A data detection method, device, apparatus and storage medium

Abstract

The application provides a data detection method and device, equipment and a storage medium. The method comprises the following steps: in response to a data detection operation, obtaining target data, the target data comprising N elements, the elements comprising any one of text, images and videos; determining first feature information corresponding to the N elements; determining risk levels corresponding to M combinations according to the first feature information corresponding to the N elements, each combination in the M combinations being composed of any two sub-elements in a plurality of sub-elements included in the N elements, the sub-elements being text or images; determining a risk level of the target data according to the risk levels corresponding to the M combinations; and sending the risk level of the target data to a display device for display. In this way, in the data detection process, the risk level of the target data can be accurately determined by comprehensively considering various element information in the target data.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence (AI) technology, and in particular to a data detection method, apparatus, device and storage medium. Background Technology

[0002] With the rapid development of internet technology, people can easily access massive amounts of information and resources, and the connection between life and production and the internet is becoming increasingly close. However, as the scale of the network gradually expands, there is also a lot of illegal internet information. How to use data detection methods to judge the risk level of complex internet information in order to filter out internet information with a higher risk level has become another hot topic in current internet research.

[0003] Current data detection methods, such as those used for advertising, determine whether an ad meets preset requirements by analyzing the text information within the ad. However, current data detection methods are inaccurate. Summary of the Invention

[0004] This application provides a data detection method, apparatus, device, and storage medium that can comprehensively and accurately determine the risk level of Internet information.

[0005] In a first aspect, embodiments of this application provide a data detection method, including:

[0006] In response to the data detection operation, acquire the target data, which includes N elements, where each element can be any one of text, image, or video, and N is a positive integer greater than 1.

[0007] Determine the first feature information corresponding to N elements, and determine the risk level corresponding to M combinations based on the first feature information corresponding to N elements. Each of the M combinations consists of any two sub-elements from the multiple sub-elements included in the N elements. The sub-elements are text or images, and M is a positive integer greater than 1.

[0008] Based on the risk levels corresponding to the M combinations, determine the risk level of the target data and send the risk level of the target data to the display device for display.

[0009] Secondly, embodiments of this application provide a data detection device, comprising:

[0010] The acquisition unit is used to acquire target data in response to the data detection operation. The target data includes N elements, which can be any one of text, image and video, where N is a positive integer greater than 1.

[0011] The first determining unit is used to determine the first feature information corresponding to N elements, and to determine the risk level corresponding to M combinations based on the first feature information corresponding to N elements. Each of the M combinations consists of any two sub-elements from the multiple sub-elements included in the N elements. The sub-elements are text or images, and M is a positive integer greater than 1.

[0012] The second determining unit is used to determine the risk level of the target data based on the risk levels corresponding to the M combinations, and send the risk level of the target data to the display device for display.

[0013] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor; the memory is used to store a computer program, and the processor is used to call and run the computer program stored in the memory to perform the methods in the first aspect or its various implementations described above.

[0014] Fourthly, a chip is provided for implementing the methods of the first aspect or its various implementations described above. Specifically, the chip includes a processor for retrieving and running a computer program from a memory, causing a device equipped with the chip to perform the methods of the first aspect or its various implementations described above.

[0015] Fifthly, a computer-readable storage medium is provided for storing a computer program that causes a computer to perform the methods described in the first aspect or its various implementations.

[0016] In a sixth aspect, a computer program product is provided, including computer program instructions that cause a computer to perform the methods described in the first aspect or its various implementations.

[0017] In a seventh aspect, a computer program is provided that, when run on a computer, causes the computer to perform the methods described in the first aspect or its various implementations.

[0018] In summary, the technical solution of this application, when performing data detection, in response to the data detection operation, acquires target data, which includes N elements, each of which can be any one of text, image, and video; determines the first feature information corresponding to the N elements, and based on the first feature information corresponding to the N elements, determines the risk level corresponding to M combinations, where each of the M combinations consists of any two sub-elements from the multiple sub-elements included in the N elements, and the sub-elements can be text or image; based on the risk level corresponding to the M combinations, determines the risk level of the target data, and sends the risk level of the target data to a display device for display. That is, in this application, when performing data detection, firstly, a comprehensive analysis is performed on each element (e.g., text, image, video, etc.) included in the target data to obtain the first feature information corresponding to each of the N elements included in the target data. Furthermore, to improve detection accuracy, the sub-elements included in the N elements are combined in pairs, and the risk level corresponding to each combination is analyzed to achieve small-granular risk analysis. Then, based on the risk levels corresponding to the pairs of sub-elements, the risk level of the target data is accurately determined, thus improving the accuracy of data detection. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of an application scenario provided by an embodiment of this application;

[0021] Figure 2 A schematic flowchart illustrating a data detection method provided in an embodiment of this application;

[0022] Figure 3 This application provides an example of an event pattern diagram.

[0023] Figure 4 A schematic diagram of the network structure of a text semantic analysis model provided in this application embodiment;

[0024] Figure 5 This is a schematic diagram of the network structure of an image recognition model provided in an embodiment of this application;

[0025] Figure 6 A schematic diagram of a DAG provided for an embodiment of this application;

[0026] Figure 7 A schematic diagram of the LTR principle provided for embodiments of this application;

[0027] Figure 8 A schematic diagram illustrating the principle of the wide & deep model provided in this application embodiment;

[0028] Figure 9 A schematic diagram of the LTR risk level prediction method provided in the embodiments of this application;

[0029] Figure 10 This application provides a schematic diagram of the execution process of a distributed system.

[0030] Figure 11 This is a schematic diagram of the structure of the data detection device provided in the embodiments of this application;

[0031] Figure 12 A schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0034] Before introducing the technical solution of this application, the relevant knowledge of the technical solution of this application will be explained below:

[0035] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0036] It should be understood that artificial intelligence (AI) technology is a comprehensive discipline involving a wide range of fields, encompassing both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0037] With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robots, smart healthcare, and smart customer service. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.

[0038] OCR (Optical Character Recognition) refers to the process by which electronic devices (such as scanners or digital cameras) examine printed characters on paper, determine their shapes by detecting dark and light patterns, and then translate those shapes into computer text using character recognition methods. In other words, for printed characters, it uses optical methods to convert the text in a paper document into a black-and-white dot matrix image file, and then uses recognition software to convert the text in the image into text format for further editing by word processing software. How to correct errors or utilize auxiliary information to improve recognition accuracy is the most important issue in OCR, hence the term ICR (Intelligent Character Recognition). The main indicators for evaluating the performance of an OCR system include: rejection rate, false recognition rate, recognition speed, user interface friendliness, product stability, ease of use, and feasibility.

[0039] Web (World Wide Web) services are independent, loosely coupled, self-contained, programmable web-based applications. These applications can be described, published, discovered, coordinated, and configured using the open XML (Extensible Markup Language) standard for developing distributed, interactive applications. Web service technology enables different applications running on different machines to exchange data or integrate without the need for additional, specialized third-party software or hardware. Applications implemented according to Web Service specifications can exchange data regardless of their language, platform, or internal protocols. Web Services are self-describing, self-contained, and available network modules that can perform specific business functions. Web Services are also easy to deploy because they are based on common industry standards and existing technologies such as XML (a subset of Standard Generalized Markup Language) and HTTP (Hypertext Transfer Protocol). Web Services reduce the cost of application interfaces. Web Services provide a common mechanism for integrating business processes across an entire enterprise or even multiple organizations.

