model of role-playing game, a computing model of music game, etc. can be provided; for the food field, a computing model of dessert type, a computing model of spicy food type, a computing model of beverage type, etc. can be provided. The embodiment of this application does not limit the specific type of the computing model. The computer can create a computing model with the data to be reviewed processed by the computer to enrich the types of computing models.
[0260] S1120, using the target calculation model to predict the fused feature information and output the audit pass rate.
[0261] The computer will input the fusion feature information into the selected target calculation model and use the model to predict or classify the audited data. The output result is the audit pass rate, that is, the model's estimated probability of the audited data passing the audit. The pass rate can be used as the basis for subsequent audit resource allocation.
[0262] In the gaming scenario, the audit pass rate can be evaluated from the dimension of whether the game content represented by the data to be audited contains any violations. Such violations may include users using third-party software to cause an imbalance in the gaming environment, deliberately not cooperating with teammates in multiplayer online games, and verbally attacking in chat-enabled games.
[0263] Optionally, the audit pass rate may be a pass probability calculated after integrating the situations of multiple violations, or multiple audit pass rates for the data to be audited may be calculated for different violations.
[0264] like Fig.12 As shown, the embodiment of the present application converts the traditional priority sorting problem into calculating the audit pass rate. The traditional priority sorting problem is mainly that it is difficult to define how to prioritize the audit data, and artificial empiricism is often included in the priority sorting, resulting in non-objective sorting. However, the present application removes the artificial empirical definition, makes full use of historical audit data, makes full use of machine recognition capabilities, and mines feature information of more factors of the audited data, so that the data to be audited can be more comprehensively characterized. In this way, the computer can estimate the pass probability of the audited data through the predictive ability of the model, and provide quantitative reference indicators for the priority sorting problem.
[0265] In one embodiment of the present application, another data processing method is provided, which can be executed by a computer. Fig.13As shown, the data processing method may include S410 to S430 and S1310 to S1320. That is, S1310 to S1320 are Figure 4 The specific implementation method of S440 is shown.
[0266] S1310 to S1320 are described below:
[0267] Allocate audit resources to the audited data based on the audit pass rate, including:
[0268] S1310, prioritizing the data to be audited based on the audit pass rate.
[0269] The computer can sort the audit data according to the calculated audit pass rate. Data with a high audit pass rate will have a higher priority, while data with a low audit pass rate will have a lower priority. Through this sorting process, the audit data can be prioritized according to the possibility of passing the audit, so as to facilitate subsequent resource allocation and processing.
[0270] S1320, allocating audit resources to the data to be audited in the order of priority, wherein the order of allocating audit resources is positively correlated with the priority of the data to be audited.
[0271] The computer sorts the priorities of the data to be audited and allocates resources to the data to be audited in order of priority. Specifically, audit resources (such as auditors, computing resources, etc.) will be allocated from high to low priority of the data to be audited, with higher priority data being processed first. This ensures that limited audit resources can be more effectively used to audit data that is more likely to pass, thereby improving the efficiency and accuracy of the audit.
[0272] Optionally, for content with a lower priority, if it has not been allocated review resources for a long time, a prompt may be output to ensure timely allocation of review resources.
[0273] Through this method, the efficiency of the audit can be improved, the audit resources can be allocated more reasonably, and the data with a higher probability of passing can be processed first.
[0274] Among them, S1310 may include S1311 to S1313.
[0275] S1311 to S1313 are described below:
[0276] S1311, obtaining an audit pass rate table, in which the audit pass rates of multiple predicted data are recorded in a preset order.
[0277] The computer can collect and record the audit pass rates of multiple predicted data. These pass rates can be previously audited data or the results of the model's prediction of sample data. The purpose of the audit pass rate table is to provide a reference for prioritizing the audited data.
[0278] S1312, comparing the audit pass rate of the data to be audited with the audit pass rates of multiple predicted data to determine the target ranking position.
