A method for triggering the preset value point of the platform algorithm

Abstract

The present invention belongs to the technical field of live broadcast data analysis, specifically a method for triggering a preset value point of a platform algorithm, comprising the following steps: setting a differentiated rule base according to the live broadcast risk type, the differentiated rule base being used to store weights corresponding to different live broadcast risk types; data acquisition and preprocessing, including video data, audio data and text data acquisition and preprocessing; feature extraction, extracting video features, audio features and text features from the preprocessed video data, audio data and text data respectively; weight allocation, based on the weights stored in the differentiated rule base, by setting the differentiated rule base, according to the identified live broadcast risk type, corresponding weight allocation can be set for different live broadcast risk types in priority, thereby generating differentiation when calculating the multimodal fusion feature vector, avoiding the same weight allocation, resulting in the same penalty standards between different live broadcast types, causing misjudgment and penalty imbalance problems.

Description

Technical Field

[0001] The present invention belongs to the technical field of live broadcast data analysis, and specifically is a method for triggering a preset value point of a platform algorithm. Background Art

[0002] Online live streaming is currently a more mainstream form of media expression. During the live broadcast process, some illegal behaviors may occur, which may cause discomfort to the audience. When the audience participates in the live broadcast, they will interact with the live broadcast room, including sending barrages and connecting, which may lead to the use of banned words or infringement of others' intellectual property rights. Therefore, in the current live broadcast environment, the live broadcast environment needs to be strictly regulated; triggering the platform algorithm can be understood as an algorithm that automatically performs specific operations based on preset conditions.

[0003] In existing technologies, banned words or illegal behaviors are generally detected by collecting single-modal features from live broadcast videos and comparing them with known illegal behaviors to achieve the purpose of regulating the live broadcast environment. For example, in an e-commerce live broadcast, the host is displaying a red product with text descriptions in the background. At this time, the host accidentally says a sentence that is similar to the illegal words but is not illegal. According to the existing technology, single-modal feature vector calculation can lead to potential misjudgment, as follows:

[0004] For the video mode, the system extracts features such as color and shape from the video and generates a video feature vector. The video feature vector is used to calculate the similarity with the preset feature vector of the illegal picture (such as the pattern of prohibited items). Since the color of the product is similar to the color feature of a certain prohibited item, although other features such as the shape do not match, the single-modal similarity calculation may misjudge it as an illegal picture due to the higher weight of the color feature; for the audio mode, the system extracts features such as spectrum and pitch from the audio and generates an audio feature vector. The audio feature vector is used to calculate the similarity with the preset illegal audio feature vector. The sentence spoken by the anchor is similar to a certain illegal word in pitch and duration. Although the specific words are different, the unimodal similarity calculation may misjudge it as illegal audio due to these similar features; for the text modality, the system extracts features such as word frequency and emotional tendency in the background text, generates a text feature vector, and uses this text feature vector to perform similarity calculation with the preset illegal text feature vector. Some words in the background text are similar to those in the illegal text. Although the overall context is different, the unimodal similarity calculation may misjudge it as illegal text due to these similar words. Based on the above, the unimodal feature vector only considers the information of a certain modality and ignores the complementary information of other modalities, which makes it easy to make misjudgments when the features are similar but not completely matched;

[0005] To this end, the present invention provides a method for triggering a preset value point of a platform algorithm. Summary of the Invention

[0006] In order to make up for the deficiencies of the prior art, at least one technical problem raised in the background technology is solved.

[0007] The technical solution adopted by the present invention to solve the technical problem is: a method for triggering a preset value point of a platform algorithm according to the present invention comprises the following steps:

[0008] S1: Setting a differentiated rule base according to the live broadcast risk type, wherein the differentiated rule base is used to store weights corresponding to different live broadcast risk types;

[0009] S2: Data collection and preprocessing, including video data, audio data, and text data collection and preprocessing;

[0010] S3: Feature extraction, extracting video features, audio features, and text features from the preprocessed video data, audio data, and text data respectively;

[0011] S4: Weight assignment, assigning dynamic weights to video, audio, and text features based on the weights stored in the differentiation rule base;

[0012] S5: Multimodal fusion, representing the extracted video, audio, and text features as feature vectors 、 、 , and perform weighted fusion based on weights;

[0013] S6: Algorithm preset point trigger judgment, including:

[0014] Represent the preset point as a preset feature vector ;

[0015] Calculate the fused feature vector With the preset feature vector Similarity ;

[0016] Set the first threshold ,like

[0017] , the preset point is triggered, otherwise it is not triggered.

