Live broadcast content auditing method and device, computing equipment and computer storage medium

Abstract

The embodiment of the invention discloses a live broadcast content auditing method and device, computing equipment and a computer storage medium, and the method comprises the steps: collecting live broadcast related data of an anchor end through an edge auditing SDK integrated in the anchor end, and auditing the live broadcast related data to obtain an edge auditing result; judging whether to trigger a cloud auditing mechanism or not according to the edge auditing result; and if yes, pushing the live broadcast related data to the cloud, and performing multi-modal analysis on the live broadcast related data by the cloud to obtain a cloud audit result. According to the method, a hierarchical collaborative auditing architecture of edge lightweight rule auditing and cloud deep analysis is adopted, so that the real-time performance, accuracy and response efficiency of live content auditing are effectively improved, the consumption of cloud resources is effectively reduced, and the auditing cost is reduced.

Description

Technical Field

[0001] Embodiments of the present application relate to the field of Internet technologies, and in particular, to a live content review method, apparatus, computing device, and computer storage medium. Background Art

[0002] With the rapid development of network technologies, many platforms have launched live broadcast services, and the number of users watching live broadcasts has increased rapidly. To maintain a good live broadcast environment, the platform needs to review the live content of the anchor. In the prior art, generally, the live video stream is sent to the cloud by pulling the stream from the anchor side or pushing the stream from the anchor side to the cloud, and the cloud reviews the content of the live video stream. This review method consumes too much cloud resources and increases the platform operation cost; or the reviewer enters the live broadcast room to watch the live content to review the live content. This review method requires a large amount of review costs and has low review efficiency, and it is difficult to review each live broadcast room. It can be seen that the existing live content review methods have problems such as excessive consumption of cloud resources and high review costs. Summary of the Invention

[0003] In view of the above problems, the present application provides a live content review method, apparatus, computing device, and computer storage medium, which are used to solve the following problems: The existing live content review methods have problems such as excessive consumption of cloud resources and high review costs.

[0004] According to one aspect of the embodiments of the present application, a live content review method is provided, including:

[0005] Using the edge review SDK integrated in the anchor side, collecting the live-related data of the anchor side, and reviewing the live-related data to obtain an edge review result;

[0006] Judging whether to trigger the cloud review mechanism according to the edge review result;

[0007] If so, pushing the live-related data to the cloud, and the cloud performs multimodal analysis on the live-related data to obtain a cloud review result.

[0008] Further, reviewing the live-related data to obtain an edge review result further includes:

[0009] Using the edge rule engine and AI inference framework in the edge review SDK, reviewing the live audio data, barrage data, and / or anchor portrait in the live-related data according to the edge rule library to obtain an edge review result.

[0010] Further, by using the edge rule engine and the AI inference framework in the edge audit SDK, the live audio data, bullet screen data, and / or host portrait in the live-related data are audited according to the edge rule library, and the obtained edge audit results further include:

[0011] Based on the AI inference framework, the live audio data is converted into text data, and sensitive word detection is performed on the text data according to the audio sensitive word detection rules in the edge rule library.

[0012] And / or, keyword detection is performed on the bullet screen data according to the bullet screen keyword detection rules in the edge rule library.

[0013] And / or, according to the host portrait audit rules in the edge rule library, it is audited whether there are abnormalities in the portrait features in the host portrait.

[0014] Further, the cloud performs multi-modal analysis on the live-related data to obtain cloud audit results, which further include:

[0015] Using the cloud rule engine in the cloud, multi-modal analysis is performed on the live video data, live audio data, live cover image, and / or interaction data between the host and the audience in the live-related data according to the cloud rule library to obtain cloud audit results.

[0016] Further, using the cloud rule engine in the cloud, multi-modal analysis is performed on the live video data, live audio data, live cover image, and / or interaction data between the host and the audience in the live-related data according to the cloud rule library to obtain cloud audit results, which further include:

[0017] According to the video detection rules in the cloud rule library, it is detected whether there is unqualified content in the frame images of the live video data.

[0018] And / or, according to the semantic detection rules in the cloud rule library, context semantic analysis is performed on the interaction data and live audio data to identify whether there is unqualified guiding behavior.

[0019] And / or, according to the consistency detection rules in the cloud rule library, it is detected whether the live cover image is consistent with the live content.

[0020] Further, the method further includes:

[0021] Obtain the audience behavior data of the audience.

[0022] The cloud analyzes whether there are abnormal behaviors in the audience behavior data based on the user behavior analysis AI model according to the audience behavior detection rules in the cloud rule library.

[0023] Further, after obtaining the cloud audit results, the method further includes:

[0024] Based on the edge audit result and the cloud audit result, determine whether to trigger the manual audit mechanism;

[0025] If so, send the live broadcast related data to the audit end for auditing to obtain the manual audit result.

[0026] Furthermore, the method further includes:

[0027] Store the edge audit result, the cloud audit result, and the manual audit result in the data warehouse for the feature learning model to perform iterative updates to generate a new edge rule library and a new cloud rule library;

[0028] Push the new edge rule library to the host end for the host end to update the edge rule library.

[0029] According to another aspect of the embodiments of the present application, there is provided a live broadcast content audit device, including:

[0030] An edge audit module, adapted to collect the live broadcast related data of the host end by using the edge audit SDK integrated in the host end, and audit the live broadcast related data to obtain an edge audit result;

[0031] A first judgment module, adapted to judge whether to trigger the cloud audit mechanism according to the edge audit result;

[0032] A cloud audit module, adapted to if the first judgment module determines that the cloud audit mechanism is triggered, push the live broadcast related data to the cloud, and the cloud performs multi-modal analysis on the live broadcast related data to obtain a cloud audit result.