[0279] The computer will compare the audit pass rate of the data to be audited with the audit pass rates of each predicted data in the audit pass rate table. By comparison, the target ranking position of the data to be audited in the audit pass rate table can be determined. Specifically, the computer will find the appropriate position of the audit pass rate of the data to be audited in the audit pass rate table to determine its relative priority in the entire ranking.
[0280] Exemplarily, the computer calculates that the audit pass rate of the data to be audited is 80%, and the audit pass rate table is shown in Table 1, with predicted data 1 to predicted data 5. Then the target position is between predicted data 2 (88%) and predicted data 3 (75%).
[0281]
[0282]
[0283] Table 1
[0284] S1313, adding the audit pass rate of the data to be audited to the target sorting position to obtain an updated audit pass rate table.
[0285] The computer will add the audit pass rate of the data to be audited to the target sorting position in the audit pass rate table, thereby obtaining an updated audit pass rate table. By adding the audit pass rate of the data to be audited to the table, it can be ensured that it is in the appropriate position in the entire sorting for subsequent resource allocation and processing.
[0286] Exemplarily, for the above example, the data to be reviewed can be added between the predicted data 2 (88%) and the predicted data 3 (75%) to obtain an updated review pass rate table as shown in Table 2.
[0287] Predicted data Audit pass rate Predicted data 1 90% Predicted data 2 88% Data awaiting review 80% Predicted data 3 75% Predicted data 4 60% Predicted data 5 55%
[0288] Table 2
[0289] This method can improve audit efficiency, improve audit accuracy, optimize resource allocation, and provide valuable reference for manual audit decisions. By focusing audit resources on data with a higher audit pass rate, the system can handle audit tasks more efficiently and reduce resource waste; at the same time, by comparing the audit pass rate of the data to be audited with the audit pass rate of the predicted data, determining the target ranking position and updating the audit pass rate table, the system can continuously optimize audit decisions and improve the accuracy and consistency of audits. This provides a more efficient and accurate processing method for content audits, which improves the efficiency of audit work.
[0290] In one embodiment of the present application, another data processing method is provided, which can be executed by a computer. Fig.14 As shown, the data processing method may include S410 to S440 and S1410 to S1420.
[0291] S1410 to S1420 are described below:
[0292] S1410, obtaining the audit result corresponding to the data to be audited.
[0293] The audit result may be an audit result of the data to be audited by the allocated audit resources, and the audit result may include information on whether the data to be audited has passed the audit.
[0294] S1420, adjusting parameters of the target computing model based on the audit results.
[0295] The computer can optimize the performance and accuracy of the target computing model based on the feedback information of the audit results. By analyzing the results of the audited data, we can understand the deviation between the audit results of the target computing model and the manual audit results. Based on this information, we can adjust the parameters of the target computing model to improve the audit accuracy and effect of the model.
[0296] Through this method, the computer can obtain and analyze the audit results of the audited data, and then optimize and adjust the target calculation model. Such a feedback loop can continuously improve the performance and accuracy of the audit system, enabling it to better respond to the audit needs of different types of data, improve audit efficiency and the accuracy of resource allocation.
[0297] The following is an explanation of the effect of this application on improving the review efficiency in conjunction with Table 3.
[0298]
[0299] Table 3
[0300] As shown in Table 3, the audit time was reduced from 72 minutes to 35 minutes, a decrease of 50%. The proportion of audits completed within half an hour increased from 61% to 71%, an increase of 16%. The overall audit pass rate increased from 12% to 30%, an increase of 150%. Therefore, the audit speedup ratio is 2.6.
[0301] In one embodiment of the present application, for easier understanding, the following description is made by taking the data to be reviewed as a game video as an example. The game video may contain text content and video content. For the game video, the following steps may be used to allocate review resources:
[0302] 1. Basic video feature extraction: For the game video, the computer will perform feature extraction, including basic attributes such as video length, resolution, frame rate, etc., and use image processing technology to extract features of the image content, such as object information, location information, and scene information.
[0303] 2. Deep video feature extraction: Computers can also perform deep feature extraction on video content. By using action recognition models and speech recognition models, they can identify action information and speech information in the video and extract deep video features.
