Video auditing content pushing method and system

Abstract

The embodiment of the invention discloses a video auditing content pushing method and system. According to the embodiment of the invention, the method comprises the steps: firstly carrying out the decision analysis of an obtained early warning event flow through a frequency reduction strategy, and obtaining a frequency reduction decision analysis result; secondly, according to an underclocking decision analysis result, generating an auditing execution instruction of the video content; if the auditing execution instruction is an auditing suppression instruction, intercepting an early warning event stream for the video content; and if the auditing execution instruction is an auditing push instruction, executing an early warning auditing operation for the video content to obtain auditing operation feedback, and updating the frequency reduction strategy according to the auditing operation feedback. Therefore, a large amount of misjudged contents can be prevented from being pushed to a manual auditing queue, so that the video content auditing efficiency of the media platform is improved.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of data processing, in particular to a video audit content pushing method and system. BACKGROUND

[0002] In the current Internet era, information dissemination presents an exponential growth trend, and the content audit supervision of media platforms faces many severe challenges: the huge data volume makes the audit system need to ensure the real-time nature when dealing with massive content information.

[0003] To cope with the above challenges, the prior art usually uses a computer to preliminarily screen potential illegal content, and then pushes the preliminary screening results to the artificial audit queue. However, the preliminary screening results often contain a large amount of misjudgment content (judged as illegal by the computer but judged as non-illegal by artificial audit), and if this part of content is pushed to the artificial audit queue in large quantities, it will significantly affect the subsequent audit efficiency.

[0004] Therefore, how to efficiently, accurately and intelligently realize flexible potential illegal content pushing to improve the content audit efficiency of media platforms is a technical problem to be solved at present. SUMMARY

[0005] Embodiments of the present application provide a video audit content pushing method and system.

[0006] In a first aspect, the embodiments of the present application provide a video audit content pushing method applied to a video audit content pushing system, and the method comprises: obtaining an early warning event stream, the early warning event stream comprising a text event stream and an image event stream for the same video content, and the text event stream and the image event stream being time-aligned; performing decision analysis on the early warning event stream based on a frequency reduction strategy to obtain a frequency reduction decision analysis result, the frequency reduction strategy comprising a sensitive word frequency reduction strategy, a regional image frequency reduction strategy and an audit feedback frequency reduction strategy, and the frequency reduction decision analysis result comprising a text frequency reduction label, an image frequency reduction label and a global frequency reduction label; generating an audit execution instruction for the video content according to the frequency reduction decision analysis result; if the audit execution instruction is an audit suppression instruction, intercepting the early warning event stream for the video content; and if the audit execution instruction is an audit pushing instruction, performing an early warning audit operation for the video content.

[0007] Optionally, performing decision analysis on the early warning event stream based on the frequency reduction strategy to obtain the frequency reduction decision analysis result comprises: performing text dimension frequency reduction decision analysis on sensitive word content of the text event stream through the sensitive word frequency reduction strategy to generate a text dimension frequency reduction decision label.

[0008] Optionally, the sensitive word content of the text type event stream is subjected to text dimension frequency reduction decision analysis by the sensitive word frequency reduction strategy, and a text dimension frequency reduction decision label is generated, including: The OCR push event stream in the video content is subjected to real-time monitoring by the stream computing strategy, and a hit frequency distribution curve of the sensitive word content is generated; The historical review results of the sensitive word content in the video content are queried through the distributed cache system; It is judged whether the sensitive word content meets the transient controlled condition based on the hit frequency distribution curve and the historical review results, and if so, the sensitive word content is determined as a transient controlled frequency reduction sensitive word; wherein the transient controlled frequency reduction sensitive word is a sensitive word pushed according to a set limited frequency; The current push timestamp of the transient controlled frequency reduction sensitive word matching the user identifier is obtained, and the latest push timestamp of the transient controlled frequency reduction sensitive word matching the user identifier is queried, and a text dimension frequency reduction decision label is generated according to the current push timestamp and the latest push timestamp; the text dimension frequency reduction decision label includes an event push interception label and an event push release label.

[0009] Optionally, the pre-warning event stream is subjected to decision analysis based on the frequency reduction strategy, and a frequency reduction decision analysis result is obtained, including: The region feature vector of the image type event stream is subjected to image dimension frequency reduction decision analysis by the region image frequency reduction strategy, and an image dimension frequency reduction decision label is generated.

[0010] Optionally, the region feature vector of the image type event stream is subjected to image dimension frequency reduction decision analysis by the region image frequency reduction strategy, and an image dimension frequency reduction decision label is generated, including: If the feature similarity of the region feature vector and any real violation image feature vector in the first feature library exceeds a first similarity threshold, an image dimension frequency reduction decision label indicating push release is generated; If the average similarity of the region feature vector and the Top-K nearest neighbor feature vectors in the second feature library exceeds a second similarity threshold but does not exceed the first similarity threshold, an image dimension frequency reduction decision label for push interception is generated; the second feature library is used to store false violation image feature vectors, and the Top-K nearest neighbor feature vectors are the K feature vectors closest to the region feature vector in the feature space in the second feature library; If the feature similarity does not exceed the first similarity threshold, and the average similarity does not exceed the second similarity threshold, an image dimension frequency reduction decision label indicating default push is generated.

[0011] Optionally, the pre-warning event stream is subjected to decision analysis based on the frequency reduction strategy, and a frequency reduction decision analysis result is obtained, including: The global frequency reduction decision label is generated by auditing the feedback frequency reduction strategy and analyzing the global frequency reduction decision based on the historical auditing operation feedback of the early warning event stream.

[0012] Optionally, the global frequency reduction decision label is generated by auditing the feedback frequency reduction strategy and analyzing the global frequency reduction decision based on the historical auditing operation feedback of the early warning event stream, including: storing the to-be-processed push event into a to-be-confirmed area of a preset double-buffer event storage structure; the to-be-confirmed area is used to store a first event record of the video content for which an auditing push instruction has been generated but for which no auditing operation feedback has been received; the to-be-processed push event includes at least one of the following: a text dimension frequency reduction decision label indicating a push release of a text push event, and an image dimension frequency reduction decision label indicating a push release or indicating a default push of an image push event; migrating the to-be-processed push event to a confirmed area of the preset double-buffer event storage structure according to a residence duration of the to-be-processed push event in the to-be-confirmed area; the confirmed area is used to store a second event record of the video content for which an auditing operation feedback has been received; generating a global frequency reduction decision label for push interception of the to-be-processed push event based on a cumulative number of false violation labels of a historical event record in the confirmed area that is identical to the user identifier and the time feature abstract of the to-be-processed push event in a preset statistical period; if it is detected that the auditing result of the historical event record is adjusted to a real violation label within an effective period of the global frequency reduction state of the to-be-processed push event, the global frequency reduction state is released, and a global frequency reduction decision label indicating a push release of the to-be-processed push event is generated.

[0013] Optionally, the method further includes: obtaining an auditing operation feedback corresponding to the early warning auditing operation; updating the frequency reduction strategy according to the auditing operation feedback.

[0014] Optionally, the frequency reduction strategy is updated according to the auditing operation feedback, including: determining a historical misjudgment sensitive word content according to the auditing operation feedback, and adding the historical misjudgment sensitive word content to a preset adaptive word table; monitoring an entry table validity period of any sensitive word in the preset adaptive word table in real time, and in response to a time difference between a current time and an entry time of the any sensitive word exceeding the entry table validity period, counting a number of newly added negative feedbacks of the any sensitive word within the entry table validity period, the number of newly added negative feedbacks being a number of times that the any sensitive word is determined to be misjudged within the entry table validity period; if the number of newly added negative feedbacks is less than a preset number threshold, the any sensitive word is deleted from the preset adaptive word table, and a frequency reduction weight of the any sensitive word is recovered; If the number of newly added negative feedbacks is greater than or equal to the preset number threshold, the validity period of any sensitive word in the preset adaptive word table is extended, or the frequency reduction weight and the entering time of any sensitive word are updated.

[0015] Optionally, the frequency reduction strategy is updated according to the audit operation feedback, including: According to the violation determination result of the image type event stream in the audit operation feedback, the real violation image feature vector in the first feature library and the false violation image feature vector in the second feature library are updated respectively; the violation determination result includes the real violation determination result and the false violation determination result of the region feature vector of the image type event stream.