[0018] Preferably, the video data acquisition and preprocessing method is: real-time acquisition of live video stream, extraction of video frames at a fixed frame rate, and obtaining a video frame sequence ; Adjust the resolution of the video frames to unify, and then perform denoising on the image;

[0019] The method of audio data acquisition and preprocessing is as follows: synchronously record the live audio stream to obtain a continuous audio signal; denoise the audio signal, and then perform frame and window processing on the audio signal to obtain an audio signal sequence. ;

[0020] The text data collection and preprocessing method is as follows: capture live subtitles and bullet screen text information, perform word segmentation on the text to obtain word units, and mark the part of speech of each word to obtain a text information sequence .

[0021] Preferably, the method for performing weighted fusion based on the weights is:

[0022] Identify live broadcast risk types and obtain differentiated weight distribution based on differentiated rule base 、 、 ;

[0023] According to the formula:

[0024]

[0025] in, ; Any live broadcast type;

[0026] According to the formula, the fusion feature vector is calculated ;

[0027] Calculate the fused feature vector With the preset feature vector Similarity The method is:

[0028]

[0029] According to the formula, the fusion feature vector is calculated With the preset feature vector Similarity .

[0030] Preferably, the similarity is calculated The method also includes:

[0031] Set time window ;

[0032] Calculation time window The cumulative eigenvector within , expressed as:

[0033]

[0034] in, express The fused feature vector at the moment;

[0035] Update fusion feature vector With the preset feature vector The similarity calculation formula is expressed as:

[0036]

[0037] Based on the calculation, the updated .

[0038] Preferably, the preset feature vector The generation method is:

[0039] Collect multimodal data on violations and extract unimodal features;

[0040] Based on single-modal features, multi-modal feature vectors are generated by fusion;

[0041] Based on the multimodal feature vector, a violation database is constructed, which also includes:

[0042] Classification multimodal feature vector and

[0043] Identify a unique sample ID for each multimodal feature vector.

[0044] Preferably, the differentiation rule base is also used to store the classification threshold, based on the live risk type matching first threshold , and dynamically adjust based on real-time feedback.

[0045] Preferably, the differentiation rule base further includes a feedback mechanism, which is used to dynamically adjust the classification threshold according to real-time feedback;

[0046] The dynamic adjustment method of the classification threshold is:

[0047] Real-time collection of audience feedback data during live broadcast;

[0048] Based on natural language technology, we differentiate audience feedback data and count the number of negative emotions in comments. , number of reports and interaction rate ;

[0049] Calculate the adjustment range of the rating threshold based on audience feedback data , according to the formula:

[0050]

[0051] in, is the emotional impact coefficient of the barrage, is the reporting impact coefficient, is the total number of barrages, is the interaction influence coefficient;

[0052] The calculated adjustment Update the first threshold , expressed as:

[0053]

[0054]

[0055] in, is the second threshold;

[0056] The calculated second threshold Applied to the updated , determine whether the preset point is triggered.

[0057] Preferably, the interaction rate The calculation method is:

[0058] Count the number of user likes per unit time , number of shares and the number of comments ;

[0059] According to the formula:

[0060]

[0061] in, The real-time number of people online in the live broadcast room.

[0062] Preferably, the time window The length is dynamically adjusted according to the live broadcast type, as follows:

[0063] Identify the risk type of live broadcast and define the time window based on the risk type of live broadcast Length, including:

[0064] First time window, s , the live broadcast type is high risk;

[0065] The second time window, s , the live broadcast type meets the medium risk;

[0066] The third time window, s , the live broadcast type is low risk;

[0067] The time window that will be matched Apply to The update calculation is in progress.

[0068] Preferably, the differentiation rule base also stores a risk level-first threshold mapping table, the grading threshold is bound to the live broadcast risk type, and the specific rule is to use the risk level-first threshold mapping table stored in the differentiation rule base to map the live broadcast risk type with the first threshold. Automatic adaptation.