[0040] Spring Boot is a new framework designed to simplify the initial setup and development process of new Spring applications. This framework uses specific configuration methods, eliminating the need for developers to define boilerplate configurations.

[0041] The server is responsible for serving the client, and its services include providing resources to the client and storing client data.

[0042] Web front-end is the process of creating front-end interfaces such as web pages or apps (applications) to present to users. It uses HTML (Hypertext Markup Language), CSS (Cascading Style Sheets), and JavaScript (JS) as well as various derived technologies, frameworks, and solutions to realize the user interface interaction of Internet products.

[0043] Hadoop (Hadoop Distributed Computing) is a distributed system infrastructure developed by the Apache Software Foundation. Users can develop distributed programs without understanding the underlying details of distributed systems, fully leveraging the power of clusters for high-speed computation and storage. Hadoop implements a distributed file system, one component of which is HDFS (Hadoop Distributed File System). HDFS is highly fault-tolerant and designed to be deployed on low-cost hardware; it also provides high throughput access to application data, making it suitable for applications with very large datasets. HDFS relaxes POSIX (Portable Operating System Interface of UNIX) requirements, allowing streaming access to data in the file system. The core design of the Hadoop framework consists of HDFS and MapReduce. HDFS provides storage for massive amounts of data, while MapReduce provides computation for massive amounts of data.

[0044] Yarn (Yet Another Resource Negotiator) is a new Hadoop resource manager. It is a general-purpose resource management system that provides unified resource management and scheduling for upper-layer applications. Its introduction has brought great benefits to the cluster in terms of utilization, unified resource management and data sharing.

[0045] Kubernetes (or K8s for short) is an open-source application used to manage containerized applications across multiple hosts in a cloud platform. Kubernetes aims to make deploying containerized applications simple and efficient, providing a mechanism for application deployment, planning, updating, and maintenance.

[0046] APK (Android Application Package) is an application package file format used by the Android operating system for distributing and installing mobile applications and middleware. For an Android application's code to run on an Android device, it must first be compiled and then packaged into a file that the Android system can recognize and run. This file format that the Android system can recognize and run is the "APK". An APK file contains compiled code files (.dex files), file resources, native resource files, certificates, and a manifest file.

[0047] Advertising creativity refers to the use of unique techniques or ingenious advertising scripts to highlight product characteristics and brand connotations, thereby promoting product sales. Advertising creativity consists of two main parts: advertising appeal and advertising execution. It encompasses creative thinking in advertising activities; anything involving creating new aspects, from strategy and image to tactics and media selection, is considered "creative." In reality, the advertising industry prefers to define advertising creativity as "the creative thinking behind advertising works." Simply put, advertising creativity is about creating unique audiovisual effects through bold and novel methods to maximize consumer attraction, thereby achieving brand awareness and product marketing goals. The English expression for advertising creativity is "idea & creative." Advertising creativity refers to the creative expression of a brand's sales message in advertising to cater to or guide consumer psychology and encourage purchasing behavior. Advertising creativity consists of two main parts: advertising appeal and advertising execution.

[0048] A database is a "repository for organizing, storing, and managing data according to a data structure." It is a large collection of organized, shareable, and uniformly managed data that is stored permanently in a computer.

[0049] Distributed storage is a data storage technology that uses the disk space on every machine in an enterprise through a network, and combines these distributed storage resources into a virtual storage device, with data stored in various parts of the enterprise.

[0050] A DAG (Directed Acyclic Graph) is a directed graph in mathematics, particularly graph theory and computer science, where there are no cycles. If a non-directed acyclic graph exists where a path from point A to B, through point C, and back to A forms a cycle, changing the direction of the edge from C to A to A to C transforms it into a directed acyclic graph. The number of spanning trees in a directed acyclic graph is equal to the product of the in-degrees of its nodes with non-zero in-degree.

[0051] TextCNN (Text Convolutional Neural Networks) was proposed in the paper "Convolutional Neural Networks for Sentence Classification." It applies Convolutional Neural Networks (CNNs) to text classification tasks, utilizing multiple convolutional kernels of different scales to extract key information from sentences (similar to multi-window Chinese language models), thus better capturing local correlations. Compared to traditional image CNN networks, TextCNN has no changes in network structure, and is even simpler. TextCNN has only one convolutional layer, one max-pooling layer, and finally, the output is externally appended with a softmax (classifier) ​​to classify n types.

[0052] LSTM (Long Short-Term Memory) networks typically outperform recurrent neural networks and Hidden Markov Models, for example, in the recognition of continuous handwriting without segmentation. In 2009, an artificial neural network model built using LSTM won the ICDAR (International Conference on Document Analysis and Recognition) handwriting recognition competition. LSTM is also widely used in autonomous speech recognition; as a nonlinear model, LSTM can be used as a complex nonlinear unit to construct larger deep neural networks.

[0053] FFmpeg (Fast Forward Moving Picture Experts Group) is an open-source computer program used to record, convert, and stream digital audio and video. It is licensed under the LGPL (Lesser General Public License). ， It is licensed under a Wide General Public License (GPL) or a General Public License (GPL). It provides a complete solution for recording, converting, and streaming audio and video. It includes highly advanced audio / video codec libraries; to ensure high portability and encoding / decoding quality, much of the code in the video decoding library was developed from scratch.

[0054] Wide&deep (Automatic High-Order Feature Combination) is a type of model for classification and regression released around June 2016, and it has been applied to app recommendations in the Google Play Store. The core idea of ​​the Wide&deep model is to combine the memorization ability of linear models with the generalization ability of DNN (Deep Neural Networks) models, optimizing the parameters of both models simultaneously during training to achieve optimal overall predictive ability.

[0055] This application can be applied to various video search error correction scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, and assisted driving.

[0056] Figure 1 This is a schematic diagram of an application scenario involved in an embodiment of this application, including a terminal device 101 and a server 102.

[0057] The terminal device 101 has interactive functionality, meaning it can receive operation commands input by an object (e.g., a user), such as receiving data detection operations input by the object. The terminal device 101 in this embodiment may include, but is not limited to, PCs (Personal Computers), PDAs (Tablet PCs), mobile phones, wearable smart devices, smart voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc. The terminal device 101 is often equipped with a display device, which may be a monitor, display screen, touchscreen, etc. The touchscreen may also be a touch screen, touch panel, etc. The display device can be used to display the final detection results, etc.

[0058] Server 102 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Server 102 can also be a node in a blockchain.

[0059] The terminal device 101 and the server 102 can be directly or indirectly connected through wired or wireless communication, and this application does not impose any restrictions on this.