[0033] Furthermore, the edge audit module is further adapted to:

[0034] Use the edge rule engine and the AI inference framework in the edge audit SDK to audit the live broadcast audio data, the bullet screen data, and / or the host portrait in the live broadcast related data according to the edge rule library to obtain an edge audit result.

[0035] Furthermore, the edge audit module is further adapted to:

[0036] Based on the AI inference framework, convert the live broadcast audio data into text data, and perform sensitive word detection on the text data according to the audio sensitive word detection rule in the edge rule library;

[0037] And / or, perform keyword detection on the bullet screen data according to the bullet screen keyword detection rule in the edge rule library;

[0038] And / or, according to the host portrait audit rule in the edge rule library, audit whether there are abnormalities in the portrait features in the host portrait.

[0039] Further, the cloud audit module is further adapted to:

[0040] Utilize the cloud rule engine in the cloud to perform multimodal analysis on the live video data, live audio data, live cover image, and / or interaction data between the host and the audience in the live-related data according to the cloud rule library, and obtain the cloud audit result.

[0041] Further, the cloud audit module is further adapted to:

[0042] Detect whether there is unqualified content in the frame images of the live video data according to the video detection rules in the cloud rule library;

[0043] And / or, perform context semantic analysis on the interaction data and live audio data according to the semantic detection rules in the cloud rule library to identify whether there is unqualified guiding behavior;

[0044] And / or, detect whether the live cover image is consistent with the live content according to the consistency detection rules in the cloud rule library.

[0045] Further, the cloud audit module is further adapted to:

[0046] Obtain the audience behavior data of the audience;

[0047] The cloud analyzes whether there is abnormal behavior in the audience behavior data based on the user behavior analysis AI model according to the audience behavior detection rules in the cloud rule library.

[0048] Further, the device further includes: a second judgment module and an artificial audit module;

[0049] The second judgment module is adapted to: judge whether to trigger the artificial audit mechanism according to the edge audit result and the cloud audit result;

[0050] The artificial audit module is adapted to: if the second judgment module determines to trigger the artificial audit mechanism, send the live-related data to the audit end for audit to obtain the artificial audit result.

[0051] Further, the device further includes: a dynamic optimization module; the dynamic optimization module is adapted to:

[0052] Store the edge audit result, the cloud audit result, and the artificial audit result in the data warehouse for the feature learning model to perform iterative update, and generate a new edge rule library and a new cloud rule library;

[0053] Push the new edge rule library to the host end for the host end to update the edge rule library.

[0054] According to another aspect of the embodiments of the present application, a computing device is provided, including: a processor, a memory, a communication interface, and a communication bus. The processor, the memory, and the communication interface complete communication with each other through the communication bus;

[0055] The memory is used to store at least one executable instruction, and the executable instruction causes the processor to perform the operations corresponding to the above live content review method.

[0056] According to still another aspect of the embodiments of the present application, a computer storage medium is provided. The storage medium stores at least one executable instruction, and the executable instruction causes the processor to perform the operations corresponding to the above live content review method.

[0057] According to yet another aspect of the embodiments of the present application, a computer program product is provided, including at least one executable instruction, and the executable instruction causes the processor to perform the operations corresponding to the above live content review method.

[0058] According to the live content review method, device, computing device, and computer storage medium provided by the embodiments of the present application, an edge review SDK is integrated at the anchor end. By using the edge review SDK, data collection and preliminary review can both be completed at the edge end, realizing fast preliminary review at the edge end; according to the edge review results, it is judged whether it is necessary to trigger the cloud review mechanism. In the case where it is necessary to trigger the cloud review mechanism, seamless docking with in-depth cloud analysis is carried out, and the live-related data is automatically pushed to the cloud, and the cloud conducts comprehensive and in-depth multi-modal analysis on the live-related data; this solution adopts a hierarchical collaborative review architecture of edge lightweight rule review and cloud in-depth analysis, effectively improving the real-time performance, accuracy, and response efficiency of live content review, optimizing the live content review method, and effectively reducing the consumption of cloud resources and the review cost.

[0059] The above description is only an overview of the technical solutions of the embodiments of the present application. In order to be able to understand the technical means of the embodiments of the present application more clearly, it can be implemented according to the content of the description. And in order to make the above and other purposes, features, and advantages of the embodiments of the present application more obvious and understandable, the following specifically describes the specific implementation manners of the embodiments of the present application. Description of the Drawings

[0060] By reading the following detailed description of the preferred embodiments, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the embodiments of the present application. And throughout the drawings, the same reference numerals are used to represent the same components. In the drawings:

[0061] Figure 1 A flowchart showing the live content review method according to an embodiment of the present application is shown;

[0062] Figure 2a Shows a schematic flowchart of a live content review method according to another embodiment of the present application;

[0063] Figure 2b Shows a schematic diagram of the principle of the live content review method according to an embodiment of the present application;

[0064] Figure 3 Shows a structural block diagram of a live content review device according to an embodiment of the present application;

[0065] Figure 4 Shows a schematic structural diagram of a computing device according to an embodiment of the present application. Detailed implementation manners

[0066] Hereinafter, exemplary embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be completely conveyed to those skilled in the art.

[0067] It should be noted that the information and data involved in the present application (including but not limited to interaction data between the anchor and the audience, anchor portraits, audience behavior data, etc.) are all information and data authorized by the users or fully authorized by all parties. The collection, use, and processing of relevant data need to comply with relevant laws, regulations, and standards of relevant countries and regions, and desensitize the data. In addition, corresponding operation entrances are provided for users to choose to authorize or reject.

[0068] First, the noun terms involved in one or more embodiments of the present application are explained.