[0304] 3. Video visual feature extraction: The computer uses pre-trained computer vision models, such as ResNet, VGG, etc., to extract the visual feature vector of the video.
[0305] 4. Operation recognition: The computer uses a target computing model that has learned the characteristic patterns of cheating operations and the differences from normal game operations during the training process. By predicting the fused feature information, the computer can determine whether there are any illegal cheating operations in the video.
[0306] 5. Calculation of review pass rate: The computer calculates the review pass rate of the game video based on the prediction results, which represents the possibility of cheating in the video.
[0307] 6. Audit resource allocation: Based on the audit pass rate, the computer can prioritize audit resources to videos with lower pass rates for further manual review or in-depth analysis to ensure accurate identification of violations.
[0308] Through the above method, the computer can combine multiple feature information to determine whether there is cheating in the game video. This comprehensive detection method can improve the accuracy and efficiency of violation detection, help game platforms to promptly discover and respond to violations, and maintain the fairness of the game and a healthy gaming environment.
[0309] In one embodiment of the present application, the data processing method of the present application can also be applied to the scenario of real-time violation detection of games. In this scenario, the data to be reviewed can be game background data, which can be obtained with the consent of the user. The following steps can be used to determine whether the user's game background data shows any illegal operations:
[0310] 1. Data collection: The computer obtains the game background data agreed by the user, which includes the player's game operations, game behaviors and other related information. Data collection can be carried out through the game client, game server or other means.
[0311] 2. Feature extraction: The computer extracts relevant feature information from the game background data. These features may include the player's operation instructions, the acquisition and use of game items, interactive behaviors in the game, game scores, network status, device information, geographic location, etc. By extracting feature information, descriptive data on player behavior can be obtained.
[0312] 3. Model training: Use existing game rules or machine learning-based methods to train the computing model through the saved historical game background data to determine whether there are any illegal operations in the game background data. The computing model can be a rule-based logical model or a model trained using a machine learning algorithm (such as decision tree, random forest, deep learning, etc.).
[0313] 4. Feature fusion: The computer can obtain the corresponding weight list based on the type of the game and perform feature fusion.
[0314] 5. Violation detection: Input the game background data to be reviewed into the trained computing model (i.e. the target computing model for the scenario) for violation detection. The model will analyze the characteristics of the game background data and output the violation probability of the game background data, that is, the probability that the data shows a violation operation.
[0315] 6. Audit resource allocation: Prioritize the violation probability of each game background data, and allocate audit resources in the order of priority to determine whether there is any violation through manual review.
[0316] By extracting the characteristic information of the above factors, we can fully understand the behavior and status of players in the game, and provide basic data support for subsequent tasks such as violation detection and behavior analysis. These characteristic information can be used to identify illegal operations and evaluate player behavior, thereby maintaining the fairness and order of the game and providing a better user experience.
[0317] In one embodiment of the present application, the data processing method of the present application can also be applied to the scenario of predicting the click-through rate (CTR) of news articles. In this scenario, the news article includes image content and text content. The click-through rate refers to the probability that the user clicks to read the news article after seeing it. The specific prediction method of the click-through rate is as follows:
[0318] 1. Text feature extraction: Use pre-trained natural language processing models (such as BERT) to extract feature information such as semantic feature vectors of article titles and text content. At the same time, algorithms such as TF-IDF or TextRank can be used to extract keywords and perform sentiment analysis.
[0319] 2. Image feature extraction: Use the image feature extraction algorithm to extract features from all image contents designed in the news article.
[0320] 3. Social feature extraction: Convert the posting time into categorical features (such as morning, afternoon, and evening), and then process it using one-hot encoding.
[0321] 4. Normalization of numerical features: Normalize numerical features such as the number of author followers, number of pictures, picture size, and historical click-through rate.
[0322] 5. Model training: Assuming that there are historical click-through rate records as shown in Table 4, features can be extracted for each historical click-through rate record and input into a computational model, such as an XGBoost model, to obtain a trained computational model, which can be used as the target computational model for predicting the click-through rate of news articles.