[0060] In some embodiments, a data detection platform is installed and runs on the terminal device 101. An object can use this platform to detect target data. For example, the object can input a data detection operation on the platform, such as triggering a data detection control on the platform. When the data detection platform detects a data detection operation, it sends a data detection request to the server 102 through the terminal device 101. After receiving the detection request, the server 102 detects the target data and returns the detection result to the data detection platform. The data detection platform then displays the detection result of the target data to the object through the terminal device.

[0061] As mentioned above, current risk level assessments of internet information are primarily based on single text, image, or video review algorithms to obtain corresponding risk level results. However, the forms in which internet information is presented today are complex and diverse. Internet information is not just a piece of text, an image, or a video, but a combination of multiple elements. Current data detection methods cannot comprehensively assess internet information that includes multiple elements such as text, images, and videos, resulting in a low accuracy rate in assessing the risk level of internet information.

[0062] To address the aforementioned technical problems, this application embodiment, in response to data detection operations, acquires target data, which includes N elements, each of which can be any one of text, image, and video. First feature information corresponding to the N elements is determined, and based on this first feature information, risk levels corresponding to M combinations are determined. Each of the M combinations consists of any two sub-elements from the multiple sub-elements included in the N elements, and the sub-elements can be text or image. Based on the risk levels corresponding to the M combinations, the risk level of the target data is determined, and the risk level of the target data is sent to a display device for display. In other words, this application, in data detection, first comprehensively analyzes each element (e.g., text, image, video, etc.) included in the target data to obtain the first feature information corresponding to each of the N elements. Furthermore, to improve detection accuracy, the sub-elements included in the N elements are combined in pairs, and the risk level corresponding to each combination is analyzed, achieving fine-grained risk analysis. Then, based on the risk levels corresponding to the pairs of sub-elements, the risk level of the target data is accurately determined, thus improving the accuracy of data detection.

[0063] It should be noted that the data detection scheme provided in this application can be applied to any scenario requiring detection of target data, whereby the target data includes at least one element from text, images, or videos. This application does not limit the target data; for example, it can be advertising data, film and television data, news data, entertainment data, etc. For instance, when the target data is advertising, this application can also be applied to any stage of advertising placement, recommendation, or review.

[0064] The technical solutions of the embodiments of this application will be described in detail below through some examples. The following embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0065] Figure 2 This is a schematic flowchart illustrating a data detection method provided in an embodiment of this application. The method can be executed by an electronic device with data processing capabilities, such as a desktop computer, laptop computer, tablet computer, or other terminal device, and / or by one or more servers.

[0066] like Figure 2 As shown, the method includes the following steps:

[0067] S210: In response to a data detection operation, acquire the target data.

[0068] The target data consists of N elements, which can be any one of text, image, and video, where N is a positive integer greater than 1.

[0069] This application does not limit the specific type of target data in its embodiments.

[0070] For example, the target data in this application embodiment can be an advertisement.

[0071] Taking advertising as an example, with the development of advertising technology, an advertisement includes a wide variety of information and presentation formats. Currently, a single advertisement may consist of multiple elements, such as text and images, or text and video.

[0072] Based on this, in detecting target data, such as advertisements, the embodiments of this application comprehensively consider all elements included in the advertisement to achieve accurate detection. For example, during the advertisement review stage, all elements included in the advertisement are comprehensively detected to determine the risk level of the advertisement, and based on the risk level, it is determined whether to remove the advertisement or to continue displaying it.

[0073] When inspecting target data, the target data is first obtained in response to a data inspection operation. For example, an object (such as a user) triggers a data inspection operation, and the target data is obtained in response to that operation.

[0074] In some embodiments, S210 includes: acquiring target data in response to a data detection operation on the displayed event mode control.

[0075] For example, Figure 3 This application provides an example of an event mode diagram, such as... Figure 3 As shown in Figure A, a data detection platform is deployed on the terminal device. When an object triggers the data detection platform, the platform starts and redirects to... Figure 3 The interface shown in B. Figure 3 As shown in Figure B, this interface displays an event-mode control. When an object (such as a user) triggers the displayed event-mode control, and the data detection platform detects that the displayed event-mode control has been triggered, in some embodiments, the data detection platform sends a data detection request to other electronic devices through a terminal device. The other electronic devices respond to the data detection request, acquire the target data, and perform detection on the target data.

[0076] In some embodiments, the target data described above is object input.

[0077] In some embodiments, the target data described above is obtained from a target data storage device.

[0078] For example, the target data storage could be HDFS (Hadoop Distributed File System), which has a large storage capacity and contains complete target data information, such as different advertising data.

[0079] In some embodiments, the event-mode control includes at least one of a pre-run mode control, an experimental mode control, a cleanup mode control, and a formal mode control. For example Figure 3 As shown in Figure B, the event-mode controls displayed by the data monitoring platform include pre-run mode controls, experimental mode controls, cleanup mode controls, and production mode controls. Optionally, the target data corresponding to different event-mode controls may not be exactly the same.

[0080] For example, the pre-run mode is used to verify whether the changed review algorithm hits the expected target data; the experimental mode is used to compare the differences brought about by using two different review algorithms for risk level assessment; the cleanup mode is used to apply the new review algorithm to historical target data, use the new review algorithm on the target data displayed on the media traffic side, and remove the non-compliant target data; the official mode is to take effect on incremental target data.

[0081] Taking target data as an example of advertising, users log in to the data detection system to process the advertisements on the media platform and select one of four event modes—pre-run, experiment, cleanup, and formal—based on the business scenario they are facing.

[0082] In pre-run mode, users input advertising data with known risk levels. Pre-trained text semantic analysis, image recognition, video detection, and risk prediction models estimate the risk level of this advertising data. This estimate is then compared to the actual risk level of the advertising data to confirm the feasibility of changing the review algorithm. If the review algorithm is not feasible, the aforementioned text semantic analysis, image recognition, video detection, and risk prediction models are retrained, and the review algorithm is adjusted. If the review algorithm accurately reflects the expected risk level of the advertising data, it can be applied to the review of both existing and new advertising.

[0083] In the experimental mode, users input advertising data into electronic devices applying two different review algorithms to observe the differences in the risk level predictions for the advertising data under the two different review algorithms. Both review algorithms include text semantic analysis models, image recognition models, video detection models, and risk prediction models, but the text semantic analysis models, image recognition models, video detection models, and risk prediction models included in different review algorithms are not entirely the same.

[0084] In cleanup mode, the electronic device executes a review algorithm that has passed feasibility verification. The user inputs historical advertising data into the electronic device, and the risk level of the historical advertising data is assessed through text semantic analysis model, image recognition model, video detection model, and risk prediction model.

[0085] In formal mode, the electronic device executes a review algorithm that has passed feasibility verification. The user inputs incremental advertising data into the electronic device, and the risk level of the incremental advertising data is evaluated through text semantic analysis model, image recognition model, video detection model, and risk prediction model.

[0086] Based on the above-mentioned technical solution of setting four event modes, this application can effectively verify the accuracy of the audit algorithm before the audit algorithm works through the pre-run mode. At the same time, through the cleanup mode and the formal mode, the risk level of the full amount of target data (such as the full amount of advertising data, including historical existing advertising data and incremental advertising data) can be estimated, thereby realizing the cleanup of the target data.