[0069] Anchor side: Refers to the terminal tool used by the anchor during the live broadcast, which is the edge side and has functions such as video capture, audio capture, and interactive functions such as chatting.

[0070] Edge review SDK: In the embodiments of the present application, it refers to a development component embedded in the anchor side, providing function interfaces such as data collection, edge rule engine, and AI inference.

[0071] Anchor portrait: Refers to the multi-dimensional data content constructed based on the historical behavior data and live broadcast performance of the anchor.

[0072] Edge rule engine: It refers to the rule judgment module integrated in the edge audit SDK, which is used to preliminarily screen and judge the collected data, and determine whether the data should be pushed to the local lightweight AI inference framework or pushed to the cloud for in-depth analysis.

[0073] AI inference framework: It refers to the real-time inference module implemented based on the lightweight AI model, which is mainly used for rapid processing and judgment at the edge side, such as audio-to-text conversion, sensitive word detection, etc.

[0074] Cloud rule engine: It refers to the rule judgment module deployed in the cloud, which is used to screen and judge the data, and determine whether the data should be pushed to large-scale and complex AI models in the cloud for review or pushed to the review side for manual review.

[0075] Figure 1 The flowchart of the live content audit method according to an embodiment of the present application is shown, as Figure 1 shown, the method includes the following steps:

[0076] Step S101, using the edge audit SDK integrated in the host side, collect the live-related data of the host side, and audit the live-related data to obtain the edge audit result.

[0077] The embodiment of the present application proposes a hierarchical collaborative audit architecture for edge lightweight rule audit and cloud in-depth analysis. The edge audit SDK is integrated in the host side. The edge audit SDK includes an edge rule engine and a lightweight AI inference framework. Using the edge audit SDK enables data collection and preliminary audit to be completed at the edge side. In the case where the cloud audit mechanism needs to be triggered, it is then seamlessly docked with the cloud in-depth analysis, and the live-related data is automatically pushed to the cloud for comprehensive and in-depth multimodal analysis.

[0078] The host side refers to the terminal tool used by the host during the live broadcast process. The host conducts the live broadcast through the host side. During the live broadcast process, the host side uses the integrated edge audit SDK to continuously collect the live-related data of the host side. Specifically, the live-related data may include live video data, live audio data, live cover image, host portrait, interaction data between the host and the audience, etc. The interaction data may include bullet screen data, comment data, like data, etc.

[0079] Among them, the edge audit SDK also provides an edge rule engine and a lightweight AI inference framework. Through the edge rule engine and the lightweight AI inference framework, the live broadcast-related data collected can be preliminarily audited at the edge side. For example, detecting sensitive words in the live audio, detecting keyword in the bullet screen, auditing the host portrait, etc., so as to obtain the edge audit result. In the embodiment of the present application, a host portrait audit mechanism is introduced in the edge audit process, and the audit is combined with the host portrait, effectively enhancing the risk identification ability, improving the audit accuracy and the intelligent decision-making effect of the edge rule engine.

[0080] Step S102: According to the edge audit result, determine whether to trigger the cloud audit mechanism; if so, execute step S103.

[0081] The edge audit result records the risk situations existing in this live broadcast found during the edge audit process. Specifically, it may include whether unqualified content is detected, the specific information of the detected unqualified content, the risk assessment result of the host portrait, etc. In the embodiment of the present application, the edge rule engine in the edge audit SDK can be used to determine whether to trigger the cloud audit mechanism according to the edge audit result.

[0082] Step S103: Push the live broadcast-related data to the cloud, and the cloud performs multimodal analysis on the live broadcast-related data to obtain the cloud audit result.

[0083] A large-scale and complex AI model for live broadcast content audit is deployed in the cloud. When it is determined in step S102 that the cloud audit mechanism needs to be triggered, the live broadcast-related data is automatically pushed to the cloud. The cloud uses the cloud rule engine to perform comprehensive and in-depth multimodal analysis through the large-scale and complex AI model in combination with the specific live broadcast scenario and multi-dimensional live broadcast-related data. For example, detecting the frame images in the live video data, identifying guiding behaviors and consistency detection by performing context semantic analysis on the interaction data and the live audio data, etc., so as to obtain the cloud audit result.

[0084] According to the live broadcast content audit method provided by the embodiment of the present application, the edge audit SDK is integrated at the host side. By using the edge audit SDK, data collection and preliminary audit can be completed at the edge side, realizing fast preliminary audit at the edge side; according to the edge audit result, determine whether to trigger the cloud audit mechanism. When it is necessary to trigger the cloud audit mechanism, then seamlessly dock with the in-depth analysis in the cloud, automatically push the live broadcast-related data to the cloud, and the cloud performs comprehensive and in-depth multimodal analysis on the live broadcast-related data; this solution adopts a hierarchical collaborative audit architecture of edge lightweight rule audit and cloud in-depth analysis, effectively improving the real-time performance, accuracy and response efficiency of live broadcast content audit, optimizing the live broadcast content audit method, and effectively reducing the consumption of cloud resources and the audit cost.

[0085] Figure 2a shows a schematic flowchart of a live content review method according to another embodiment of the present application, as Figure 2a shown, the method includes the following steps:

[0086] Step S201, using the edge review SDK integrated in the host end, collect live-related data of the host end.

[0087] The embodiments of the present application fully integrate the review capabilities of the edge side and the cloud side, adopt a hierarchical collaborative review architecture of edge lightweight rule review and cloud deep analysis, and effectively improve the real-time performance, accuracy and response efficiency of live content review. The embodiments of the present application propose an integrated SDK solution. Among them, the edge review SDK is integrated in the host end. The edge review SDK includes an edge rule engine and a lightweight AI inference framework, and the AI inference framework can be implemented based on a lightweight AI model. By integrating the edge review SDK in the host end, not only can the data collection of live-related data be conveniently realized, but also the rapid preliminary review of the edge side is realized through lightweight rule judgment and real-time AI inference. During the live broadcast process of the host end, the live-related data of the host end is continuously collected by using the integrated edge review SDK.