[0323]
[0324] Table 4
[0325] Among them, after extracting the features of the three historical click rate records shown in Table 3, we can get the following
[0326] Multiple feature vectors shown in Table 5:
[0327]
[0328] Table 5
[0329] 6. Click-through rate prediction: Input the news article to be reviewed into the trained computing model to predict the click-through rate. The model will analyze the characteristics of the news article and output the estimated click-through rate of the news article, that is, the probability that the user will click to read the news article after seeing it.
[0330] It should be noted that since it is applied in the click rate prediction scenario of news articles, such as Fig.12The target calculation model shown will output the click-through rate estimate. Accordingly, the content scheduling module will not work. The lighthouse data channel can collect the actual click-through rate and update the parameters of the model to optimize the target calculation model.
[0331] By extracting the characteristic information of the above factors, we can fully understand the characteristics of news articles and authors, and provide basic data support for subsequent tasks such as click-through rate prediction.
[0332] Fig.15 FIG. 1 is a block diagram of a data processing device according to an embodiment of the present application. Fig.15 As shown, the device comprises:
[0333] The processing unit 1510 is used to perform feature extraction processing on the data to be audited to obtain multiple feature information;
[0334] A fusion unit 1520 is used to fuse multiple feature information based on weight information corresponding to the type of the data to be reviewed to obtain fused feature information;
[0335] A calculation unit 1530, configured to calculate the review pass rate of the data to be reviewed based on the fused feature information;
[0336] The allocation unit 1540 is used to allocate audit resources to the data to be audited based on the audit pass rate.
[0337] In one embodiment of the present application, based on the aforementioned scheme, the data to be reviewed includes game text content; the processing unit 1510 is also used to analyze at least one of the text length, text structure and word frequency information of the game text content to obtain basic text feature information; feature extract the naming text in the game text content to obtain deep text feature information; wherein the naming text includes at least one of the following information of the game text content: game character name information, equipment name information, map location information, game music information, game task information and skill name information; keyword extraction is performed on the game text content to obtain the keyword text information. In one embodiment of the present application, based on the aforementioned scheme, the data to be reviewed includes game image content; the processing unit 1510 is also used to analyze at least one of the number of images, image size and image type of the game image content to obtain basic image feature information; identify the game static features in the game image content to obtain depth image feature information; wherein the game static features include at least one of the following information of the game image content: equipment information, map location information, character modeling information, game plot information and character status information; perform visual detection calculations on the game image content to obtain image visual feature information.
[0338] In one embodiment of the present application, based on the aforementioned scheme, the data to be reviewed includes game video content; the processing unit 1510 is also used to extract basic attributes of the game video content to obtain basic video feature information; wherein the basic attributes include at least one of the following information of the game video content: duration, resolution, frame rate and format; the game dynamic features in the game video content are identified to obtain deep video feature information; wherein the game dynamic features include at least one of the following information of the game video content: character action information, voice information, map scene information, character modeling information, character status information, game music information and game asset information; visual detection calculations are performed on the game video content to obtain video visual feature information.
[0339] In one embodiment of the present application, based on the aforementioned scheme, the calculation unit 1530 is also used to perform type calculation based on multiple feature information to determine the type of the data to be reviewed; the processing unit 1510 is also used to obtain a weight list based on the determined type, and the weight list includes the weight of each feature information; the fusion unit 1520 is also used to fuse multiple feature information based on the weight list to obtain fused feature information.
[0340] In one embodiment of the present application, based on the aforementioned scheme, the processing unit 1510 is further used to filter out feature information used for type judgment from multiple feature information to obtain at least one type feature information; perform type identification on at least one type feature information to obtain a type probability distribution of at least one type feature information; and calculate the type of the data to be reviewed based on the type probability distribution.
[0341] In one embodiment of the present application, based on the aforementioned scheme, the processing unit 1510 is further used to perform type identification on the at least one type feature information to obtain a type probability distribution of the at least one type feature information, including: obtaining a keyword list, the keyword list including a plurality of game keywords; based on the plurality of game keywords, performing game keyword matching on the at least one type feature information to obtain a plurality of game keyword matching rates; and using the plurality of game keyword matching rates as the type probability distribution of the at least one type feature information.