[0016] Optionally, according to the frequency reduction decision analysis result, an audit execution instruction of the video content is generated, including: If there is at least one decision label in the frequency reduction decision analysis result for indicating push interception, an audit suppression instruction is generated, and the audit suppression instruction includes at least one of a text interception identifier, an image interception identifier or a global interception identifier; If all decision labels in the frequency reduction decision analysis result are used to indicate push release, an audit push instruction is generated.

[0017] In a second aspect, the embodiments of the present application provide a video audit content push system, including: a processor; a storage system, which has a computer program stored thereon, When the computer program is executed by the processor, the processor implements any of the video audit content push methods.

[0018] The embodiments of the present application provide a readable storage medium, which has a program or instruction stored thereon, and the program or instruction is executed by a processor to implement the steps of the video audit content push method.

[0019] The embodiments of the present application first perform decision analysis on the obtained early warning event stream through the frequency reduction strategy to obtain a frequency reduction decision analysis result; secondly, according to the frequency reduction decision analysis result, an audit execution instruction of the video content is generated; if the audit execution instruction is an audit suppression instruction, the early warning event stream for the video content is intercepted; if the audit execution instruction is an audit push instruction, an early warning audit operation for the video content is performed to obtain an audit operation feedback, and then the frequency reduction strategy is updated according to the audit operation feedback. On the one hand, it can efficiently, accurately and intelligently realize flexible potential violation content push, and on the other hand, it can realize feedback closed-loop update of the frequency reduction strategy, and strengthen the dynamic perception and adaptive adjustment capability of the frequency reduction strategy. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1A flowchart of a video review content pushing method provided by an embodiment of the present application.

[0021] Figure 2 An input-output flow direction schematic diagram of a video review content pushing method provided by an embodiment of the present application.

[0022] Figure 3 A schematic diagram of a basic structure of a video review content pushing system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0023] In order to make the above objectives, characteristics and advantages of the present application more apparent, further detailed description will be made to the embodiments of the present application in combination with the drawings and specific implementation modes.

[0024] Reference is made to Figure 1 The figure is a flowchart of a video review content pushing method provided by an embodiment of the present application, and the method can be applied to a video review content pushing system. As shown in the figure, Figure 1 The method can include steps 110-142.

[0025] Step 110: Obtain a pre-warning event stream.

[0026] In the embodiment of the present application, the pre-warning event stream includes a text type event stream and an image type event stream for the same video content, and the text type event stream and the image type event stream are time-sequentially aligned.

[0027] Taking a popular fashion dress live broadcast as the video content for example, the fashion dress live broadcast lasts for several hours, during which the host will continuously show various fashion clothes and communicate with the audience. In order to review the video content, it is necessary to obtain the pre-warning event stream. The text type event stream is mainly derived from the real-time explanation of the host and the text information generated by the audience in the bullet screen, and these text information will be continuously generated and updated over time. The image type event stream is composed of live broadcast pictures, and each live broadcast picture frame can be processed as an image type event, and the image type event also continuously changes as the live broadcast progresses.

[0028] Further, the text type event stream and the image type event stream are time-sequentially aligned, in other words, at one time, the host explains a certain dress, and the corresponding live broadcast picture also shows the dress at the same time (that is, the text information and the image information correspond to each other in time).

[0029] As an exemplary implementation mode, the voice of the host can be converted into text, and the text type event stream can be generated in combination with the audience's speech in the bullet screen, and the image type event stream can be obtained by frame-by-frame interception of the live broadcast picture, so as to form a complete pre-warning event stream through time-sequential alignment processing.

[0030] In the embodiments of the present application, the acquisition of the text event stream and the image event stream is authorized by the relevant anchors and audiences. For example, before the anchor starts the live broadcast, the video review content pushing system (hereinafter referred to as the system) will explicitly inform the anchor that the video content will be reviewed, and the text information and image information will be collected and processed during the review process. The anchor agrees to the relevant agreement before starting the live broadcast. For the audience, when entering the live broadcast page, the system will also prompt the audience in a conspicuous manner that the audience's pop-up speech and viewing screen will be used for review purposes. The audience is deemed to have authorized the platform to obtain the relevant information in the process of knowing and continuing to participate in the live broadcast. Therefore, the text event stream and the image event stream can also be understood as the early warning event stream obtained by preliminary screening.

[0031] Step 120: performing decision analysis on the early warning event stream based on the frequency reduction strategy to obtain a frequency reduction decision analysis result.

[0032] After obtaining the early warning event stream of the fashion dressing live broadcast, the frequency reduction strategy can be used to perform decision analysis. In the embodiments of the present application, the frequency reduction strategy includes a sensitive word frequency reduction strategy, a regional image frequency reduction strategy, and an audit feedback frequency reduction strategy. Based on this, the frequency reduction decision analysis result includes a text frequency reduction label, an image frequency reduction label, and a global frequency reduction label.

[0033] The text frequency reduction label can be understood as a text dimension frequency reduction decision label corresponding to the sensitive word frequency reduction strategy, the image frequency reduction label can be understood as an image dimension frequency reduction decision label corresponding to the regional image frequency reduction strategy, and the global frequency reduction label can be understood as a global frequency reduction decision label corresponding to the audit feedback frequency reduction strategy.

[0034] It can be understood that the frequency reduction strategy comprehensively considers the information in multiple dimensions such as text, image, and human review feedback driving, as well as the dynamic changes of these information over time, aiming to reduce unnecessary review pushing and improve review efficiency.

[0035] Specifically, based on the above sensitive word frequency reduction strategy, regional image frequency reduction strategy, and audit feedback frequency reduction strategy, the early warning event stream can be evaluated from different angles to determine the risk degree and actual violation possibility of each event (text event / image event), so as to determine whether to perform frequency reduction processing, and finally obtain the frequency reduction decision analysis result. The frequency reduction decision analysis result can be used to generate an audit execution instruction.

[0036] It can be understood that before performing decision analysis on the early warning event stream based on the frequency reduction strategy to obtain the frequency reduction decision analysis result, the sensitive word content of the text event stream and the regional feature vector of the image event stream need to be extracted.

[0037] For the text event stream of fashion live streaming, sensitive word content needs to be extracted. The definition of sensitive word content is the word related to violation and sensitive information, such as some words related to false propaganda, exaggerated propaganda and inappropriate speech. During the live streaming process, the host may introduce the material and function of the clothing, and if some exaggerated efficacy description is mentioned, these related words may be regarded as sensitive words. The system will use natural language processing (NLP) technology to process the text event stream, such as word segmentation and part-of-speech tagging, and then compare it with the preset sensitive word library to extract the sensitive word content.

[0038] For the image event stream of fashion live streaming, the region feature vector (which can be understood as a Box-level high-risk feature vector) needs to be extracted. In the fashion live streaming picture, there may be overall display of clothing, local close-up and other different regions. The system will divide and identify these different regions, such as the front, back and detail parts of the clothing as different local image regions. Then, using image feature extraction algorithm, each local image region is processed to extract the corresponding region feature vector, which can include color, texture, shape and other information of the image.

[0039] As an optional embodiment, the decision analysis of the early warning event stream based on the frequency reduction strategy described in step 120 obtains a frequency reduction decision analysis result, including: Step 121: performing text dimension frequency reduction decision analysis on the sensitive word content of the text event stream through the sensitive word frequency reduction strategy, and generating a text dimension frequency reduction decision label.

[0040] After extracting the sensitive word content in the text event stream of fashion live streaming, the sensitive words need to be analyzed in the text dimension for frequency reduction decision, which will consider the hit frequency of sensitive words, historical audit results and other factors to determine whether the events containing these sensitive words need to be processed for frequency reduction, and generate the corresponding text dimension frequency reduction decision label.

[0041] In an optional embodiment, the sensitive word content of the text event stream is analyzed in the text dimension for frequency reduction decision through the sensitive word frequency reduction strategy described in step 121, and a text dimension frequency reduction decision label is generated, including: Step 1211: using stream computing strategy to listen to the OCR push event stream in the video content in real time, and generating a hit frequency distribution curve of sensitive word content.