[0088] Specifically, the live-related data may include live video data, live audio data, live cover image, host portrait, interaction data between the host and the audience, live room title, etc. The interaction data may include bullet screen data, comment data, like data, etc. Among them, the host portrait is a multi-dimensional data content constructed based on the host's historical behavior data and live performance, etc., and specifically may include portrait features such as the host's historical review result records, live content preferences, setting behaviors, live behaviors, interaction behaviors, etc.

[0089] Step S202, using the edge rule engine and AI inference framework in the edge review SDK, review the live audio data, bullet screen data and / or host portrait in the live-related data according to the edge rule library, and obtain the edge review result.

[0090] Among them, the host end can use the edge rule engine and AI inference framework in the integrated edge review SDK to quickly and preliminarily review the live audio data, bullet screen data and other data in the live-related data. The live-related data collected in step S201 is sent to the edge rule engine for processing, and live audio sensitive word detection, bullet screen keyword detection and / or host portrait review are performed. The edge rule library records the review rules required in the edge review process, which may include audio sensitive word detection rules, bullet screen keyword detection rules, host portrait review rules, etc.

[0091] For live audio sensitive word detection, based on the AI inference framework, live audio data can be converted into text data, and sensitive word detection can be performed on the text data according to the audio sensitive word detection rules in the edge rule library. For example, based on the speech conversion model in the AI inference framework, live audio data can be quickly and accurately converted into text data; then, according to the audio sensitive word detection rules, each word in the text data is quickly compared with the sensitive words in the preset sensitive word library to detect whether the text data contains sensitive words, which sensitive words are included, and the occurrence frequency of each sensitive word, etc.

[0092] For barrage keyword detection, keyword detection can be performed on barrage data according to the barrage keyword detection rules in the edge rule library. For example, according to the barrage keyword detection rules, the text content in the barrage data posted by the audience is quickly compared with the keywords in the preset keyword library to detect whether the barrage data contains high-risk content.

[0093] For host portrait review, according to the host portrait review rules in the edge rule library, it can be reviewed whether there are abnormalities in the portrait features in the host portrait. Among them, the host portrait includes portrait features such as the host's historical review result records, live content preferences, setting behaviors, live behaviors, and interaction behaviors. Specifically, the historical review result records record the unqualified content that the host had in historical live broadcasts; the live content preference is the classification of the host's historical live content. For example, game classification, store visit classification, etc.; the setting behaviors may include the live partition selection and live cover image replacement situation selected by the host for their own live content; the live behaviors may include the host's language, clothing, actions, etc. during the live broadcast; the interaction behaviors may include the host's bullet screen responses, comment replies, etc. For example, according to the host portrait review rules, it is reviewed whether there are abnormal situations in the host portrait such as the host frequently changing the live cover image, whether there are ambiguous guiding words in the language, and whether the live partition selection does not match the actual live content. In the actual application scenario, the review threshold, review criteria, etc. in the host portrait review rules can be dynamically adjusted according to the review requirements of the specific scenario.

[0094] In the embodiment of this application, through the combination of the edge rule engine and the lightweight AI inference framework, fast preliminary review is achieved; and, a host portrait review mechanism is introduced in the edge review process, and the review is combined with the host portrait, effectively enhancing the risk recognition ability and improving the review accuracy and the intelligent decision-making effect of the edge rule engine.

[0095] Step S203, according to the edge review result, judge whether to trigger the cloud review mechanism; if so, execute step S204.

[0096] Among them, the edge audit results may include whether unqualified content is detected, specific information about the detected unqualified content, the risk assessment results of the host portrait, etc. Based on the edge audit results, it is determined whether it is necessary to trigger the cloud audit mechanism.

[0097] Specifically, when it is known from the edge audit results that no unqualified content is detected and the risk assessment of the host portrait passes (for example, the risk assessment level of the host portrait is a low-risk level), it indicates that the live content of this live broadcast meets the requirements and there is no need to trigger the cloud audit mechanism.

[0098] When it is known from the edge audit results that no unqualified content is detected but the risk assessment of the host portrait fails (for example, the risk assessment level of the host portrait is a medium-risk level or above), it indicates that the live content of this live broadcast may not meet the requirements. In this case, it can be determined that it is necessary to trigger the cloud audit mechanism in order to perform comprehensive and in-depth multimodal analysis through the cloud to achieve accurate auditing.

[0099] When it is known from the edge audit results that unqualified content is detected, then it can be comprehensively determined whether it is necessary to trigger the cloud audit mechanism to perform comprehensive and in-depth multimodal analysis based on the specific information of the detected unqualified content and the risk assessment results of the host portrait. For example, when the detected unqualified content belongs to the content of the low-risk level and the risk assessment level of the host portrait is the low-risk level, corresponding processing can be directly performed, such as giving a warning, etc., and there is no need to trigger the cloud audit mechanism anymore; when the detected unqualified content belongs to the content of the medium-risk level and the risk assessment level of the host portrait is the low-risk level, it can be determined that it is necessary to trigger the cloud audit mechanism; in addition, when the detected unqualified content belongs to the content of the high-risk level, it indicates that the live content of this live broadcast is significantly seriously unqualified. In this case, corresponding processing can be directly performed, such as closing the live broadcast, etc., and there is no need to trigger the cloud audit mechanism anymore.