[0342] In one embodiment of the present application, based on the aforementioned scheme, the processing unit 1510 is also used to obtain historical audit result information that matches the type of the data to be audited; based on the type of the data to be audited and the historical audit result information, a weight is assigned to each feature information to obtain a weight list.
[0343] In one embodiment of the present application, based on the aforementioned scheme, the processing unit 1510 is further used to obtain the weight of each feature information in the weight list; wherein the weight of the feature information of the potential abnormal game data in the multiple feature information contained in the weight list is greater than the weight of other feature information in the multiple feature information; the potential abnormal game data includes at least one of the following information of the data to be reviewed: user text information, user voice information, character behavior information and equipment transaction information; based on the weight of each feature information, a weighted sum calculation is performed on the multiple feature information to obtain the fused feature information.
[0344] In one embodiment of the present application, based on the aforementioned solution, the processing unit 1510 is further used to select a target computing model based on the type of data to be reviewed; use the target computing model to predict the fused feature information, and output the review pass rate.
[0345] In one embodiment of the present application, based on the aforementioned scheme, after allocating audit resources to the data to be audited based on the audit pass rate, the processing unit 1510 is also used to obtain the audit results corresponding to the data to be audited; and adjust the parameters of the target computing model based on the audit results.
[0346] In one embodiment of the present application, based on the aforementioned scheme, the processing unit 1510 is also used to prioritize the data to be audited based on the audit pass rate; and to allocate audit resources to the data to be audited in the order of priority ranking, wherein the order of allocating audit resources is positively correlated with the priority ranking of the data to be audited.
[0347] In one embodiment of the present application, based on the aforementioned scheme, the processing unit 1510 is also used to obtain an audit pass rate table, in which the audit pass rates of multiple predicted data are recorded in a preset order; the audit pass rate of the data to be audited is compared with the audit pass rates of multiple predicted data to determine the target sorting position; the audit pass rate of the data to be audited is added to the target sorting position to obtain an updated audit pass rate table.
[0348] It should be noted that the apparatus provided in the aforementioned embodiment and the method provided in the aforementioned embodiment belong to the same concept, wherein the specific manner in which each module and unit performs the operation has been described in detail in the method embodiment.
[0349] An embodiment of the present application also provides a data processing device, including: one or more processors; a memory for storing one or more programs, when the one or more programs are executed by the one or more processors, the electronic device implements the above data processing method.
[0350] Fig.16It is a structural diagram of a computer system suitable for implementing a data processing device of an embodiment of the present application.
[0351] It should be noted that Fig.16 The computer system 1600 of the electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.
[0352] like Fig.16 As shown, the computer system 1600 includes a central processing unit (CPU) 1601, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 1602 or the program loaded from the storage part 1608 to the random access memory (RAM) 1603, such as executing the method in the above embodiment. In RAM 1603, various programs and data required for system operation are also stored. CPU 1601, ROM 1602 and RAM 1603 are connected to each other through bus 1604. Input / output (I / O) interface 1605 is also connected to bus 1604.
[0353] The following components are connected to the I / O interface 1605: an input section 1606 including a keyboard, a mouse, etc.; an output section 1607 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1608 including a hard disk, etc.; and a communication section 1609 including a network interface card such as a LAN (Local Area Network) card, a modem, etc. The communication section 1609 performs communication processing via a network such as the Internet. A drive 1610 is also connected to the I / O interface 1605 as needed. A removable medium 1611, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 1610 as needed so that a computer program read therefrom is installed into the storage section 1608 as needed.
[0354] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication section 1609, and / or installed from a removable medium 1611. When the computer program is executed by a central processing unit (CPU) 1601, various functions defined in the system of the present application are executed.
[0355] It should be noted that the computer-readable medium shown in the embodiment of the present application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. Computer-readable media may be, for example, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples of computer-readable media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable medium may be any tangible medium containing or storing a program, which may be used by an instruction execution system, device or device or used in combination with it. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, wherein a computer-readable computer program is carried. This propagated data signal may take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. A computer program contained on a computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.