[0087] S220: Determine the first feature information corresponding to the N elements.

[0088] This application embodiment does not limit the specific method of determining the first feature information corresponding to the N elements in the above S220.

[0089] In some embodiments, for each of the N elements, the first feature information of that element is extracted using the neural network model corresponding to that element. For example, if the element is text, the text is input into textCNN for feature extraction to obtain the first feature information of the text.

[0090] In some embodiments, determining the first feature information corresponding to the N elements in S220 above includes the following steps S220-A and S220-B:

[0091] S220-A: Determine the second feature information corresponding to N elements;

[0092] In this embodiment, the second feature information corresponding to N elements is determined in order to transform the N elements into the input form of a preset model, such as transforming text into a multi-dimensional vector or video into a three-dimensional matrix of [X,Y,Z].

[0093] In this embodiment, in order to improve the speed of determining the first feature information and reduce the computational load of the model, before determining the first feature information of the elements, the second feature information corresponding to N elements is first determined, and the second feature information corresponding to the N elements is directly input into the model for first feature extraction. The model does not need to process the input data, thereby improving the computational speed.

[0094] The methods for determining the second feature information corresponding to the N elements in S220-A above include, but are not limited to, the following:

[0095] Method 1: Obtain the second feature information from the feature memory, i.e., the above S220-A includes the following steps S220-A1 and S220-A2:

[0096] S220-A1: Retrieve the identification information corresponding to N elements from the ID memory;

[0097] S220-A2: Based on the identification information corresponding to each of the N elements, obtain the second feature information corresponding to the N elements from the feature memory.

[0098] Since the calculation process for obtaining the second feature information corresponding to the N elements involves a huge amount of computation, a feature memory is used to associate the second feature information with the identification information corresponding to the N elements in the ID memory. The second feature information is obtained based on the identification information, which greatly reduces the amount of computation in the extraction process of the second feature information and saves detection time.

[0099] This application does not limit the specific form of the identification information.

[0100] In one example, the identification information is name information, such as text name, image name, video name, advertisement name, and material name.

[0101] In another example, the identification information is ID (Identity) information, such as text ID, image ID, video ID, advertisement ID, and material ID.

[0102] To improve the response speed of storage by more than 100 times when dealing with large amounts of target data, only ID information is recorded for complex and diverse target data, and ID association is provided. At this time, the data storage space is only about one percent of the original, which greatly saves data storage space.

[0103] For example, the identification information of an element is its id information.

[0104] For example, the ID storage described above is used to store ID information related to recorded materials. Taking advertising as an example, it can store ad IDs, material IDs, element IDs, etc. Because of the need to process large amounts of data (such as advertising), in order to improve the storage response speed by n times, for example, more than 100 times, only ID information is recorded for complex and diverse target data, i.e., providing ID association relationships. In this case, the data storage space is only 1 / n of the original, for example, about 1 / 100.

[0105] For example, the feature memory stores the correspondence between the identification information and the second characteristic information of different elements.

[0106] In this way, when determining the second characteristic information of an element, the element's identification information can be obtained from the ID memory, and the element's second characteristic information can be retrieved from the feature memory based on this identification information. The whole process is simple and computationally inefficient, thus improving the efficiency of data detection.

[0107] Method 2: For each of the N elements, if the second feature information corresponding to the element is not found in the feature memory, then feature extraction is performed on the element to obtain the second feature information of the element.

[0108] S220-B: For each of the N elements, the second feature information of the element is processed through the preset model corresponding to that element to obtain the first feature information of that element.

[0109] In this application embodiment, different elements correspond to different preset models. In this way, when determining the first feature information of each element, the preset model that matches the element is used for extraction, which can improve the accuracy of extraction.

[0110] It should be noted that the preset models corresponding to the above elements are pre-trained models.

[0111] For example, the preset model corresponding to the above text can be a text semantic analysis model.

[0112] This application does not restrict the specific network structure of the text semantic analysis model; it can be any network structure with corresponding feature extraction or extraction functions.

[0113] In one example, the text semantic analysis model is an LSTM model.

[0114] In another example, the text semantic analysis model is the TextCNN model (Text Convolutional Neural Networks).

[0115] In another example, the text semantic analysis model mentioned above includes LSTM (Long Short-Term Memory), an attention layer, and a classifier (softmax). The text semantic analysis model corresponding to the text element processes the second feature information of the text element to obtain the first feature information of the text element.

[0116] Figure 4 This is a schematic diagram of the network structure of a text semantic analysis model provided in an embodiment of this application.

[0117] like Figure 4 As shown, the second feature information of the text element is processed by LSTM to obtain the hidden feature information and output feature information corresponding to each character in the text element.

[0118] For each character in the text element, the hidden feature information and output feature information of the character are processed through the attention layer to obtain the third feature information of the text element.

[0119] The third feature information of the text element is processed by a classifier to obtain the first feature information of the text element.

[0120] For example, for the promotional text "Group buying activities are underway in XXX region, with a continuous stream of great deals," an LSTM+Attention (attention layer) approach is used for risk prediction. A word segmenter is used to segment the sentence, extracting key words: "group buying," "great deals," and "group buying." These three words are input into the representation layer for vectorization and weighting. The weighted vectors are then iteratively trained through an LSTM network to obtain the hidden feature information h1, h2, h3 and the output feature information h1', h2', h3' for the three words "group buying," "great deals," and "group buying." The attention layer processes the hidden and output feature information of the text to obtain the third feature information. A classifier processes this third feature information to output a 512-dimensional feature vector. To facilitate subsequent calculations, this feature vector is stored in a feature memory, avoiding multiple time-consuming repetitive calculations.

[0121] This approach abandons the traditional multi-classification method using the sigmoid activation function and loss function. Instead, it utilizes a softmax classifier to treat the first feature information output by the attention layer as a vector. This allows for integrated use with the processing results of images and videos to determine the risk level in the risk prediction model.

[0122] For example, the preset model corresponding to the above image can be an image recognition model.

[0123] This application does not impose any restrictions on the specific network structure of the image recognition model.

[0124] In one example, the image recognition model includes OCR (Optical Character Recognition) and DNN (Deep Neural Networks) models.

[0125] In another example, the image recognition model mentioned above includes three convolutional layers and a classifier (softmax). The second feature information of the image element is processed by the image recognition model corresponding to the image element to obtain the first feature information of the image element.

[0126] Figure 5 This is a schematic diagram of the network structure of an image recognition model provided in an embodiment of this application.

[0127] like Figure 5 As shown, for the input image element, a modified CNN (Convolutional Neural Networks) is used to obtain the vector result of the image element.

[0128] For example, when the image element is input into the network structure of the image recognition model, the Convolution 1 (Conv1) in the network structure performs vectorization weighting on the image element to obtain a 2048-dimensional vector result value. After two more convolution operations, a smaller 512-dimensional feature vector is obtained. Finally, it is input into the softmax classifier for feature vector recording and classification operations.