[0100] After obtaining the edge audit results, the edge audit results can be directly fed back to the host side so that the host can know the edge audit results and correct the unqualified content in a timely manner.

[0101] Step S204, push the live broadcast related data to the cloud, and use the cloud rule engine in the cloud to perform multimodal analysis on the live video data, live audio data, live cover image, and / or the interaction data between the host and the audience in the live broadcast related data to obtain the cloud audit results.

[0102] In the case of determining that the cloud audit mechanism needs to be triggered, the live broadcast related data is automatically pushed to the cloud for in-depth cloud analysis. Using the cloud rule engine in the cloud, multi-modal analysis is performed on the live video data, live audio data, live cover image, and / or the interaction data between the host and the audience in the live broadcast related data according to the cloud rule library. By synthesizing the multi-modal analysis results of video, audio, image, etc., the cloud audit result is finally obtained. The cloud rule library records the audit rules required in the cloud audit process, which may include video detection rules, semantic detection rules, consistency detection rules, etc.

[0103] For video detection, according to the video detection rules in the cloud rule library, it can be detected whether there is unqualified content in the frame images of the live video data. For example, each frame image in the live video data is extracted, and based on the image detection AI model, according to the video detection rules, it is quickly and accurately detected whether there is unqualified content in each frame image, such as whether there is unqualified content in the host's clothing, actions, slogans in the background, and items placed in the frame image.

[0104] For semantic detection, according to the semantic detection rules in the cloud rule library, context semantic analysis can be performed on the interaction data and live audio data to identify whether there is unqualified guiding behavior. For example, based on the semantic analysis AI model, according to the semantic detection rules, context semantic analysis is performed on the interaction data such as bullet screens, comments, likes, and live audio data to identify whether there is unqualified guiding behavior.

[0105] For consistency detection, according to the consistency detection rules in the cloud rule library, it can be detected whether the live cover image is consistent with the live content. For example, based on the live content AI model, the live video data and live audio data can be analyzed to identify the live content; and the content of the live cover image can be identified, and then the content of the live cover image is compared with the live content for consistency detection to identify whether there is unqualified drainage behavior. In addition, the title of the live broadcast room can also be compared with the live content for consistency detection to identify whether there is unqualified drainage behavior.

[0106] Considering that the viewer behavior data of live stream viewers can also reflect the situation of the live stream content to a certain extent, during the in-depth analysis process in the cloud, the viewer behavior data can also be analyzed to conduct viewer behavior detection. Then, the cloud rule library can also include viewer behavior detection rules. Obtain the viewer behavior data of the viewer, and based on the user behavior analysis AI model in the cloud, and according to the viewer behavior detection rules in the cloud rule library, analyze whether there are abnormal behaviors in the viewer behavior data and evaluate the risk level of the live stream content. Among them, the viewer behavior data can include viewer click-through rate, viewer stay duration, the number of viewers entering the live stream room, viewer portraits, etc. The viewer portrait is a multi-dimensional data content constructed based on the viewer's historical behavior data, etc., and specifically can include portrait features such as the viewer's historical interaction behaviors (such as bullet screens, comments, gift giving, etc.) and the preference for the live stream content watched.

[0107] Step S205, based on the edge audit result and the cloud audit result, determine whether to trigger the manual audit mechanism; if so, execute step S206.

[0108] To further improve the audit accuracy, in the embodiment of the present application, on the basis of the hierarchical collaborative audit architecture of edge lightweight rule audit and cloud in-depth analysis, a manual audit mechanism is also set up. Based on the specific content in the edge audit result and the cloud audit result, combined with the audit requirements of the specific scenario, automatically determine whether to trigger the manual audit mechanism to ensure that in the case of high-risk live stream content, the final audit is carried out manually. Through this processing method, it is possible to effectively reduce the investment in human resources and reduce the audit cost while ensuring the audit accuracy.

[0109] Step S206, send the live stream related data to the audit end for audit to obtain the manual audit result.

[0110] In the case of determining that the manual audit mechanism needs to be triggered, send the live stream related data to the audit end, and the audit personnel on the audit end side conduct a manual audit to obtain the manual audit result.

[0111] Optionally, the edge audit result, the cloud audit result, and the manual audit result can be fed back to the host side, so that the host side can display the audit status in real time and perform corresponding processing when necessary, such as content interception, prompt warning, live stream closing processing, etc.

[0112] Step S207, store the edge audit result, the cloud audit result, and the manual audit result in the data warehouse for the feature learning model to perform iterative updates to generate a new edge rule library and a new cloud rule library.

[0113] The edge rule library and the cloud rule library can flexibly set different rules for different scenarios (such as game scenarios, information interpretation scenarios, etc.), providing an extensible and configurable audit architecture that can well adapt to the needs of multiple scenarios. Moreover, the edge rule library and the cloud rule library can be dynamically updated, thereby further improving the accuracy of live content audit.

[0114] Store the generated edge audit results, cloud audit results, and manual audit results in the data warehouse, enabling the feature learning model to iteratively update according to each audit result in the data warehouse, learn the latest live content features, and generate a new edge rule library and a new cloud rule library.

[0115] Step S208, push the new edge rule library to the host side for the host side to update the edge rule library.

[0116] Push the generated new edge rule library to the host side, enabling the host side to update the edge rule library and realizing the dynamic update of the edge audit policy. In addition, provide the new cloud rule library to the cloud to update the cloud rule library in the cloud, realizing the dynamic update of the cloud deep analysis policy. By dynamically updating the edge rule library and the cloud rule library, the construction of the closed-loop optimization system is completed, effectively improving the accuracy of live content audit.