[0042] In the embodiments of the present application, a streaming computing strategy is adopted to listen to the OCR push event stream in real time. The OCR (Optical Character Recognition) service identifies the text information in the live picture and pushes it in the form of an event stream. The push system of the video review content continuously listens to the OCR push event stream. Whenever a sensitive word is identified, the hit information of the sensitive word is recorded, including the hit time, the hit text content, etc. With the passage of time, a large amount of sensitive word hit data is collected.

[0043] Then, according to the above sensitive word hit data, a hit frequency distribution curve of the sensitive word content is generated. The hit frequency distribution curve takes time as the horizontal axis and the hit frequency of the sensitive word as the vertical axis, and can intuitively show the hit frequency change of the sensitive word in different time periods. By analyzing the hit frequency distribution curve, the peak and trough of the hit frequency of the sensitive word can be determined. For example, if the hit frequency of a sensitive word increases sharply in a short period of time, it may mean that there is an abnormal situation that needs further attention.

[0044] For convenience of distinction, the text event stream in step 110 and the OCR push event stream in step 1211 are described here.

[0045] In the embodiments of the present application, the source of the text event stream in step 110 is the text information generated by the real-time commentary (speech-to-text) of the host and the remarks of the audience in the bullet screen, that is, the core is the speech-to-text content (host commentary) and the bullet screen text input by the user. These text information directly comes from speech recognition (ASR) and user input, and does not need to be processed by image recognition technology.

[0046] The OCR push event stream in step 1211 is the text information in the live picture recognized by the OCR (Optical Character Recognition) service and pushed in the form of an event stream. Its source is the image content in the live picture (such as the product label displayed by the host, the screen caption, the background text, etc.), which needs to be extracted from the picture frame by image recognition technology (OCR).

[0047] Based on this, the content of the text event stream mainly includes: the commentary content of the host speech-to-text (such as “this coat uses top-grade fabric”); the bullet screen content sent by the audience in real time (such as “link please” and “color is beautiful”). The content of the text event stream is “pure text of human voice or user input”, which does not need to rely on images.

[0048] The content of the OCR push event stream is the text information carried by the image in the live picture (such as “material: 90% wool” on the product tag and “limited-time discount” on the background screen, etc.). The content of the OCR push event stream is hidden in the image and needs to be extracted from the picture frame by OCR technology before forming the event stream.

[0049] It is worth mentioning that although the text event stream and the OCR push event stream have different sources, they jointly constitute complete text review materials in the video content review scene. On the one hand, the text event stream covers the text risks of “voice interaction” and “user input” (such as inappropriate remarks of the host and sensitive content of the bullet screen); on the other hand, the OCR push event stream covers the risks of “text in the picture image” (such as false propaganda of product labels and illegal information of background subtitles). Therefore, the text event stream and the OCR push event stream are two different sub-streams of text review, which jointly supplement all possible text risk points in the video content.

[0050] Step 1212: Query the historical review result of the sensitive word content in the video content through the distributed cache system.

[0051] In the embodiment of the present application, the distributed cache system has efficient data storage and query capability, which can quickly obtain the required historical review result. The push system of the video review content will store the result of each previous review in the distributed cache system as the historical review result. When a certain sensitive word needs to be analyzed, the historical review result of the sensitive word can be obtained by querying the distributed cache system according to the related identifier of the sensitive word. These historical review results can reflect how the sensitive word performs in the past review and whether it often leads to violation judgment. For example, if a sensitive word has a few times of true submission and violation judgment but a lot of total hits in many past reviews, it may need to be processed for frequency reduction.

[0052] In the embodiment of the present application, the historical review result includes the user identifier corresponding to the sensitive word content, the target number of true submission and violation judgment, and the total hit number. In the embodiment of the present application, the user identifier is used to distinguish different hosts or audiences. Each user participating in the live broadcast has a unique identifier, and through this identifier, the hit situation of the sensitive word can be associated with a specific user. The target number of true submission and violation judgment refers to the number of times that are actually sent to artificial review and finally judged as violation in all hits of the sensitive word. The total hit number refers to the total number of times that the sensitive word is identified in the video content.

[0053] For example, in a live broadcast, the sensitive word “top quality” is hit many times. Through the query of the historical review result, it is found that the corresponding user identifier is host A, the target number of true submission and violation judgment is 2 times, and the total hit number is 10 times. These information can help the system to evaluate the actual violation probability of the sensitive word, so as to make more accurate frequency reduction decision.

[0054] Step 1213: judging whether the sensitive word content meets the transient controlled condition based on the hit number distribution curve and the historical review result, and if yes, determining the sensitive word content as a transient controlled frequency reduction sensitive word.

[0055] In the embodiments of the present application, the transient controlled condition includes a combination of the hit number sudden increase condition and the historical violation rate condition.

[0056] In detail, the historical violation rate can be determined according to the target number of real submissions and the total hit number of the queried sensitive word. The historical violation rate is an important indicator for measuring the possibility of violation of a sensitive word in historical review. Specifically, the target number is divided by the total hit number, and the result is the historical violation rate of the sensitive word. For example, for the sensitive word "top quality", the target number is 2 times and the total hit number is 10 times, and the historical violation rate is 2 / 10, that is, the possibility of violation of the sensitive word in historical review is a certain percentage. By calculating the historical violation rate, the risk level of the sensitive word can be quantitatively evaluated.

[0057] On the basis of the above, if the slope value of the hit number distribution curve in the preset time window exceeds the preset rising threshold and the historical violation rate is lower than the preset violation rate, the sensitive word content is determined as a transient controlled frequency reduction sensitive word, and the transient controlled frequency reduction sensitive word is a sensitive word that is pushed according to a set limit frequency.

[0058] In the embodiments of the present application, for the generated hit number distribution curve, a preset time window and a preset rising threshold are set. The preset time window is an adjustable time period, for example, a certain time interval after the start of live broadcast. The preset rising threshold is a pre-set slope standard for judging whether the rising speed of the hit number of the sensitive word is too fast.

[0059] If the slope value of the hit number distribution curve of the sensitive word exceeds the preset rising threshold in the preset time window, it means that the hit number of the sensitive word increases sharply in a short time. At the same time, the historical violation rate of the sensitive word is also considered. If the historical violation rate is lower than the preset violation rate, it means that the sensitive word has a small possibility of actual violation although the hit number increases. In this case, the sensitive word content is determined as a transient controlled frequency reduction sensitive word.

[0060] For example, the sensitive word "luxury fabric" has a hit number distribution curve with a slope exceeding the preset rising threshold in the preset time window, and its historical violation rate is lower than the preset violation rate. Therefore, it will be determined as a transient controlled frequency reduction sensitive word, and subsequent events containing the sensitive word will be subjected to frequency reduction processing.

[0061] Step 1214: Obtain the current push timestamp of the transient controlled frequency reduction sensitive word matching user identifier, and query the latest push timestamp of the transient controlled frequency reduction sensitive word matching user identifier, and generate a text dimension frequency reduction decision label according to the current push timestamp and the latest push timestamp; the text dimension frequency reduction decision label includes an event push interception label and an event push release label.

[0062] In the embodiments of the present application, in response to the current push timestamp of the transient controlled frequency reduction sensitive word matching user identifier, the latest push timestamp of the transient controlled frequency reduction sensitive word matching user identifier in the distributed cache system is queried; if the time interval between the latest push timestamp and the current push timestamp is less than the preset cooling interval, a text dimension frequency reduction decision label for push interception is generated; if the time interval is greater than or equal to the preset cooling interval, a text dimension frequency reduction decision label for indicating frequency reduction push release is generated.

[0063] In the embodiments of the present application, when a transient controlled frequency reduction sensitive word triggers a push, the push system of the video review content will record the current push timestamp of the sensitive word matching user identifier. Then, by querying the distributed cache system, the latest push timestamp of the same user identifier matching the transient controlled frequency reduction sensitive word is obtained. In addition, the preset cooling interval is a pre-set time interval for controlling the push frequency of the sensitive word.