[0356] The flowchart and block diagram in the accompanying drawings illustrate the possible architecture, functions and operations of the system, method and computer program product according to various embodiments of the present application. Wherein, each box in the flowchart or block diagram can represent a module, a program segment, or a part of the code, and the above-mentioned module, program segment, or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.
[0357] The units involved in the embodiments described in this application may be implemented by software or hardware, and the units described may also be set in a processor. The names of these units do not, in some cases, constitute limitations on the units themselves.
[0358] Another aspect of the present application also provides a computer readable medium having a computer program stored thereon, which implements the above data processing method when the computer program is executed by a processor. The computer readable medium may be included in the electronic device described in the above embodiment, or may exist independently without being assembled into the electronic device.
[0359] Another aspect of the present application also provides a computer program product or a computer program, which includes a computer instruction stored in a computer-readable medium. A processor of a computer device reads the computer instruction from the computer-readable medium, and the processor executes the computer instruction, so that the computer device executes the data processing method provided in each of the above embodiments.
[0360] The above content is only a preferred exemplary embodiment of the present application and is not intended to limit the implementation scheme of the present application. A person skilled in the art can easily make corresponding changes or modifications based on the main concept and spirit of the present application. Therefore, the scope of protection of the present application shall be based on the scope of protection required by the claims.
Claims
1. A data processing method, It is characterized in that include: Perform feature extraction processing on the data to be audited to obtain multiple feature information; Based on the weight information corresponding to the type of the data to be reviewed, the plurality of feature information are fused to obtain fused feature information; Calculating the review pass rate of the data to be reviewed based on the fused feature information; Audit resources are allocated to the data to be audited based on the audit pass rate.
2. The method according to claim 1, It is characterized in that The data to be reviewed includes game text content; The feature extraction process is performed on the data to be audited to obtain multiple feature information, including: Analyze at least one of the text length, text structure and word frequency information of the game text content to obtain basic text feature information; Performing feature extraction on the named text in the game text content to obtain deep text feature information; wherein the named text includes at least one of the following information of the game text content: game character name information, equipment name information, map location information, game music information, game task information and skill name information; Keyword extraction is performed on the game text content to obtain the keyword text information.
3. The method according to claim 1, It is characterized in that The data to be reviewed includes game image content; The feature extraction process is performed on the data to be audited to obtain multiple feature information, including at least two of the following methods: Analyze at least one of the number of images, image size, and image type of the game image content to obtain basic image feature information; Identify the game static features in the game image content to obtain depth image feature information; wherein the game static features include at least one of the following information of the game image content: equipment information, map location information, character modeling information, game plot information, and character status information; Visual detection calculation is performed on the game image content to obtain image visual feature information.
4. The method according to claim 1, It is characterized in that The data to be reviewed includes game video content; The feature extraction process is performed on the data to be audited to obtain multiple feature information, including at least two of the following methods: Extracting basic attributes of the game video content to obtain basic video feature information; wherein the basic attributes include at least one of the following information of the game video content: duration, resolution, frame rate, and format; Identify the game dynamic features in the game video content to obtain deep video feature information; wherein the game dynamic features include at least one of the following information of the game video content: character action information, voice information, map scene information, character modeling information, character status information, game music information and game asset information; Visual detection calculation is performed on the game video content to obtain video visual feature information.
5. The method according to claim 1, It is characterized in that The step of fusing the plurality of feature information based on the weight information corresponding to the type of the data to be reviewed to obtain fused feature information includes: Performing type calculation based on the multiple feature information to determine the type of the data to be reviewed; Acquire a weight list based on the determined type, the weight list including a weight of each feature information; The plurality of feature information are fused based on the weight list to obtain the fused feature information.
6. The method according to claim 5, It is characterized in that The performing type calculation based on the plurality of feature information to determine the type of the data to be reviewed includes: Filtering the feature information used for type determination from the plurality of feature information to obtain at least one type of feature information; Performing type identification on the at least one type feature information to obtain a type probability distribution of the at least one type feature information; The type of the data to be reviewed is calculated based on the type probability distribution.