[0069] The beneficial effects of the present invention are as follows:

[0070] 1. The method of triggering the preset value point of the platform algorithm described in the present invention, by setting a differentiated rule library, can give priority to setting corresponding weight distribution for different live broadcast risk types according to the identified live broadcast risk type, so that differentiation is generated when calculating the multimodal fusion feature vector, avoiding the same weight distribution, resulting in the same penalty standards between different live broadcast types, causing misjudgment and penalty imbalance.

[0071] 2. The method of triggering the preset value point of the platform algorithm described in the present invention is based on multimodal fusion features, and then calculates the similarity with known or preset violations in sequence to obtain the similarity , according to the similarity The size of With the first threshold A comparison is made to determine whether there is any violation during the live broadcast process. Based on the similarity calculation based on the preset violations, the violation corresponding to the maximum similarity is selected, which represents the possible violation in the live broadcast room. The multimodal fusion feature vector based on the same time axis is used to reduce the probability of misjudgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The present invention will be further described below with reference to the accompanying drawings.

[0073] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0074] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0075] like Figure 1 As shown, a method for triggering a preset value point of a platform algorithm according to an embodiment of the present invention includes the following steps:

[0076] S1: Setting a differentiated rule base according to the live broadcast risk type, wherein the differentiated rule base is used to store weights corresponding to different live broadcast risk types;

[0077] S2: Data collection and preprocessing, including video data, audio data, and text data collection and preprocessing;

[0078] S3: Feature extraction, extracting video features, audio features, and text features from the preprocessed video data, audio data, and text data respectively;

[0079] S4: Weight assignment, assigning dynamic weights to video, audio, and text features based on the weights stored in the differentiation rule base;

[0080] S5: Multimodal fusion, representing the extracted video, audio, and text features as feature vectors 、 、 , and perform weighted fusion based on weights;

[0081] S6: Algorithm preset point trigger judgment, including:

[0082] Represent the preset point as a preset feature vector ;

[0083] Calculate the fused feature vector With the preset feature vector Similarity ;

[0084] Set the first threshold ,like

[0085] , the preset point is triggered, otherwise it is not triggered.

[0086] According to the existing technology of single-mode feature vector calculation, potential misjudgment will occur as follows:

[0087] For the video mode, the system extracts features such as color and shape from the video and generates a video feature vector. The video feature vector is used to perform similarity calculation with the preset feature vector of the illegal picture (such as the pattern of prohibited items). Since the color of the product is similar to the color feature of a certain prohibited item, although other features such as shape do not match, the single-modal similarity calculation may misjudge it as an illegal picture due to the higher weight of the color feature. For the audio mode, the system extracts features such as spectrum and tone from the audio and generates an audio feature vector. The audio feature vector is used to perform similarity calculation with the preset illegal audio feature vector. In the similarity calculation, the sentence spoken by the anchor is similar to a certain illegal word in pitch and duration. Although the specific words are different, the unimodal similarity calculation may misjudge it as illegal audio due to these similar features; for the text modality, the system extracts features such as word frequency and emotional tendency in the background text, generates a text feature vector, and uses this text feature vector to perform similarity calculation with the preset illegal text feature vector. Some words in the background text are similar to those in the illegal text. Although the overall context is different, the unimodal similarity calculation may misjudge it as illegal text due to these similar words.

[0088] Based on the above, the single-modal feature vector only considers the information of a certain modality and ignores the complementary information of other modalities, which leads to misjudgment when the features are similar but not completely matched; Based on this, in one embodiment of the present invention, when analyzing the live video, the multimodal features of the live video are collected separately, and are pre-processed separately and then fused to obtain the multimodal fusion features of the live video. Based on the multimodal fusion features, the similarity is calculated in sequence with the known or preset violations to obtain the similarity , according to the similarity The size of With the first threshold A comparison is made to determine whether there is any violation during the live broadcast process, and based on the similarity calculation based on the preset violations, the violation corresponding to the maximum similarity is selected, which represents the possible violation in the live broadcast room. In addition, due to different live broadcast risk types, the corresponding violations may exist in the video, audio and barrage respectively. Therefore, when calculating the multimodal fusion features, it is also necessary to distinguish the weight distribution corresponding to different live broadcast risk types. It is understandable that in high-risk live broadcasts, such as game live broadcasts, there may be a higher probability of banned words or uncivilized language. Then, for game live broadcasts, the weight ratio of its audio is higher. On the contrary, for low-risk live broadcasts, such as educational live broadcasts, the probability of violations in audio or video is lower, and the probability of violations in text may be higher. Therefore, the weight ratio of text in educational live broadcasts is higher. Based on the above-mentioned differentiated rule library, according to the identified live broadcast risk type, the corresponding weight distribution can be set for different live broadcast risk types first, so that differentiation is generated when calculating the multimodal fusion feature vector, avoiding the same weight distribution, resulting in the same penalty standards between different live broadcast types, causing misjudgment and penalty imbalance.