[0064] It can be understood that if the time interval between the latest push timestamp and the current push timestamp is less than the preset cooling interval, it means that the sensitive word has been pushed in a short time, in order to avoid excessive push, a text dimension frequency reduction decision label for push interception is generated, that is, the event containing the sensitive word is not pushed to the artificial review. For example, for the transient controlled frequency reduction sensitive word "top quality" used by the host A, the current push timestamp is a certain time, and the latest push timestamp is close to the current time, which is less than the preset cooling interval, so the push interception label is generated.

[0065] On the contrary, if the time interval is greater than or equal to the preset cooling interval, it means that enough time has passed since the last push, so the frequency reduction push can be performed, and a text dimension frequency reduction decision label for indicating frequency reduction push release is generated. For example, the text dimension frequency reduction decision label for indicating frequency reduction push release can be: {top quality, host A, push once every X minutes}.

[0066] As another optional embodiment, the step 120 described based on the frequency reduction strategy to analyze the pre-warning event stream to obtain the frequency reduction decision analysis result, including: Step 122: performing image dimension frequency reduction decision analysis on the region feature vector of the image type event stream through the region image frequency reduction strategy, and generating an image dimension frequency reduction decision label.

[0067] After the regional feature vectors of the fashion dress live broadcast image class event stream are extracted, a regional image frequency reduction strategy needs to be combined to perform frequency reduction decision analysis on the regional feature vectors from the image dimension to generate image dimension frequency reduction decision labels. The process mainly judges whether the image region has a violation risk and whether it needs to be reduced in frequency.

[0068] As a preferred embodiment, in step 122, the regional feature vectors of the image class event stream are subjected to frequency reduction decision analysis from the image dimension by the regional image frequency reduction strategy to generate image dimension frequency reduction decision labels, which includes: Step 1221: If the feature similarity between the regional feature vector and any real violation image feature vector in the first feature library exceeds the first similarity threshold, an image dimension frequency reduction decision label indicating push release is generated.

[0069] In the embodiments of the present application, the first feature library stores real violation image feature vectors confirmed by manual audit. For the extracted regional feature vector, similarity comparison is performed with the real violation image feature vectors in the first feature library (for example, by cosine similarity calculation or Euclidean distance calculation, etc.). If the feature similarity between the regional feature vector and any real violation image feature vector in the first feature library exceeds the first similarity threshold, it means that the image region corresponding to the regional feature vector is likely to have a violation risk. In this case, an image dimension frequency reduction decision label indicating push release is generated, that is, the image region needs to be pushed to manual audit for further confirmation of its violation situation. For example, a local region of a piece of clothing displayed in a live broadcast picture has a regional feature vector whose similarity with the feature vector of a certain piece of violation clothing in the first feature library exceeds the first similarity threshold, and a push release image dimension frequency reduction decision label is generated.

[0070] Step 1222: If the average similarity between the regional feature vector and the Top-K nearest neighbor feature vectors in the second feature library exceeds the second similarity threshold but does not exceed the first similarity threshold, an image dimension frequency reduction decision label for push interception is generated.

[0071] In the embodiments of the present application, the second feature library is used to store false violation image feature vectors, and the Top-K nearest neighbor feature vectors are the K feature vectors closest to the region feature vector in the feature space in the second feature library. When analyzing the extracted region feature vector, the Top-K nearest neighbor feature vectors of the region feature vector are searched in the second feature library, and the average similarity is calculated. If the average similarity exceeds the second similarity threshold but does not exceed the first similarity threshold, it is determined that the image region corresponding to the region feature vector has a certain similarity with some known false violation images, but it cannot be determined as a real violation. In order to reduce unnecessary review push, an image dimension reduction decision label for push interception is generated, that is, the image region is not reported as the main basis for manual review. For example, the region feature vector of a clothing pattern displayed in a live picture exceeds the second similarity threshold but does not exceed the first similarity threshold with the average similarity of the Top-K nearest neighbor feature vectors in the second feature library, and a push interception label is generated.

[0072] In the embodiments of the present application, the second feature library is used to store those image feature vectors confirmed by manual review as non-violation or low risk, but marked as high risk in AI early warning. These false violation image feature vectors are obtained through a large amount of manual review and data accumulation. For example, some normal wearing postures of clothes and specific light and shadow effects displayed in a live picture may be misjudged as violations by the AI model, but confirmed as non-violations by manual review. The feature vectors of these images are stored in the second feature library. By establishing and maintaining the second feature library, known false violation patterns can be identified and intercepted, and unnecessary review push can be reduced.

[0073] Step 1223: If the feature similarity does not exceed the first similarity threshold, and the average similarity does not exceed the second similarity threshold, an image dimension reduction decision label for indicating default push is generated.

[0074] It can be understood that if the feature similarity of the extracted region feature vector with the real violation image feature vectors in the first feature library does not exceed the first similarity threshold, and the average similarity with the Top-K nearest neighbor feature vectors in the second feature library also does not exceed the second similarity threshold, it is determined that the image region corresponding to the region feature vector has neither obvious violation features nor similarity with known false violation patterns. In this case, an image dimension reduction decision label for indicating default push is generated, that is, the image region is continued to be used for reporting manual review according to the default strategy, so as to ensure that no violation may be missed. For example, a normal clothing detail region displayed in a live picture has no significant similarity hit in the two feature libraries, and a default push label is generated.

[0075] As another optional embodiment, the step 120 describes the decision analysis of the early warning event stream based on the frequency reduction strategy, and the frequency reduction decision analysis result includes: Step 123: Feedback frequency reduction strategy by auditing, and global frequency reduction decision analysis is performed by combining the historical auditing operation feedback of the early warning event stream to generate a global frequency reduction decision label.

[0076] In the embodiments of the present application, in addition to the frequency reduction decision analysis from the text dimension and the image dimension, global frequency reduction decision analysis based on the historical auditing operation feedback of the early warning event stream is also needed. The historical auditing operation feedback contains the results and related information of previous audits. Through the analysis of these feedback information, the risk degree of the event and whether it needs to be processed by frequency reduction can be judged from a more macro perspective, and finally a global frequency reduction decision label is generated.

[0077] In the embodiments of the present application, the auditing feedback frequency reduction strategy can also be understood as a dynamic rule strategy driven by human review feedback, and the dynamic rule strategy can be realized based on the real-time adjustment of the double-flow event (early warning event stream and human review submission event stream) and the state machine.

[0078] In one implementation, the step 123 described that the global frequency reduction decision analysis is performed by feedback frequency reduction strategy through auditing, and combining the historical auditing operation feedback of the early warning event stream to generate a global frequency reduction decision label, includes: Step 1231: Store the to-be-processed push event into the to-be-confirmed area of the preset double-buffer event storage structure.

[0079] In this step, the preset double-buffer event storage structure (double-buffer state machine) includes a to-be-confirmed area and a confirmed area. The to-be-confirmed area is used to store the first event record of the video content which has generated an auditing push instruction but has not received an auditing operation feedback. Each first event record includes a user identifier, an event feature abstract, and a push timestamp. The to-be-processed push event includes a text type push event indicated by the text dimension frequency reduction decision label to push and release, and an image type push event indicated by the image dimension frequency reduction decision label to push and release or indicated by the default push. When these to-be-processed push events are generated, they will be stored in the to-be-confirmed area.

[0080] For example, for the text type push event mentioned by the anchor A which contains a sensitive word and is indicated by the text dimension frequency reduction decision label to push and release, and the image type push event indicated by the image dimension frequency reduction decision label to push and release which is generated by the region feature vector in the live picture, they will be stored in the to-be-confirmed area, waiting for the feedback of the manual auditing result.

[0081] Step 1232: According to the residence time length of the to-be-processed push event in the to-be-confirmed area, the to-be-processed push event is migrated to the confirmed area of the preset double-buffer event storage structure.

[0082] For example, in response to the length of time that the to-be-processed push event stays in the to-be-confirmed area exceeding a preset length of time, a potential false violation label is set for the to-be-processed push event and the to-be-processed push event is migrated to the confirmed area of the preset double-buffered event storage structure.