[0129] For convolutional methods, common methods such as VGG-16 (Visual Geometry Group Network-16), VGG-19 (Visual Geometry Group Network-19), and ResNet-101 (Residual Neural Network-101) can be used. In terms of feasibility, the convolutional output layer of the ResNet-101 network has more than 1000 dimensions, which, although it has higher accuracy, makes it difficult to obtain the same 512-dimensional result as the text elements mentioned above. Therefore, the VGG-19 + TripletNet (Triplet Network) scheme is used to vectorize the image. Similarly, since the vectorization process takes a long time, the vectorized result is stored in the feature memory (layer memory) for easy reuse later.

[0130] For example, the preset model corresponding to the video is a video detection model.

[0131] This application does not impose restrictions on the specific network structure of the video detection model.

[0132] In one example, the video detection model may include the FFmpeg algorithm and the autofilp (automatic flipping) algorithm.

[0133] It should be noted that the numerical parameters in the above model are only an example. The embodiments of this application include, but are not limited to, the above numerical parameters. For example, they are not limited to 512-dimensional feature vectors, but can also be feature vectors of other different dimensions.

[0134] Based on the text semantic analysis model, image recognition model, and video detection model set above, for target data containing at least one element from text, image, and video, after the action of the above three algorithm models, different elements such as text, image, and video can be expressed in the form of feature vectors of the same dimension. The risk prediction model processes the first feature information of different elements to accurately obtain the comprehensive risk level of target data containing elements such as text, image, and video.

[0135] In some embodiments, the method for determining the first feature information corresponding to the N elements in S220 above includes the following steps S220-C and S220-D:

[0136] S220-C: Based on the relationship between the N elements, generate at least one DAG with each of the N elements as a vertex, and the DAG includes at least one of the N elements;

[0137] S220-D: For each DAG in at least one DAG, send the DAG as a subtask to the target computing device so that the target computing device can determine the first characteristic information corresponding to each element in the DAG. The target computing device is any computing device in the distributed system.

[0138] In this embodiment, in order to improve the speed of determining the first feature information, the first feature information of the elements is extracted through a distributed system. Specifically, the target data is first decomposed to obtain N elements. Then, based on the relationship between the N elements, at least one DAG is generated with each of the N elements as a vertex. A DAG is sent as a subtask to different target computing devices in the distributed system for parallel computing to quickly obtain the first feature information of each element.

[0139] Figure 6 This is a schematic diagram of a Directed Acyclic Graph (DAG) provided in this embodiment. For example, based on the parallelism and dependencies of the task computation, at least one DAG is generated using each of the N elements as vertices. Figure 6 As shown: There is a dependency between text 1 and text 2 in Task 1. Text 1 must be calculated first, and then text 2 can be calculated. Therefore, the execution order indicated by DAG is to calculate text 1 and image 1 simultaneously, and then calculate text 2. Task 2 calculates text 1 and image 1 simultaneously.

[0140] In this embodiment of the application, before determining the first feature information of each of the N elements included in the target data, the second feature information of each element is first determined, and then the second feature information of the element is input into the preset model corresponding to the element to obtain the first feature information of the element. This not only improves the determination speed of the first feature information, but also improves the determination accuracy. Thus, when performing subsequent detection based on the accurate first feature information, the detection accuracy can be improved.

[0141] After determining the first feature information corresponding to each of the N elements according to the above method, the following S230 is executed.

[0142] S230: Based on the first feature information corresponding to the N elements, determine the risk level corresponding to the M combinations.

[0143] Each of the above M combinations consists of any two sub-elements from the multiple sub-elements included in the N elements. The sub-elements are text or images, and M is a positive integer greater than 1.

[0144] For example, suppose the target data includes text 1 and video 1, where video 1 consists of 3 frames. Treating each frame as an image, video 1 comprises 3 images. Therefore, the target data includes text 1 and 3 images, for a total of 4 sub-elements. Combining any two of these 4 sub-elements creates multiple combinations. For instance, combining the 4 sub-elements in pairs yields 6 combinations, such as (text 1, image 1), (text 1, image 2), (text 1, image 3), (image 1, image 2), (image 1, image 3), and (image 2, image 3).

[0145] In order to improve the accuracy of risk level determination, this application embodiment combines the sub-elements included in the N elements in pairs to obtain M combinations, and determines the risk level corresponding to each of the M combinations. Then, based on the risk level corresponding to each of the M combinations, the risk level of the target data is accurately determined.

[0146] This application embodiment does not limit the specific method of determining the risk level corresponding to the M combinations based on the first feature information corresponding to the N elements in the above S230.

[0147] In some embodiments, taking one of the M combinations as an example, the first feature information corresponding to the sub-elements included in the combination is obtained from the first feature information corresponding to the N elements. Referring to the example above, assuming that the combination includes (text 1, image 1), the first feature information of image 1 is extracted from the first feature information corresponding to video 1. Then, based on the first feature information of text 1 and the first feature information of image 1, the risk level corresponding to the combination is determined. For example, the first feature information of text 1 and the first feature information of image 1 are input into a pre-trained risk prediction model to predict the risk level corresponding to the combination.

[0148] In some embodiments, the risk prediction model described above is an LTR (Learning to Rank, a supervised learning ranking algorithm) model.

[0149] Figure 7 This is a schematic diagram of an LTR principle involved in an embodiment of this application.

[0150] The entire working principle of LTR is as follows Figure 7As shown, the combination of text and image is considered a form of material combination, where the learning system is used to generate the risk level of the image & text combination. For example, Figure 7 As shown, m texts and n images are combined. For example, text 1 is combined with n images, resulting in combinations denoted as text 1, image 1, text 1, image 2, and so on up to text 1, image n. Similarly, text m is combined with n images, resulting in combinations denoted as text m, image 1, text m, image 2, and so on up to text m, image n. The learning system is trained using these combinations to predict the risk level corresponding to different image and text combinations. In practical use, it is necessary to calculate the risk level of each of the n combinations of text m+1 and n images. The ranking system loads the trained learning system, which predicts the risk levels f(text m+1, image 1), f(text m+1, image 2), and so on up to text m+1, image n for the n combinations of text m+1 and n images, i.e., text m+1, image 1, text m+1, image 2, and so on up to text m+1, image n.

[0151] Thus, when determining the risk level corresponding to M combinations, these M combinations can be input into the trained LTR model to predict the risk level corresponding to the M combinations.

[0152] In some embodiments, determining the risk level corresponding to the M combinations based on the first feature information corresponding to the N elements in S230 above includes the following steps S230-A and S230-B:

[0153] S230-A: For each of the M combinations, determine the identification information corresponding to each element in that combination;

[0154] S230-B: Determine the risk level of the combination based on the identification information corresponding to each element in the combination and the first characteristic information corresponding to each element in the combination.

[0155] In this approach, to improve the accuracy of determining the risk level of a portfolio, the identification information of each element in the portfolio is considered during the process of determining the risk level of the portfolio.

[0156] For example, the identification information corresponding to each element in the combination, as well as the first feature information corresponding to each element in the combination, are input into a pre-trained risk prediction model to predict the risk level corresponding to the combination.