[0117] Figure 2b Shows a schematic diagram of the principle of the live content audit method according to an embodiment of the present application, as Figure 2b shown, this solution can be divided into three parts: data collection, dynamic decision-making, and feedback optimization.

[0118] Among them, in the data collection part, use the edge audit SDK integrated in the host side to collect live-related data, and the live-related data may include live video data, live audio data, interaction data, etc.

[0119] In the dynamic decision-making part, introduce the specific scenario and the host portrait into the processing of the rule engine, and the rule engine includes an edge rule engine and a cloud rule engine; the host side uses the edge rule engine and the AI inference framework to preliminarily audit the live-related data according to the edge rule library; the cloud uses the cloud rule engine to conduct a comprehensive and in-depth multi-modal analysis of the live-related data according to the cloud rule library; during the multi-modal analysis process, task allocation is performed, and it is decided to send the data to the AI model of the specific modality for audit, and the manual audit mechanism can also be triggered; the obtained audit results (including edge audit results, cloud audit results, and manual audit results) are fed back to the host side and stored in the data warehouse to achieve result reach.

[0120] In the feedback optimization section, the feature learning model iteratively updates according to each audit result in the data warehouse to generate a new edge rule library and a new cloud rule library; then the new edge rule library is pushed to the host side for the host side to update the edge rule library.

[0121] According to the live content audit method provided by the embodiments of the present application, the audit capabilities of the edge side and the cloud side are fully integrated, and a hierarchical collaborative audit architecture of edge lightweight rule audit and cloud deep analysis is adopted, which effectively improves the real-time performance, accuracy and response efficiency of live content audit, effectively reduces the consumption of cloud resources, and reduces the audit cost; by integrating the edge audit SDK on the host side, not only can the data collection of live-related data be conveniently realized, but also through the combination of the edge rule engine and the lightweight AI inference framework in the edge audit SDK, live audio sensitive word detection, bullet screen keyword detection, host portrait audit, etc. can be carried out, realizing the rapid preliminary audit of the edge side; moreover, a host portrait audit mechanism is introduced in the edge audit process, and the audit is carried out in combination with the host portrait, which effectively enhances the risk identification ability, further improves the audit accuracy and the intelligent decision-making effect of the edge rule engine; according to the edge audit result, it is judged whether it is necessary to trigger the cloud audit mechanism. In the case where it is necessary to trigger the cloud audit mechanism, the live-related data is automatically pushed to the cloud, and the cloud uses large-scale and complex AI models to perform multi-modal analysis on live video data, live audio data, live cover images, interactive data, etc., and comprehensively combines the multi-modal analysis results of video, audio, images, etc. to finally obtain the cloud audit result; on the basis of the hierarchical collaborative audit architecture of edge lightweight rule audit and cloud deep analysis, an artificial audit mechanism is also set up. According to the specific content in the edge audit result and the cloud audit result, combined with the audit requirements of the specific scenario, it is automatically judged whether it is necessary to trigger the artificial audit mechanism, which can effectively reduce the human resource input while ensuring the audit accuracy and reduce the audit cost; in addition, it can also be iteratively updated according to each obtained audit result to generate a new edge rule library and a new cloud rule library, realizing the dynamic update of the audit strategy and completing the construction of the closed-loop optimization system.

[0122] Figure 3 The structural block diagram of a live content audit device according to an embodiment of the present application is shown, as Figure 3 shown, the device includes: an edge audit module 310, a first judgment module 320 and a cloud audit module 330.

[0123] The edge audit module 310 is adapted to: use the edge audit SDK integrated in the host side to collect live-related data of the host side and audit the live-related data to obtain an edge audit result.

[0124] The first judgment module 320 is adapted to: determine whether to trigger the cloud audit mechanism according to the edge audit result.

[0125] The cloud audit module 330 is adapted to: if the first judgment module 320 determines to trigger the cloud audit mechanism, push the live broadcast related data to the cloud, and the cloud performs multimodal analysis on the live broadcast related data to obtain the cloud audit result.

[0126] Furthermore, the edge audit module 310 is further adapted to: utilize the edge rule engine and the AI inference framework in the edge audit SDK to audit the live audio data, bullet screen data, and / or the host portrait in the live broadcast related data according to the edge rule library to obtain the edge audit result.

[0127] Furthermore, the edge audit module 310 is further adapted to: based on the AI inference framework, convert the live audio data into text data, and perform sensitive word detection on the text data according to the audio sensitive word detection rule in the edge rule library; and / or, perform keyword detection on the bullet screen data according to the bullet screen keyword detection rule in the edge rule library; and / or, audit whether there are abnormalities in the portrait features in the host portrait according to the host portrait audit rule in the edge rule library.

[0128] Furthermore, the cloud audit module 330 is further adapted to: utilize the cloud rule engine in the cloud to perform multimodal analysis on the live video data, live audio data, live cover image, and / or the interaction data between the host and the audience in the live broadcast related data according to the cloud rule library to obtain the cloud audit result.

[0129] Furthermore, the cloud audit module 330 is further adapted to: detect whether there is unqualified content in the frame images of the live video data according to the video detection rule in the cloud rule library; and / or, perform context semantic analysis on the interaction data and the live audio data according to the semantic detection rule in the cloud rule library to identify whether there is unqualified guiding behavior; and / or, detect whether the live cover image is consistent with the live content according to the consistency detection rule in the cloud rule library.

[0130] Furthermore, the cloud audit module 330 is further adapted to: obtain the audience behavior data of the audience; and the cloud analyzes whether there is abnormal behavior in the audience behavior data according to the audience behavior detection rule in the cloud rule library based on the user behavior analysis AI model.