[0083] In this step, the preset length of time is a pre-set time standard. If the length of time that the to-be-processed push event stays in the to-be-confirmed area exceeds the preset length of time, it indicates that the event has not received feedback from manual review for a long time. Based on the logic of "high probability of non-urgent high risk due to timeout", a potential false violation label is set for the to-be-processed push event. Then, the to-be-processed push event is migrated from the to-be-confirmed area to the confirmed area of the preset double-buffered event storage structure. The confirmed area is used to store second event records of video content that have received feedback from review operations, and each second event record contains a user identifier, an event feature abstract, a review result, and a review timestamp. The review result includes a true violation label or a false violation label. For example, a to-be-processed push event of an image type stays in the to-be-confirmed area for more than a preset length of time, and is set with a potential false violation label and migrated to the confirmed area.

[0084] Step 1233: Based on the cumulative number of false violation labels of historical event records in the confirmed area that are the same as the user identifier and the time feature abstract of the to-be-processed push event in a preset statistical period, a global frequency reduction decision label for pushing and intercepting the to-be-processed push event is generated.

[0085] In detail, the cumulative number of false violation labels of historical event records in the confirmed area that are the same as the user identifier and the time feature abstract of the to-be-processed push event in a preset statistical period is counted. If the cumulative number of false violation labels reaches a set cumulative number, a global frequency reduction state for the user identifier and the time feature abstract of the to-be-processed push event is triggered, and a global frequency reduction decision label for indicating pushing and intercepting the to-be-processed push event is generated.

[0086] In this step, the cumulative number of false violation labels of historical event records in the confirmed area that are the same as the user identifier and the time feature abstract of the to-be-processed push event in a preset statistical period is counted. The preset statistical period can be a certain time range after the start of live streaming, and the set cumulative number is a pre-set number standard.

[0087] If the accumulated number of false violation labels reaches the set accumulated number, it indicates that the event corresponding to the user identifier and time feature summary is determined to be a false violation multiple times in the statistical period, and there is a high possibility of misjudgment. In this case, the global frequency reduction state for the user identifier and time feature summary of the to-be-processed push event is triggered, and a global frequency reduction decision label indicating push interception of the to-be-processed push event is generated, that is, the event will no longer be pushed to manual review. For example, for the events of anchor A in a certain time period, if the accumulated number of false violation labels reaches the set accumulated number, the global frequency reduction state is triggered, and a label indicating push interception is generated.

[0088] Step 1234: If the audit result of the historical event record is adjusted to the real violation label within the effective period of the global frequency reduction state of the to-be-processed push event, the global frequency reduction state is released, and a global frequency reduction decision label indicating the push release of the to-be-processed push event is generated.

[0089] In the fashion dress live broadcast scene, when in the effective period of the global frequency reduction state, the system will continue to monitor the audit result of the historical event record. Once it is detected that the audit result of the historical event record previously marked as a false violation is adjusted to a real violation label, it indicates that such events actually have violation risks. At this time, in order to ensure the recall ability of real risks, the system will release the global frequency reduction state for the user identifier and time feature summary of the to-be-processed push event. At the same time, a global frequency reduction decision label indicating the push release of the to-be-processed push event is generated, that is, the event is restored to be pushed to manual review, so as to timely discover and handle the violation content. For example, during the effective period of the global frequency reduction state, it is found that the video content of anchor A in a certain time period previously determined to be a false violation actually has a violation, and the global frequency reduction state is released and a label indicating push release is generated.

[0090] It can be understood that after the text dimension frequency reduction decision label, the image dimension frequency reduction decision label, and the global frequency reduction decision label are determined, a frequency reduction decision analysis result can be generated based on the text dimension frequency reduction decision label, the image dimension frequency reduction decision label, and the global frequency reduction decision label.

[0091] In the embodiments of the present application, the text dimension frequency reduction decision label, the image dimension frequency reduction decision label, and the global frequency reduction decision label are comprehensively considered to generate a frequency reduction decision analysis result. The labels of the three dimensions respectively reflect the risk situation of the video content and whether it needs to be processed from different aspects. By integrating and analyzing these labels, a comprehensive frequency reduction decision can be obtained.

[0092] For example, if the text dimension de-rating decision label indicates push interception, the image dimension de-rating decision label also indicates push interception, and the global de-rating decision label likewise indicates that there is at least one of push interception, the de-rating decision analysis result will tend to push interception for the related event of the video content; conversely, if the labels of the above three dimensions all indicate push release, then the de-rating decision analysis result will allow the push of the related event.

[0093] Step 130: generating an audit execution instruction of the video content according to the de-rating decision analysis result.

[0094] After obtaining the de-rating decision analysis result of the fashion dress live broadcast, an audit execution instruction of the video content needs to be generated according to the result. The audit execution instruction is a key basis for deciding whether to push the video content to manual audit, which directly affects the efficiency and accuracy of the audit process.

[0095] In an alternative embodiment, the step 130 of generating an audit execution instruction of the video content according to the de-rating decision analysis result includes: Step 131: if there is at least one decision label in the de-rating decision analysis result for indicating push interception, generating an audit suppression instruction, the audit suppression instruction including at least one of a text interception identifier, an image interception identifier or a global interception identifier.

[0096] In this step, if there is at least one of the text dimension de-rating decision label, the image dimension de-rating decision label or the global de-rating decision label in the de-rating decision analysis result indicating push interception, an audit suppression instruction will be generated. The audit suppression instruction will contain at least one of the text interception identifier, the image interception identifier or the global interception identifier according to the specific situation.

[0097] For example, if the text dimension de-rating decision label indicates push interception, the audit suppression instruction will contain the text interception identifier, meaning that the text event containing the related sensitive word will not be pushed to manual audit; if the image dimension de-rating decision label indicates push interception, the audit suppression instruction will contain the image interception identifier, i.e. the related image area will not be reported to manual audit as the main basis; if the global de-rating decision label indicates push interception, the audit suppression instruction will contain the global interception identifier, and the related event of the entire video content will be pushed to interception.

[0098] Step 132: if all decision labels in the de-rating decision analysis result are used to indicate push release, generating an audit push instruction.

[0099] In this step, when the text dimension down frequency decision label, the image dimension down frequency decision label and the global down frequency decision label in the down frequency decision analysis result all indicate push release, it indicates that there is no obvious risk prompt or the video content meets the push requirements in each dimension. At this time, an audit push instruction is generated to push the related video content event to artificial audit to further confirm whether there is a violation. For example, the text description of the costume introduced by the host in the live broadcast, the image displayed and the three dimension decision labels corresponding to the overall live broadcast situation all indicate push release, and then the audit push instruction is generated to push these contents to artificial audit.

[0100] Step 141: If the audit execution instruction is an audit suppression instruction, intercept the pre-warning event stream of the video content.

[0101] In the embodiment of the present application, if the generated audit execution instruction is an audit suppression instruction, the video audit content pushing system will perform a push interception operation on the video content, including intercepting the pre-warning event stream of the video content, or reducing the push frequency of the pre-warning event stream of the video content, etc.

[0102] Specifically, if the audit suppression instruction contains a text interception identifier, the video audit content pushing system will prevent the text event containing the related sensitive words from entering the artificial audit queue; if it contains an image interception identifier, it will not report the related image area to artificial audit as the main basis; if it contains a global interception identifier, it will comprehensively suppress the push of the video content related event. For example, when the audit suppression instruction contains a text interception identifier, the text information related to some sensitive words mentioned by the host will not be pushed to artificial audit, thereby reducing unnecessary audit workload.

[0103] Step 142: If the audit execution instruction is an audit push instruction, perform a pre-warning audit operation on the video content.

[0104] In the embodiment of the present application, when the audit execution instruction is an audit push instruction, the video audit content pushing system will push the related video content event to artificial audit for pre-warning audit operation. Such design can efficiently, accurately and intelligently realize flexible potential violation content pushing, avoid a large number of misjudgment contents being pushed to the artificial audit queue, and thus improve the video content audit efficiency of the media platform.

[0105] On the basis of step 142, it further includes: Step 143: Obtain an audit operation feedback corresponding to the pre-warning audit operation; and update the down frequency strategy according to the audit operation feedback.

[0106] Further, human auditors will carefully review the video content to determine whether there are violations and give the audit results, that is, audit operation feedback. Audit operation feedback can be divided into two cases: real violations and false violations. According to this audit operation feedback, the frequency reduction strategy will be updated to improve the accuracy and adaptability of the strategy. For example, if the audit operation feedback indicates that there is a violation of a certain sensitive word, the relevant text dimension frequency reduction strategy will be adjusted; if the feedback indicates that a certain image region is a false violation, the feature library and frequency reduction strategy of the image dimension will be updated.