[0157] In some embodiments, S230-B above includes the following steps S230-B1 and S230-B2:

[0158] S230-B1: Perform K nonlinear activation processes on the first feature information corresponding to each element in the combination to obtain the feature vector corresponding to the combination, where K is a positive integer;

[0159] S230-B2: Classify the risk level of the combination based on the identification information and feature vectors corresponding to each element in the combination.

[0160] In some embodiments, S230-B above determines the risk level corresponding to the above combination using a wide & deep model.

[0161] Figure 8 This is a schematic diagram illustrating the principle of the wide & deep model involved in the embodiments of this application.

[0162] like Figure 8 As shown, the wide&deep model is divided into a wide side and a deep side. The input to the wide side is the identification information corresponding to each element in the combination, such as text ID, image ID, video ID, advertisement ID, and material ID. The input to the deep side is the first feature information corresponding to each element in the combination. The deep side first performs full-length feature concatenation on the first feature information corresponding to each element in the combination, and then performs two ReLU (Rectified Linear Activation Function) processes to obtain the feature vector corresponding to the combination. Based on the identification information and feature vector corresponding to each element in the combination, classification is performed, for example, after softmax classification, to obtain the risk level corresponding to the combination.

[0163] S240: Based on the risk levels corresponding to the M combinations, determine the risk level of the target data and send the risk level of the target data to the display device for display.

[0164] This application does not limit the specific method for determining the risk level of target data based on the risk levels corresponding to M combinations.

[0165] In one example, S240 above includes the following steps S240-A and S240-B:

[0166] S240-A: Based on the risk levels corresponding to M combinations, determine the probability that the risk level of the i-th combination is greater than the risk level of the j-th combination. The i-th combination and the j-th combination are any combination among the M combinations, and the i-th combination is different from the j-th combination.

[0167] S240-B: Determine the risk level of the target data based on the probability that the risk level of the i-th combination is greater than the risk level of the j-th combination.

[0168] For example, suppose there are four combinations. After step S240-A, the probability that combination 1 has a higher risk level than combination 2 is probability 1, the probability that combination 2 has a higher risk level than combination 3 is probability 2, and the probability that combination 3 has a higher risk level than combination 4 is probability 3. The combinations are then sorted according to their probability values. Combinations 2 and 3 are considered safe. If either probability 1 or probability 3 is greater than probability 2, the target data is considered risky. If both probability 1 and probability 3 are less than probability 2, the target data is considered safe.

[0169] In another example, S240 above includes the following steps S240-C and S240-D:

[0170] S240-C: Determine the product of the risk levels corresponding to the M combinations;

[0171] S240-D: Determine the risk level of the target data based on the product.

[0172] The specific process of determining the risk level of the target data in S240-D above includes: if the product of the risk levels corresponding to the M combinations is higher than a set threshold, it means that there is a combination with a high risk coefficient among the M combinations, and the risk level of the target data is high; if the product is lower than the set threshold, it means that the M combinations are relatively safe, and the risk level of the target data is low.

[0173] In some embodiments, when training the LTR model described above, the problem of machine learning using the LTR ranking method can be divided into the following three types: the risk level of text & image, the risk level of (text 1 & image 1) > (text 1 & image 2), and the risk level of text 1 combined with n images. These are referred to below as point-wise (single document method), pair-wise (document pair method), and list-wise (document list method). It is assumed that the risk rating generated after a text is matched with multiple image inputs is the true risk of the target data.

[0174] Figure 9 This is a schematic diagram of three risk level prediction methods for LTR involved in the embodiments of this application.

[0175] like Figure 8 As shown, LTR working methods are mainly divided into three categories:

[0176] 1) Point-wise LTR prediction models. These models take as input features of a single item and request, and output a relevance score, ranking the items based on that score. Point-wise models do not consider the interactions between elements of the target data; therefore, they cannot guarantee that the sequence is optimal.

[0177] In this embodiment of the application, the risk levels of M combinations can be processed based on a point-wise LTR prediction model to obtain the risk level correlation scores of the M combinations, and then the risk level of the target data can be determined based on the correlation scores.

[0178] 2) Pair-wise LTR prediction models. These models predict the relative positions of item pairs using a loss function. If the relative positions are correct, it's a positive example; otherwise, it's a negative example. The goal of the model is to reduce the number of incorrect item pairs. Pair-wise models ignore global information about the list and significantly increase the complexity of model training and prediction.

[0179] In this embodiment of the application, the risk levels of M combinations can be processed based on a pair-wise LTR prediction model to obtain the probability that the risk level of the i-th combination is greater than the risk level of the j-th combination. The i-th combination and the j-th combination are any combination among the M combinations, and the i-th combination is different from the j-th combination. The risk level of the target data is determined based on this probability.

[0180] 3) List-wise LTR prediction model. This type of model models the entire item list under each request, considers the relationships between all samples within it, and uses a list-wise loss function to compare the relationships between different items.

[0181] In this embodiment of the application, the risk level of M combinations is processed by the list-wise LTR prediction model to obtain the product of the risk levels of the M combinations. If the product is higher than a set threshold, it means that there is a combination with a high risk coefficient among the M combinations, and the risk level of the target data is high. If the product is lower than the set threshold, it means that the M combinations are relatively safe and the risk level of the target data is low.

[0182] In some embodiments, the above method can also be executed by a distributed system, which includes multiple computing devices.

[0183] Taking the target data as advertising data as an example, the advertising data includes N elements, which include any one of text, image and video, where N is a positive integer greater than 1.

[0184] Figure 10 The diagram shown is a schematic representation of the execution process of a distributed system according to an embodiment of this application. Figure 10 As shown, the system includes a terminal device, an ID storage device, and a distributed system. The distributed system includes a first computing device, a second computing device, a third computing device, a fourth computing device, and a fifth computing device. The second, third, and fourth computing devices form a computing module, and the fifth computing device forms an algorithm module.

[0185] Users select an event mode on their terminal device, which includes pre-run mode, experimental mode, cleanup mode, and formal mode. The terminal device sends a data detection request to the first computing device in the distributed system. The first computing device responds to the data detection request, obtains the advertising data, reads the corresponding ID information (including advertising ID, text ID, image ID, video ID, and material ID, etc.) from the ID storage, and sends the advertising data and ID information to the second computing device in the distributed system.

[0186] The second computing device receives the advertising data and ID information sent by the first computing device, and decomposes the advertising data to obtain N elements. Then, based on the relationship between the N elements, the second computing device generates at least one DAG with each of the N elements as a vertex, and sends the DAG and ID information to the third computing device in the distributed system.

[0187] The third computing device receives the DAG and ID information sent by the second computing device, obtains the OP operator corresponding to the DAG, and sends the OP operator and ID information to the fourth computing device in the distributed system.

[0188] The fourth computing device receives the OP operator and ID information sent by the third computing device, and obtains the subtask to be executed based on the OP operator. According to the ID information corresponding to each element in the subtask, it retrieves the second feature information corresponding to each element from the layer memory. If the second feature information of an element is not found in the layer memory, feature extraction is performed on the element to obtain the second feature information of the element, and the second feature information corresponding to N elements is sent to the fifth computing device in the distributed system.