7. The method according to claim 6, It is characterized in that The performing type identification on the at least one type feature information to obtain a type probability distribution of the at least one type feature information includes: Obtain a keyword list, wherein the keyword list includes a plurality of game keywords; Based on the multiple game keywords, performing game keyword matching on the at least one type feature information to obtain multiple game keyword matching rates; The plurality of game keyword matching rates are used as type probability distribution of the at least one type feature information.
8. The method according to claim 5, It is characterized in that The obtaining of a weight list based on the determined type includes: Obtaining historical audit result information that matches the type of the data to be audited; Based on the type of the data to be reviewed and the historical review result information, a weight is assigned to each feature information to obtain the weight list.
9. The method according to claim 5, It is characterized in that The fusing the plurality of feature information based on the weight list to obtain the fused feature information includes: Obtaining the weight of each feature information in the weight list; wherein the weight of the feature information of the potential abnormal game data in the plurality of feature information contained in the weight list is greater than the weight of other feature information in the plurality of feature information; the potential abnormal game data includes at least one of the following information of the data to be reviewed: user text information, user voice information, character behavior information, and equipment transaction information; The plurality of feature information are weighted and summed based on the weight of each feature information to obtain the fused feature information.
10. The method according to claim 5, It is characterized in that The calculating the audit pass rate of the data to be audited based on the fused feature information includes: Selecting a target computing model based on the type of the data to be reviewed; The target calculation model is used to predict the fused feature information and output the audit pass rate.
11. The method according to claim 10, It is characterized in that After allocating audit resources to the data to be audited based on the audit pass rate, the method further includes: Obtaining the audit result corresponding to the data to be audited; Parameters of the target computing model are adjusted based on the audit results.
12. The method according to claim 1, It is characterized in that The allocating audit resources to the data to be audited based on the audit pass rate includes: Prioritizing the data to be reviewed based on the review pass rate; Audit resources are allocated to the data to be audited in the order of priority, wherein the order of allocating audit resources is positively correlated with the priority of the data to be audited.
13. The method according to claim 12, It is characterized in that The prioritizing the data to be reviewed based on the review pass rate includes: Obtaining an audit pass rate table, wherein the audit pass rate table records the audit pass rates of a plurality of predicted data in a preset order; Compare the audit pass rate of the data to be audited with the audit pass rates of the plurality of predicted data to determine the target ranking position; The audit pass rate of the data to be audited is added to the target sorting position to obtain an updated audit pass rate table.
14. A data processing device, It is characterized in that include: A processing unit, used for performing feature extraction processing on the data to be audited to obtain multiple feature information; A fusion unit, configured to fuse the plurality of feature information based on weight information corresponding to the type of the data to be reviewed to obtain fused feature information; A calculation unit, used for calculating the review pass rate of the data to be reviewed based on the fused feature information; An allocation unit is used to allocate audit resources to the data to be audited based on the audit pass rate.
15. A data processing device, It is characterized in that include: a memory storing computer-readable instructions; A processor reads the computer-readable instructions stored in the memory to execute the method according to any one of claims 1 to 13.
16. A computer-readable storage medium, It is characterized in that Computer-readable instructions are stored thereon, and when the computer-readable instructions are executed by a processor of a computer, the computer is caused to execute the method according to any one of claims 1 to 13.
17. A computer program product, It is characterized in that The computer program product comprises a computer program, wherein the computer program is stored in a computer-readable storage medium, and a processor of an electronic device reads and executes the computer program from the computer-readable storage medium, so that the electronic device executes the method according to any one of claims 1 to 13.
Citation Information
Patent Citations
Video auditing method and device, equipment and storage medium
CN109495766A
Game data processing method, device and system, electronic equipment and storage medium
CN113181634A
Intelligent delivery of actionable content
WO2015200407A1
Cited By
Instruction approval data processing method and system based on multiple neural networks
CN120013457A
Comment sorting method and device, electronic equipment and storage medium
CN120407962A