[0107] As an alternative embodiment, based on steps 1211-1214, the updating of the frequency reduction strategy according to the audit operation feedback in step 143 includes: Step 14211: determining the historical misjudgment sensitive word content according to the audit operation feedback, and adding the historical misjudgment sensitive word content to the preset adaptive word table.

[0108] In the embodiments of the present application, the historical misjudgment sensitive word content can be determined according to the audit operation feedback given by human auditors. These sensitive words may have been incorrectly determined as high-risk in previous audits, but in fact they did not cause violations. These historical misjudgment sensitive word contents are added to the preset adaptive word table. For example, if it is found in the audit operation feedback that the sensitive word "fashion trend" was misjudged before, it will be added to the preset adaptive word table, and subsequent events containing this sensitive word will be subject to more stringent frequency reduction processing.

[0109] In the embodiments of the present application, the frequency reduction weight of the sensitive word in the preset adaptive word table is greater than the frequency reduction weight of the transient controlled frequency reduction sensitive word. Illustratively, the role of the preset adaptive word table is to more strictly manage those verified historical misjudgment sensitive words. Since these sensitive words have been proven to have a lower probability of causing violations in actual applications, they are given greater frequency reduction weight.

[0110] Compared with the transient controlled frequency reduction sensitive word, the sensitive word in the preset adaptive word table will be subject to more stringent frequency reduction restrictions, thereby further reducing unnecessary audit push. For example, the transient controlled frequency reduction sensitive word may still have a small amount of push within a certain time interval, while the push frequency of the sensitive word in the preset adaptive word table will be lower.

[0111] Step 14212: monitoring the entry table validity period of any sensitive word in the preset adaptive word table in real time, and in response to the time difference between the current time and the entry time of any sensitive word exceeding the entry table validity period, counting the number of new negative feedbacks of any sensitive word within the entry table validity period. The number of new negative feedbacks refers to the number of times the sensitive word is determined to be misjudged within the entry table validity period.

[0112] In the embodiments of the present application, the entry table validity period of each sensitive word in the preset adaptive word table is monitored in real time. The entry table validity period is a preset time period for specifying the valid time of the sensitive word in the preset adaptive word table. When the time difference between the current time and the entry time of a sensitive word exceeds the entry table validity period, the video review content pushing system will count the number of newly added negative feedbacks of the sensitive word within the entry table validity period. The number of newly added negative feedbacks refers to the number of times that the sensitive word is determined to still have problems or cause misjudgments within the entry table validity period. For example, for the sensitive word "personal matching", the number of times that it is determined to possibly have misjudgments or inaccuracies is counted after the end of its entry table validity period.

[0113] Step 14213: If the number of newly added negative feedbacks is less than the preset number threshold, any sensitive word is deleted from the preset adaptive word table, and the frequency reduction weight recovery is performed on any sensitive word.

[0114] It can be understood that if the number of newly added negative feedbacks of a sensitive word in the preset adaptive word table obtained by counting is less than the preset number threshold, it indicates that the sensitive word performs well within the entry table validity period and may no longer need special frequency reduction processing. At this time, the video review content pushing system deletes the sensitive word from the preset adaptive word table and performs frequency reduction weight recovery on it. Frequency reduction weight recovery means that the sensitive word will no longer enjoy the strict frequency reduction strategy in the preset adaptive word table, but will be restored to the normal frequency reduction processing mode. For example, the sensitive word "unique style" has a number of newly added negative feedbacks within the entry table validity period that is less than the preset number threshold, so it is deleted from the preset adaptive word table and the frequency reduction weight is restored.

[0115] Step 14214: If the number of newly added negative feedbacks is greater than or equal to the preset number threshold, the validity period of any sensitive word in the preset adaptive word table is extended, or the frequency reduction weight and entry time of any sensitive word are updated.

[0116] In this step, when the number of newly added negative feedbacks of a sensitive word in the preset adaptive word table is greater than or equal to the preset number threshold, it indicates that the sensitive word still has a high risk of misjudgment or a high possibility of causing problems. In order to continue to strictly manage it, the video review content pushing system will take corresponding measures. It can choose to extend the validity period of the sensitive word in the preset adaptive word table, so that it continues to be strictly limited in frequency reduction; or update the frequency reduction weight and entry time of the sensitive word, adjust its frequency reduction strategy, and better adapt to its actual situation. For example, the sensitive word "luxury fashion" has a number of newly added negative feedbacks within the entry table validity period that is greater than or equal to the preset number threshold, so its validity period in the preset adaptive word table is extended or its frequency reduction weight and entry time are updated.

[0117] As another alternative embodiment, on the basis of steps 1221-1223, the update of the downshifting strategy in step 143 according to the audit operation feedback includes: Step 14221: According to the violation determination result of the image type event stream in the audit operation feedback, update the real violation image feature vector in the first feature library and the false violation image feature vector in the second feature library respectively; the violation determination result includes the real violation determination result and the false violation determination result of the region feature vector of the image type event stream.

[0118] On the one hand, in the case that the violation determination result of the image type event stream in the audit operation feedback indicates that the region feature vector is a real violation, if the third feature similarity between the region feature vector and the first false violation image feature vector in the second feature library exceeds the first maintenance similarity, the first false violation image feature vector is weakened.

[0119] In detail, when the audit operation feedback indicates that a certain region feature vector is a real violation, the third feature similarity between the region feature vector and the first false violation image feature vector in the second feature library is further analyzed. If the third feature similarity exceeds the first maintenance similarity, it means that the previous determination of the first false violation image feature vector as a false violation may be inaccurate. At this time, in order to ensure the accuracy of the feature library, the first false violation image feature vector is weakened, which can be to reduce its weight in the second feature library or delete it from the second feature library, so that it is no longer easily hit in subsequent matching. For example, the feature vector of a certain clothing region in the live picture is determined to be a real violation, and the third feature similarity with a false violation image feature vector in the second feature library exceeds the first maintenance similarity, then the false violation image feature vector is weakened.

[0120] If the fourth feature similarity between the region feature vector and the first real violation image feature vector in the first feature library exceeds the second maintenance similarity, the first real violation image feature vector is enhanced, and if the fourth feature similarity does not exceed the second maintenance similarity, the region feature vector is added to the first feature library.

[0121] In detail, when the audit operation feedback indicates that a region feature vector is a true violation, a fourth feature similarity is obtained by comparing the region feature vector with a first true violation image feature vector in the first feature library. If the fourth feature similarity exceeds a second maintenance similarity, it indicates that the region feature vector has a high similarity with a known true violation feature. In order to enhance the recognition ability of the true violation feature, the first true violation image feature vector is enhanced, which can be increasing the weight of the first true violation image feature vector in the first feature library or updating the time stamp of the first true violation image feature vector. If the fourth feature similarity does not exceed the second maintenance similarity, it indicates that the region feature vector is a new true violation feature, and the region feature vector is added to the first feature library to expand the coverage of the first feature library.

[0122] For example, a region feature vector of a certain clothing pattern in a live picture is determined as a true violation, and the similarity between the region feature vector and a certain true violation image feature vector in the first feature library exceeds the second maintenance similarity. The true violation image feature vector is enhanced. If the similarity does not exceed the second maintenance similarity, the region feature vector is added to the first feature library.

[0123] On the other hand, in the case that the audit operation feedback indicates that a region feature vector of a picture class event flow is a false violation, if a fifth feature similarity between the region feature vector and a second false violation image feature vector in the second feature library exceeds a third maintenance similarity, the second false violation image feature vector is enhanced. If the fifth feature similarity does not exceed the third maintenance similarity, the region feature vector is added to the second feature library.

[0124] In detail, when the audit operation feedback indicates that a region feature vector is a false violation, a fifth feature similarity is obtained by comparing the region feature vector with a second false violation image feature vector in the second feature library. If the fifth feature similarity exceeds a third maintenance similarity, it indicates that the region feature vector has a high similarity with a known false violation feature. In order to better identify and intercept the false violation mode, the second false violation image feature vector is enhanced, which can be increasing the inhibition weight of the second false violation image feature vector in the second feature library or updating the time stamp of the second false violation image feature vector. If the fifth feature similarity does not exceed the third maintenance similarity, it indicates that the region feature vector is a new false violation feature, and the region feature vector is added to the second feature library to enrich the content of the second feature library.