[0189] The fifth computing device receives the second feature information of N elements sent by the fourth computing device. Based on the category of the N elements, it inputs them into the corresponding text semantic analysis model, image recognition model, and video detection model. Through these three models, it obtains the first feature information corresponding to the N elements. Then, based on the first feature information of the N elements, the fifth computing device, according to steps S230 and S240, obtains the risk levels corresponding to M combinations. Each of the M combinations consists of any two sub-elements from the multiple sub-elements included in the N elements. These sub-elements can be text or images, and M is a positive integer greater than 1. Based on the risk levels of the M combinations, the fifth computing device determines the risk level of the advertising data and sends the execution results to the terminal device for collection and display.

[0190] In pre-run mode, the fifth computing device sends the result of whether the review algorithm hits the expected risk level of the advertising data to the terminal device for display; in experimental mode, the fifth computing device sends the risk level prediction result of the advertising data to the terminal device for display when using two different review algorithms; in cleanup mode, the fifth computing device sends the cleanup result of historical existing advertisements to the terminal device for display; in formal mode, the fifth computing device sends the cleanup result of incremental advertisements to the terminal device for display.

[0191] The data detection method provided in this application, in response to a data detection operation, acquires target data, which includes N elements, each of which can be any one of text, image, and video. It determines first feature information corresponding to the N elements and, based on the first feature information, determines the risk level corresponding to M combinations. Each of the M combinations consists of any two sub-elements from the multiple sub-elements included in the N elements, and the sub-elements can be text or image. Based on the risk level corresponding to the M combinations, it determines the risk level of the target data and sends the risk level of the target data to a display device for display. In other words, in this application embodiment, during data detection, it first comprehensively analyzes each element (e.g., text, image, video, etc.) included in the target data to obtain the first feature information corresponding to each of the N elements. Furthermore, to improve detection accuracy, it combines the sub-elements included in the N elements in pairs, analyzes the risk level corresponding to each combination, and achieves fine-grained risk analysis. Then, based on the risk levels corresponding to the pairs of sub-elements, it accurately determines the risk level of the target data, thereby improving the accuracy of data detection.

[0192] The preferred embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this application, various simple modifications can be made to the technical solutions of this application, and these simple modifications all fall within the protection scope of this application. For example, the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, this application will not describe the various possible combinations separately. Furthermore, various different embodiments of this application can also be arbitrarily combined, as long as they do not violate the spirit of this application, they should also be considered as the content disclosed in this application.

[0193] It should also be understood that, in the various method embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0194] Figure 11 A schematic diagram of the structure of a data detection device provided in one embodiment of this application. (See attached diagram.) Figure 9 As shown, the data detection device 10 includes:

[0195] Acquisition unit 11 is used to acquire target data in response to a data detection operation. The target data includes N elements, and the elements include any one of text, image, and video, where N is a positive integer greater than 1.

[0196] The first determining unit 12 is used to determine the first feature information corresponding to the N elements, and to determine the risk level corresponding to the M combinations based on the first feature information corresponding to the N elements. Each of the M combinations is a combination composed of any two sub-elements among the multiple sub-elements included in the N elements. The sub-elements are text or images, and M is a positive integer greater than 1.

[0197] The second determining unit 13 is used to determine the risk level of the target data according to the risk levels corresponding to the M combinations, and send the risk level of the target data to the display device for display.

[0198] In some embodiments, the first determining unit 12 is specifically used to determine the identification information corresponding to each element in each of the M combinations; and to determine the risk level corresponding to the combination based on the identification information corresponding to each element in the combination and the first feature information corresponding to each element in the combination.

[0199] In some embodiments, the first determining unit 12 is further configured to perform K nonlinear activation processing on the first feature information corresponding to each element in the combination to obtain the feature vector corresponding to the combination, where K is a positive integer; and to classify the risk level corresponding to the combination according to the identification information corresponding to each element in the combination and the feature vector.

[0200] In some embodiments, the second determining unit 13 is specifically used to determine the probability that the risk level of the i-th combination is greater than the risk level of the j-th combination based on the risk levels corresponding to the M combinations, wherein the i-th combination and the j-th combination are any combination among the M combinations, and the i-th combination is different from the j-th combination; and to determine the risk level of the target data based on the probability that the risk level of the i-th combination is greater than the risk level of the j-th combination.

[0201] In some embodiments, the second determining unit 13 is specifically used to determine the product of the risk levels corresponding to the M combinations; and to determine the risk level of the target data based on the product.

[0202] In some embodiments, the first determining unit 12 is further configured to determine the second feature information corresponding to the N elements; and for each of the N elements, process the second feature information of the element through a preset model corresponding to the element to obtain the first feature information of the element.

[0203] In some embodiments, the first determining unit 12 is further configured to obtain the identification information corresponding to the N elements from the ID memory; and obtain the second feature information corresponding to the N elements from the feature memory based on the identification information corresponding to the N elements.

[0204] In some embodiments, the first determining unit 12 is further configured to perform feature extraction on each of the N elements if the second feature information corresponding to the element is not found in the feature memory, and obtain the second feature information of the element.

[0205] In some embodiments, the first determining unit 12 is specifically configured to, if the element is text and the preset model includes LSTM, an attention layer, and a classifier, process the second feature information of the element through the preset model corresponding to the element to obtain the first feature information of the element, including: processing the second feature information of the text through the LSTM to obtain hidden feature information and output feature information corresponding to each character in the text; for each character in the text, processing the hidden feature information and output feature information of the character through the attention layer to obtain the third feature information of the text; and processing the third feature information of the text through the classifier to obtain the first feature information of the text.

[0206] In some embodiments, the acquisition unit 11 is specifically used to acquire the target data in response to a data detection operation on the displayed event mode control.

[0207] In some embodiments, the event mode control that the acquisition unit 11 specifically responds to includes at least one of a pre-run mode control, an experimental mode control, a cleanup mode control, and a formal mode control.

[0208] In some embodiments, the first determining unit 12 is further configured to generate at least one DAG based on the relationship between the N elements, with each of the N elements as a vertex, wherein the DAG includes at least one element among the N elements; and for each of the at least one DAG, send the DAG as a subtask to a target computing device so that the target computing device determines the first characteristic information corresponding to each element in the DAG, wherein the target computing device is any computing device in the distributed system.

[0209] It should be understood that the device embodiments and method embodiments can correspond to each other, and similar descriptions can be referred to the method embodiments. To avoid repetition, further details will not be provided here. Specifically, Figure 10 The apparatus shown can execute the above-described model training method embodiments, and the foregoing and other operations and / or functions of each module in the apparatus are respectively for implementing the model training method embodiments corresponding to the computing device. For the sake of brevity, they will not be described in detail here.

[0210] The apparatus of this application embodiment has been described above from the perspective of functional modules in conjunction with the accompanying drawings. It should be understood that this functional module can be implemented in hardware, in software instructions, or in a combination of hardware and software modules. Specifically, the steps of the method embodiments in this application can be completed by integrated logic circuits in the processor's hardware and / or by software instructions. The steps of the method disclosed in this application embodiment can be directly embodied as being executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. Optionally, the software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps in the above method embodiments.