[0131] Further, the device further includes: a second determination module 340 and a manual review module 350. The second determination module 340 is adapted to: determine whether to trigger a manual review mechanism according to the edge review result and the cloud review result; the manual review module 350 is adapted to: if the second determination module 340 determines to trigger the manual review mechanism, send the live broadcast related data to the review end for review to obtain a manual review result.

[0132] Further, the device further includes: a dynamic optimization module 360; the dynamic optimization module 360 is adapted to: store the edge review result, the cloud review result, and the manual review result in a data warehouse for the feature learning model to perform iterative updates to generate a new edge rule library and a new cloud rule library; push the new edge rule library to the host end for the host end to update the edge rule library.

[0133] The descriptions of the above modules refer to the corresponding descriptions in the method embodiments and will not be elaborated herein.

[0134] The live content review device provided by the embodiments of the present application fully integrates the review capabilities of the edge side and the cloud side, adopts a hierarchical collaborative review architecture of edge lightweight rule review and cloud deep analysis, effectively improves the real-time performance, accuracy and response efficiency of live content review, effectively reduces the consumption of cloud resources, and reduces the review cost; by integrating the edge review SDK on the host side, it can not only conveniently realize the data collection of live-related data, but also through the combination of the edge rule engine and the lightweight AI inference framework in the edge review SDK, it can perform live audio sensitive word detection, barrage keyword detection, host portrait review, etc., realizing the rapid preliminary review of the edge side; moreover, in the edge review process, a host portrait review mechanism is introduced, and the review is carried out in combination with the host portrait, effectively enhancing the risk identification ability, further improving the review accuracy and the intelligent decision-making effect of the edge rule engine; according to the edge review results, it is judged whether it is necessary to trigger the cloud review mechanism. In the case where it is necessary to trigger the cloud review mechanism, the live-related data is automatically pushed to the cloud, and the cloud uses large-scale and complex AI models to perform multimodal analysis on live video data, live audio data, live cover images, interaction data, etc., and comprehensively combines the multimodal analysis results of video, audio, images, etc., and finally obtains the cloud review result; on the basis of the hierarchical collaborative review architecture of edge lightweight rule review and cloud deep analysis, an artificial review mechanism is also set up, and according to the specific content in the edge review result and the cloud review result, combined with the review requirements of the specific scenario, it is automatically judged whether it is necessary to trigger the artificial review mechanism, which can effectively reduce the human resource input while ensuring the review accuracy and reduce the review cost; in addition, it can also be iteratively updated according to the obtained review results to generate a new edge rule library and a new cloud rule library, realizing the dynamic update of the review strategy and completing the construction of the closed-loop optimization system.

[0135] The embodiments of the present application provide a non-volatile computer storage medium, and the computer storage medium stores at least one executable instruction or computer program, and the executable instruction or computer program can enable the processor to perform the operations corresponding to the live content review method in any of the above method embodiments.

[0136] The embodiments of the present application provide a computer program product, and the computer program product includes at least one executable instruction or computer program, and the executable instruction or computer program can enable the processor to perform the operations corresponding to the live content review method in any of the above method embodiments.

[0137] Figure 4 The structural schematic diagram of a computing device according to an embodiment of the present application is shown, and the specific implementation of the computing device is not limited in the specific embodiments of the present application.

[0138] As Figure 4As shown, the computing device may include: a processor 402, a communications interface 404, a memory 406, and a communication bus 408.

[0139] Among them: The processor 402, the communications interface 404, and the memory 406 communicate with each other through the communication bus 408. The communications interface 404 is used to communicate with network elements of other devices such as clients or other servers. The processor 402 is used to execute the program 410, and specifically can execute the relevant steps in the above-described embodiments of the live content review method for the computing device.

[0140] Specifically, the program 410 may include program code, and the program code includes computer operation instructions.

[0141] The processor 402 may be a central processing unit CPU, or a specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application. One or more processors included in the computing device may be of the same type of processor, such as one or more CPUs; or may be of different types of processors, such as one or more CPUs and one or more ASICs.

[0142] The memory 406 is used to store the program 410. The memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk memory.

[0143] The program 410 is specifically used to cause the processor 402 to execute the live content review method in any of the above method embodiments. For the specific implementation of each step in the program 410, reference may be made to the corresponding steps and descriptions in the corresponding units in the above live content review embodiments, which will not be elaborated here. Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described devices and modules can refer to the corresponding process descriptions in the foregoing method embodiments, which will not be elaborated here.

[0144] The algorithms and displays provided herein are not inherently related to any specific computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The structure required to construct such a system is obvious from the above description. In addition, the embodiments of the present application are not directed to any specific programming language. It should be understood that the content of the embodiments of the present application described herein can be implemented using various programming languages, and the description of the specific language above is to disclose the best mode of the embodiments of the present application.

[0145] In the specification provided herein, a large number of specific details are set forth. However, it will be understood that embodiments of the present application may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0146] Similarly, it should be understood that in order to streamline the present disclosure and assist in understanding one or more of the various inventive aspects, in the foregoing description of the exemplary embodiments of the present application, the various features of the present application are sometimes grouped together in a single embodiment, figure, or description thereof. However, the disclosed method should not be construed as reflecting an intention that the claimed embodiments of the present application require more features than are expressly recited in each claim. Rather, as reflected in the following claims, the inventive aspects lie in less than all the features of the single foregoing disclosed embodiment. Thus, the claims following the detailed description are hereby expressly incorporated into the detailed description, with each claim standing on its own as a separate embodiment of the present application.