[0125] For example, a region feature vector of a certain clothing pattern in a live picture is determined as a true violation, and the similarity between the region feature vector and a certain true violation image feature vector in the first feature library exceeds the second maintenance similarity. The true violation image feature vector is enhanced. If the similarity does not exceed the second maintenance similarity, the region feature vector is added to the first feature library.

[0126] If the sixth feature similarity of the region feature vector to the second real violation image feature vector in the first feature library exceeds the fourth maintenance similarity, and the storage time of the second real violation image feature vector exceeds the preset aging time, the second real violation image feature vector is subjected to weakening processing.

[0127] In detail, when analyzing a region feature vector, if the sixth feature similarity of the region feature vector to the second real violation image feature vector in the first feature library exceeds the fourth maintenance similarity, and the storage time of the second real violation image feature vector exceeds the preset aging time, it indicates that the second real violation image feature vector may no longer have high representativeness due to long time or content change. In order to ensure the effectiveness and accuracy of the first feature library, the second real violation image feature vector is subjected to weakening processing, which can be reducing its weight in the first feature library or moving it to an observation list, etc. For example, if the sixth feature similarity of a region feature vector of an old-fashioned dress in a live picture to a real violation image feature vector in the first feature library with a long storage time exceeds the fourth maintenance similarity, the real violation image feature vector is subjected to weakening processing.

[0128] In actual application process, the first feature library and the second feature library can also be periodically simplified and managed in life cycle. In detail, in order to prevent the feature library from expanding indefinitely, resulting in decreased query efficiency and wasted storage resources, a rolling deletion mechanism based on time and access frequency can be used.

[0129] For the second feature library (“false high-risk” feature library), the features in the last N days (for example, 7 days) can be retained, or cold data can be eliminated based on access frequency, and features exceeding the period or with low frequency are automatically deleted or archived. Specifically, a time threshold and an access frequency threshold can be set, and each feature vector in the second feature library is checked. If the creation time of the feature vector exceeds N days or the access frequency in a certain period of time is lower than the access frequency threshold, the feature vector is considered to be a feature exceeding the period or with low frequency, and is subjected to automatic deletion or archiving processing. In this way, it can be ensured that the second feature library always stores relatively new and active feature vectors, and the accuracy and efficiency of feature matching are improved.

[0130] For the first feature library ("high-risk" feature library), the retention time is relatively long, for example, M days (M>N, for example, 30 days), or a relatively complex weight decay mechanism is adopted to ensure the recall ability of long-term high-risk content. For the weight decay mechanism, an initial weight can be assigned to each feature vector according to the storage time and historical hit times of the feature vector. With the passage of time, these weights are adjusted periodically. For those feature vectors with long storage time and gradually decreasing hit times, their weights are reduced. When the weight is lower than a certain threshold, special processing can be considered, such as moving to the observation list or deleting it. For some feature vectors that continue to hit and have a recent time, their weights are appropriately increased. In this way, the first feature library can retain important high-risk features and adapt to the dynamic changes of video content during long-term operation.

[0131] As can be seen, the embodiments of the present application combine the frequency reduction strategy and its feedback update processing to form a closed-loop review mechanism for the early warning event stream. This closed-loop review mechanism can effectively reduce the interference of "false positive" content on the manual review process, improve the accuracy and efficiency of review decisions, and realize intelligent, dynamic and fine-grained management of high-risk content push frequency, thereby building a more robust and timely video content security protection system. In actual application, through the frequency reduction processing of sensitive words and the dynamic maintenance of image feature library, unnecessary review push can be reduced, and real violation content can be discovered and processed in time, thereby ensuring the compliance and safety of video content.

[0132] Please refer to Figure 1 and Figure 2 The multi-dimensional intelligent dynamic frequency reduction processing scheme provided by the embodiments of the present application deeply combines the business scenario characteristics and massive historical data of live streaming platform content review, innovatively optimizes the key bottlenecks in the review link, and forms a complete and efficient intelligent content management solution.

[0133] Firstly, the embodiments of the present application greatly enhance the stability of the video content review system. The traditional video content review system (platform) often lacks an effective processing mechanism for early warning event streams. When a large number of early warning events flow in at the same time, the system is prone to insufficient processing capacity, slow response speed, and even collapse. The time-aligned text event stream and image event stream obtained by the embodiments of the present application can integrate various types of early warning information of video content comprehensively and orderly. Based on the decision analysis of the frequency reduction strategy, the system is prevented from being overwhelmed by a large number of redundant early warning events, effectively reducing the processing burden of the system. In this way, the system can still maintain a stable running state when facing a large number of early warning events, improving the reliability of the system and reducing the risk of review errors and delays caused by system overload.

[0134] Secondly, the embodiment of the present application can accurately identify the video content that really needs to be audited by performing multi-dimensional dynamic frequency reduction decision analysis on the early warning event stream, and effectively filters and suppresses the "false high-risk" early warning events. When generating an audit execution instruction, if it is judged as an audit suppression instruction, the push interception operation is performed to avoid the generation of invalid audit work orders, so that the audit personnel can focus more energy on the video content with high risk. When it is judged as an audit push instruction, the early warning audit operation is performed in time to improve the timeliness of the audit. At the same time, the frequency reduction strategy is updated according to the audit operation feedback, so that the strategy can be continuously optimized, further improving the efficiency of the audit. This dynamic audit mechanism can quickly respond to changes in video content, reduce the time cost of the audit, improve the processing amount of the audit per unit of time, and make the entire audit process more efficient.

[0135] In addition, in the traditional video content audit mode, the audit personnel often need to spend a lot of time and energy to process a large number of early warning events, among which there are some invalid audit work orders due to the misjudgment of the AI model. However, the application embodiment effectively reduces the "false high-risk" false positives by applying the frequency reduction strategy, so that the artificial audit resources can be more accurately invested in the audit of real high-risk content, not only improving the effectiveness of artificial audit, but also reducing the work burden of the audit personnel, so that the human resources are more reasonably configured.

[0136] Furthermore, the embodiment of the present application updates the frequency reduction strategy based on real-time audit operation feedback, so that the system can adapt to the continuous changes of the video content ecosystem, the emergence of new risk patterns, and the iterative evolution of the AI model. The type and risk pattern of the video content are constantly changing, and the traditional audit technology often cannot quickly adapt to these changes. However, the embodiment of the present application can dynamically adjust and optimize the frequency control strategy by learning historical data and real-time feedback, automatically identify and respond to new risk situations, and this adaptability and intelligent level reduces the dependence on artificial rule maintenance, so that the system can make more autonomous audit decisions, improving the flexibility and adaptability of the system.

[0137] Finally, the frequency reduction strategy of the embodiment of the present application supports flexible configuration and dynamic optimization of frequency control strategies of different dimensions (users, content types, sensitive words, feature patterns, etc.) and different granularities (second-level, minute-level, hour-level). In the field of Internet video content safety, different video contents may have different risk characteristics and control requirements. The technical solution of the embodiment of the present application can set individualized frequency control strategies for different dimensions and granularities according to specific needs, realizing fine-grained control of video content. This flexible strategy configuration capability enables the system to meet the diversified control needs in different scenarios, improving the versatility and practicality of the system.

[0138] To sum up, the embodiment of the application first performs decision analysis on the obtained early warning event stream through the frequency reduction strategy to obtain a frequency reduction decision analysis result; secondly, an audit execution instruction of video content is generated according to the frequency reduction decision analysis result; if the audit execution instruction is an audit suppression instruction, the early warning event stream for the video content is intercepted; if the audit execution instruction is an audit push instruction, an early warning audit operation for the video content is performed to obtain an audit operation feedback, and then the frequency reduction strategy is updated according to the audit operation feedback. On the one hand, it can efficiently, accurately and intelligently realize flexible potential illegal content pushing, and on the other hand, it can realize feedback closed-loop updating of the frequency reduction strategy, and strengthen the dynamic perception and adaptive adjustment capability of the frequency reduction strategy. Therefore, it can avoid a large number of misjudgment contents being pushed to the artificial audit queue, thereby improving the video content audit efficiency of the media platform.