[0211] Figure 12 This is a schematic block diagram of an electronic device provided in an embodiment of this application. The electronic device can be used to perform the above-described data detection method.

[0212] like Figure 12 As shown, the electronic device 60 may include:

[0213] The system includes a memory 61 and a processor 62. The memory 61 stores computer programs and transfers the program code to the processor 62. In other words, the processor 62 can retrieve and run the computer programs from the memory 61 to implement the methods described in the embodiments of this application.

[0214] For example, the processor 62 can be used to execute the above-described method embodiments according to instructions in the computer program.

[0215] In some embodiments of this application, the processor 62 may include, but is not limited to:

[0216] General-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0217] In some embodiments of this application, the memory 61 includes, but is not limited to:

[0218] Volatile memory and / or non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0219] In some embodiments of this application, the computer program may be divided into one or more modules, which are stored in the memory 61 and executed by the processor 62 to perform the method provided in this application. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the video production device.

[0220] like Figure 12 As shown, the electronic device 60 may further include:

[0221] Transceiver 63, which can be connected to processor 62 or memory 61.

[0222] The processor 62 can control the transceiver 63 to communicate with other devices; specifically, it can send information or data to other devices or receive information or data sent by other devices. The transceiver 63 may include a transmitter and a receiver. The transceiver 63 may further include antennas, and the number of antennas may be one or more.

[0223] It should be understood that the various components in the device are connected through a bus system, which includes a data bus, a power bus, a control bus, and a status signal bus.

[0224] According to one aspect of this application, a computer storage medium is provided that stores a computer program thereon, which, when executed by a computer, enables the computer to perform the methods of the above-described method embodiments. Alternatively, embodiments of this application also provide a computer program product containing instructions that, when executed by a computer, cause the computer to perform the methods of the above-described method embodiments.

[0225] According to another aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computing device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computing device to perform the method of the above-described method embodiments.

[0226] In other words, when implemented using software, it can be implemented wholly or partially in the form of a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video disc (DVD)), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0227] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0228] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0229] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. For example, the functional modules in the various embodiments of this application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0230] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data detection method, characterized in that, include: In response to a data detection operation, target data is acquired, the target data comprising N elements, wherein the elements include any one of text, image and video, and N is a positive integer greater than 1; Determine the first feature information corresponding to the N elements, and determine the risk level corresponding to the M combinations based on the first feature information corresponding to the N elements. Each of the M combinations consists of any two sub-elements from the multiple sub-elements included in the N elements. The sub-elements are text or images, and M is a positive integer greater than 1. Based on the risk levels corresponding to the M combinations, the risk level of the target data is determined, and the risk level of the target data is sent to the display device for display.

2. The method according to claim 1, characterized in that, The step of determining the risk level corresponding to the M combinations based on the first feature information corresponding to the N elements includes: For each of the M combinations, determine the identification information corresponding to each sub-element in the combination; The risk level of the combination is determined based on the identification information corresponding to each sub-element in the combination and the first feature information corresponding to each sub-element in the combination.

3. The method according to claim 2, characterized in that, The step of determining the risk level corresponding to the combination based on the identification information corresponding to each element in the combination and the first feature information corresponding to each sub-element in the combination includes: The first feature information corresponding to each sub-element in the combination is subjected to K nonlinear activation processes to obtain the feature vector corresponding to the combination, where K is a positive integer; The risk level corresponding to the combination is obtained by classifying the combination based on the identification information corresponding to each sub-element in the combination and the feature vector.

4. The method according to any one of claims 1-3, characterized in that, The step of determining the risk level of the target data based on the risk levels corresponding to the M combinations includes: Based on the risk levels corresponding to the M combinations, determine the probability that the risk level of the i-th combination is greater than that of the j-th combination, wherein the i-th combination and the j-th combination are any combination among the M combinations, and the i-th combination is different from the j-th combination; The risk level of the target data is determined based on the probability that the risk level of the i-th combination is greater than the risk level of the j-th combination.

5. The method according to any one of claims 1-3, characterized in that, The step of determining the risk level of the target data based on the risk levels corresponding to the M combinations includes: Determine the product of the risk levels corresponding to the M combinations; The risk level of the target data is determined based on the product.

6. The method according to any one of claims 1-3, characterized in that, Determining the first feature information corresponding to the N elements includes: Determine the second feature information corresponding to the N elements; For each of the N elements, the second feature information of the element is processed through the preset model corresponding to the element to obtain the first feature information of the element.

7. The method according to claim 6, characterized in that, Determining the second feature information corresponding to the N elements includes: Retrieve the identification information corresponding to each of the N elements from the ID memory; Based on the identification information corresponding to the N elements, the second feature information corresponding to the N elements is obtained from the feature memory.

8. The method according to claim 7, characterized in that, The method further includes: For each of the N elements, if the second feature information corresponding to the element is not found in the feature memory, then feature extraction is performed on the element to obtain the second feature information of the element.

9. The method according to claim 6, characterized in that, If the element is text, and the preset model includes a Long Short-Term Memory (LSTM) network, an attention layer, and a classifier, then the second feature information of the element is processed through the preset model corresponding to the element to obtain the first feature information of the element, including: The second feature information of the text is processed by the LSTM to obtain the hidden feature information and output feature information corresponding to each character in the text; For each character in the text, the attention layer processes the hidden feature information and output feature information of the character to obtain the third feature information of the text; The classifier processes the third feature information of the text to obtain the first feature information of the text.

10. The method according to any one of claims 1-3, characterized in that, The step of acquiring target data in response to a data detection operation includes: In response to a data detection operation on the displayed event-mode control, the target data is acquired.

11. The method according to claim 10, characterized in that, The event mode control includes at least one of the following: pre-run mode control, experimental mode control, cleanup mode control, and formal mode control.

12. The method according to any one of claims 1-3, characterized in that, Determining the first feature information corresponding to the N elements includes: Based on the relationship between the N elements, at least one directed acyclic graph (DAG) is generated using each of the N elements as a vertex. The DAG includes at least one of the N elements. For each of the at least one DAGs, the DAG is sent as a subtask to a target computing device so that the target computing device can determine the first characteristic information corresponding to each element in the DAG. The target computing device can be any computing device in the distributed system.

13. A data detection device, characterized in that, include: An acquisition unit is configured to acquire target data in response to a data detection operation. The target data includes N elements, where each element includes any one of text, image, and video, and N is a positive integer greater than 1. The first determining unit is used to determine the first feature information corresponding to the N elements, and to determine the risk level corresponding to the M combinations based on the first feature information corresponding to the N elements. Each of the M combinations is a combination composed of any two sub-elements from the plurality of sub-elements included in the N elements. The sub-elements are text or images, and M is a positive integer greater than 1. The second determining unit is used to determine the risk level of the target data according to the risk levels corresponding to the M combinations, and send the risk level of the target data to the display device for display.

14. An electronic device, characterized in that, include: Memory, processor; The memory is used to store computer programs; The processor is configured to execute the computer program to implement the method as described in any one of claims 1 to 12.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 12.

Images