[0147] Those skilled in the art will appreciate that the modules in the devices in the embodiments can be adaptively changed and disposed in one or more devices different from those of the embodiments. The modules or units or components in the embodiments can be combined into one module or unit or component, and in addition, they can be divided into multiple sub-modules or sub-units or sub-components. Except that at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all the features disclosed in this specification (including the accompanying claims, abstract and drawings) and all the processes or units of any method or device so disclosed. Unless otherwise expressly stated, each feature disclosed in this specification (including the accompanying claims, abstract and drawings) can be replaced by an alternative feature providing the same, equivalent or similar purpose.

[0148] In addition, those skilled in the art will be able to understand that although some of the embodiments described herein include certain features included in other embodiments but not other features, the combination of the features of different embodiments means that it is within the scope of the embodiments of the present application and forms different embodiments. For example, in the following claims, any one of the claimed embodiments can be used in any combination.

[0149] Each component embodiment of the embodiments of the present application may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. Those skilled in the art should understand that a microprocessor or a digital signal processor (DSP) may be used in practice to implement some or all of the functions of some or all of the components according to the embodiments of the present application. The embodiments of the present application may also be implemented as a device or apparatus program (for example, a computer program and a computer program product) for executing part or all of the methods described herein. Such a program implementing the embodiments of the present application may be stored on a computer-readable medium, or may be in the form of one or more signals. Such signals may be downloaded from an Internet website, or provided on a carrier signal, or in any other form.

[0150] It should be noted that the above embodiments illustrate the embodiments of the present application rather than limit the embodiments of the present application, and those skilled in the art can design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in the claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The embodiments of the present application may be implemented by means of hardware including several different elements and by means of a suitably programmed computer. In the unit claims listing several devices, several of these devices may be embodied by the same item of hardware. The use of the words first, second, and third, etc. does not denote any order. These words may be interpreted as names.

Claims

1. A live content review method, comprising: Using the edge review SDK integrated in the host end to collect the live-related data of the host end, and reviewing the live-related data to obtain an edge review result; Judging whether to trigger the cloud review mechanism according to the edge review result; If so, pushing the live-related data to the cloud, and performing multimodal analysis on the live-related data by the cloud to obtain a cloud review result.

2. The method according to claim 1, wherein the reviewing the live-related data to obtain an edge review result further comprises: Using the edge rule engine and the AI inference framework in the edge review SDK to review the live audio data, barrage data, and / or host portrait in the live-related data according to the edge rule library to obtain the edge review result.

3. The method according to claim 2, wherein the using the edge rule engine and the AI inference framework in the edge review SDK to review the live audio data, barrage data, and / or host portrait in the live-related data according to the edge rule library to obtain the edge review result further comprises: Based on the AI inference framework, converting the live audio data into text data, and performing sensitive word detection on the text data according to the audio sensitive word detection rule in the edge rule library; And / or, performing keyword detection on the barrage data according to the barrage keyword detection rule in the edge rule library; And / or, reviewing whether there are abnormalities in the portrait features in the host portrait according to the host portrait review rule in the edge rule library.

4. The method according to any one of claims 1-3, wherein the performing multimodal analysis on the live-related data by the cloud to obtain a cloud review result further comprises: Using the cloud rule engine of the cloud to perform multimodal analysis on the live video data, live audio data, live cover image, and / or interaction data between the host and the audience in the live-related data according to the cloud rule library to obtain the cloud review result.

5. The method according to claim 4, wherein the using the cloud rule engine of the cloud to perform multimodal analysis on the live video data, live audio data, live cover image, and / or interaction data between the host and the audience in the live-related data according to the cloud rule library to obtain the cloud review result further comprises: Detecting whether there is unqualified content in the frame images of the live video data according to the video detection rule in the cloud rule library; And / or, performing context semantic analysis on the interaction data and the live audio data according to the semantic detection rule in the cloud rule library to identify whether there is unqualified guiding behavior; And / or, detecting whether the live cover image is consistent with the live content according to the consistency detection rule in the cloud rule library.

6. The method according to claim 4 or 5, wherein the method further comprises: Obtaining the audience behavior data of the audience; Based on the user behavior analysis AI model by the cloud, analyze whether there are abnormal behaviors in the audience behavior data according to the audience behavior detection rules in the cloud rule library.

7. The method according to any one of claims 1-6, after obtaining the cloud audit result, the method further includes: Judging whether to trigger the manual audit mechanism according to the edge audit result and the cloud audit result; If so, send the live broadcast related data to the audit end for audit to obtain the manual audit result.

8. The method according to any one of claims 1-7, the method further includes: Store the edge audit result, the cloud audit result, and the manual audit result in the data warehouse for the feature learning model to perform iterative updates to generate a new edge rule library and a new cloud rule library; Push the new edge rule library to the host end for the host end to update the edge rule library.

9. A live content audit device, including: An edge audit module, adapted to collect the live broadcast related data of the host end by using the edge audit SDK integrated in the host end, and audit the live broadcast related data to obtain an edge audit result; A first judgment module, adapted to judge whether to trigger the cloud audit mechanism according to the edge audit result; A cloud audit module, adapted to if the first judgment module determines that the cloud audit mechanism is triggered, push the live broadcast related data to the cloud, and the cloud performs multimodal analysis on the live broadcast related data to obtain a cloud audit result.

10. A computing device, comprising: A processor, a memory, a communication interface, and a communication bus, and the processor, the memory, and the communication interface complete communication with each other through the communication bus; The memory is used to store at least one executable instruction, and the executable instruction causes the processor to execute the operations corresponding to the live content audit method according to any one of claims 1-8.

11. A computer storage medium, in which at least one executable instruction is stored, and the executable instruction causes the processor to execute the operations corresponding to the live content audit method according to any one of claims 1-8.

12. A computer program product, including at least one executable instruction, and the executable instruction causes the processor to execute the operations corresponding to the live content audit method according to any one of claims 1-8.

Images