[0139] Referring to Figure 3 As shown in the figure, the figure is a schematic diagram of the basic structure of a video audit content pushing system 200 provided by the embodiment of the application, and the video audit content pushing system 200 comprises: a processor 201; a storage system 202, which stores a computer program 2020 thereon; When the computer program 2020 is executed by the processor 201, the processor 201 realizes any video audit content pushing method.

[0140] On the basis of the above, a readable storage medium is provided, and the readable storage medium stores a program or instruction, and the program or instruction is executed by the processor to realize the steps of the above method.

[0141] It should be noted that the embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts of each embodiment can be referred to. For the system or system disclosed by the embodiment, since it corresponds to the method disclosed by the embodiment, the description is relatively simple, and the relevant part can be referred to the method part.

Claims

1. A method for pushing video-reviewed content, characterized in that, The method includes: Acquire an early warning event stream, which includes a text-based event stream and an image-based event stream for the same video content, and the text-based event stream and the image-based event stream are time-aligned; Based on the frequency reduction strategy, the early warning event stream is subjected to decision analysis to obtain the frequency reduction decision analysis result. The frequency reduction strategy includes sensitive word frequency reduction strategy, regional image frequency reduction strategy and review feedback frequency reduction strategy. The frequency reduction decision analysis result includes text frequency reduction label, image frequency reduction label and global frequency reduction label. Based on the frequency reduction decision analysis results, an execution instruction for reviewing the video content is generated; If the review execution instruction is a review suppression instruction, then the warning event stream for the video content is blocked; if the review execution instruction is a review push instruction, then the warning review operation for the video content is executed.

2. The method according to claim 1, characterized in that, The decision analysis of the early warning event stream based on the frequency reduction strategy, to obtain the frequency reduction decision analysis results, includes: The sensitive word reduction strategy is used to perform text-level reduction decision analysis on the sensitive word content of the text event stream, and generate text-level reduction decision tags.

3. The method according to claim 2, characterized in that, The step of performing text-level frequency reduction decision analysis on the sensitive word content of the text-type event stream using a sensitive word frequency reduction strategy, and generating text-level frequency reduction decision tags, includes: A streaming computing strategy is used to monitor the OCR push event stream in the video content in real time, and generate a hit distribution curve of the sensitive word content. The historical review results of the sensitive words in the video content are queried through a distributed caching system. Based on the hit count distribution curve and the historical review results, it is determined whether the sensitive word content meets the transient control conditions. If so, the sensitive word content is identified as a transiently controlled frequency-reduced sensitive word; wherein, the transiently controlled frequency-reduced sensitive word is a sensitive word pushed according to a set limited frequency. Obtain the current push timestamp of the user identifier that matches the transiently controlled frequency reduction sensitive word, and query the most recent push timestamp of the user identifier that matches the transiently controlled frequency reduction sensitive word. Generate the text-dimensional frequency reduction decision label based on the current push timestamp and the most recent push timestamp. The text-dimensional frequency reduction decision label includes an event push interception label and an event push allow label.

4. The method according to claim 1, characterized in that, The decision analysis of the early warning event stream based on the frequency reduction strategy, to obtain the frequency reduction decision analysis results, includes: The image-level frequency reduction decision analysis is performed on the regional feature vector of the image event stream using the regional image frequency reduction strategy to generate image-level frequency reduction decision labels.

5. The method according to claim 4, characterized in that, The step of performing image-dimensional down-frequency decision analysis on the regional feature vectors of the image-type event stream using the regional image down-frequency strategy, and generating image-dimensional down-frequency decision labels, includes: If the feature similarity between the region feature vector and any real violation image feature vector in the first feature library exceeds the first similarity threshold, then an image dimension down-frequency decision label is generated to indicate whether the push can be approved. If the average similarity between the region feature vector and the Top-K nearest neighbor feature vectors in the second feature library exceeds the second similarity threshold but does not exceed the first similarity threshold, then an image dimension reduction decision label for push interception is generated; the second feature library is used to store false violation image feature vectors, and the Top-K nearest neighbor feature vectors are the K feature vectors in the second feature library that are closest to the region feature vector in the feature space; If the feature similarity does not exceed the first similarity threshold and the average similarity does not exceed the second similarity threshold, then an image dimension downsampling decision label is generated to indicate the default push.

6. The method according to any one of claims 1-5, characterized in that, The decision analysis of the early warning event stream based on the frequency reduction strategy, to obtain the frequency reduction decision analysis results, includes: By combining the aforementioned review feedback frequency reduction strategy with the historical review operation feedback of the early warning event stream, a global frequency reduction decision analysis is performed to generate a global frequency reduction decision label.

7. The method according to claim 6, characterized in that, The process of using the review feedback frequency reduction strategy, combined with the historical review operation feedback of the early warning event stream, to perform global frequency reduction decision analysis and generate global frequency reduction decision tags includes: The pending push events are stored in the confirmation area of ​​a preset double-buffered event storage structure; the confirmation area is used to store the first event record in the video content in which a review push instruction has been generated but no review operation feedback has been received; the pending push events include at least one of the following: text-level push events with a text-dimensional frequency reduction decision label indicating push release, and image-level push events with an image-dimensional frequency reduction decision label indicating push release or indicating default push. Based on the duration of the pending push event in the pending confirmation area, the pending push event is migrated to the confirmed area of ​​the preset double-buffered event storage structure; the confirmed area is used to store the second event record in the video content that has received review operation feedback; Based on the cumulative number of false violation tags of historical event records in the confirmed area that have the same user identifier and time feature summary as the push event to be processed in the preset statistical period, a global frequency reduction decision tag is generated for pushing and intercepting the push event to be processed. If, during the effective period of the global frequency reduction state of the pending push event, the review result of the historical event record is detected to be adjusted to a real violation label, the global frequency reduction state is lifted, and a global frequency reduction decision label is generated to indicate the release of the push event to be processed.

8. The method according to claim 1, characterized in that, The method further includes: Obtain the feedback from the review operation corresponding to the warning review operation; Based on the feedback from the audit operation, the frequency reduction strategy is updated.

9. The method according to claim 8, characterized in that, The step of updating the frequency reduction strategy based on the feedback from the audit operation includes: Based on the feedback from the review operation, determine the content of historically misjudged sensitive words and add the content of historically misjudged sensitive words to the preset adaptive word list; The system monitors the validity period of any sensitive word in the preset adaptive word list in real time. If the time difference between the current time and the entry time of the sensitive word exceeds the validity period, the system counts the number of new negative feedbacks for the sensitive word within the validity period. The number of new negative feedbacks refers to the number of times the sensitive word is judged as a misjudgment within the validity period. If the number of newly added negative feedbacks is less than a preset threshold, then any sensitive word will be deleted from the preset adaptive word list, and the frequency reduction weight of any sensitive word will be restored. If the number of new negative feedbacks is greater than or equal to the preset number threshold, then the validity period of any sensitive word in the preset adaptive word list is extended, or the frequency reduction weight and entry time of any sensitive word are updated.

10. The method according to claim 8, characterized in that, The step of updating the frequency reduction strategy based on the feedback from the audit operation includes: Based on the violation determination results of the image event stream in the review operation feedback, the feature vectors of real violation images in the first feature library and the feature vectors of fake violation images in the second feature library are updated respectively; the violation determination results include the real violation determination results and the fake violation determination results of the region feature vectors of the image event stream.

11. The method according to claim 1, 2, 3, 4, 5, 8, 9 or 10, characterized in that, The step of generating the video content review execution instruction based on the frequency reduction decision analysis result includes: If at least one decision label exists in the frequency reduction decision analysis result to indicate push blocking, then an audit suppression instruction is generated. The audit suppression instruction includes at least one of the following: text blocking identifier, image blocking identifier, or global blocking identifier. If all decision tags in the frequency reduction decision analysis results are used to indicate push release, then an approval push instruction is generated.

12. A video content review and push system, characterized in that, include: processor; A storage system on which computer programs are stored. When the computer program is executed by the processor, the processor implements the video review content push method as described in any one of claims 1-11.

Images