[0089] In one embodiment, the video data acquisition and preprocessing method is: real-time acquisition of live video stream, extraction of video frames at a fixed frame rate, and obtaining a video frame sequence. ; Adjust the resolution of the video frames to unify, and then perform denoising on the image;

[0090] The method of audio data acquisition and preprocessing is as follows: synchronously record the live audio stream to obtain a continuous audio signal; denoise the audio signal, and then perform frame and window processing on the audio signal to obtain an audio signal sequence. ;

[0091] The text data collection and preprocessing method is as follows: capture live subtitles and bullet screen text information, perform word segmentation on the text to obtain word units, and mark the part of speech of each word to obtain a text information sequence .

[0092] When collecting live broadcast data, video data, audio data and text data are mainly collected. The preprocessing of video data, audio data and text data is based on a sequence established on the time axis, which can ensure the synchronization of the time axis between multimodal features when calculating the multimodal fusion feature vector in the later stage, thereby reducing the probability of misjudgment.

[0093] In one embodiment, the method for performing weighted fusion based on the weights is:

[0094] Identify live broadcast risk types and obtain differentiated weight distribution based on differentiated rule base 、 、 ;

[0095] According to the formula:

[0096]

[0097] in, ; Any live broadcast type;

[0098] According to the formula, the fusion feature vector is calculated ;

[0099] Calculate the fused feature vector With the preset feature vector Similarity The method is:

[0100]

[0101] According to the formula, the fusion feature vector is calculated With the preset feature vector Similarity .

[0102] Since the traditional method of comparing single-modal feature vectors with illegal behaviors has a high misjudgment rate, in one embodiment of the present invention, a multi-modal fusion feature vector synchronized with the time axis is used. Preset feature vectors corresponding to violations By making comparisons, we can reduce the misjudgment and penalty imbalance caused by single-modal feature vectors. Specifically, we can obtain the weight distribution corresponding to the live broadcast risk type based on the differentiated rule base. 、 、 , and then calculate the multimodal fusion feature vector based on the formula ,in 、 and All at the same time The collected features are calculated according to the formula to obtain the fused feature vector , applied to the similarity calculation, and then obtain the fusion feature vector With the preset feature vector Similarity , based on the above, calculate the similarity After that, with the preset first threshold A comparison is made to determine whether a preset value point is triggered, and the preset value point can be understood as whether any violation is triggered.

[0103] In one embodiment, the similarity is calculated The method also includes:

[0104] Set time window ;

[0105] Calculation time window The cumulative eigenvector within , expressed as:

[0106]

[0107] in, express The fused feature vector at the moment;

[0108] Update fusion feature vector With the preset feature vector The similarity calculation formula is expressed as:

[0109]

[0110] Based on the calculation, the updated .

[0111] Since the fusion feature vector calculated above With the preset feature vector Similarity ,in 、 and All at the same time The collected features may be misjudged due to instantaneous fluctuations. In one embodiment of the present invention, a time window is introduced. , smoothing instantaneous fluctuations that cause temporary abnormal misjudgments, taking the detection of violent scenes as an example:

[0112] Set time window ;

[0113] Real-time collection of video, audio, and text data within 5 seconds;

[0114] Extract the average color histogram and motion vector for each frame of video, extract the average audio energy and spectrum features for each audio segment, and extract the frequency of violence-related keywords for text;

[0115] Average the eigenvectors within 5 seconds to form a cumulative eigenvector ;

[0116] calculate Preset feature vectors with violent scenes Similarity ;

[0117] If the similarity Exceeding the first threshold , an alarm is triggered, indicating that a violent scene has been detected; in the above embodiment, by introducing a time window By transforming the corresponding formula, it is possible to more accurately determine whether the preset points are triggered during the live broadcast, thereby reducing the misjudgment rate.

[0118] In one embodiment, the preset feature vector The generation method is:

[0119] Collect multimodal data on violations and extract unimodal features;

[0120] Based on single-modal features, multi-modal feature vectors are generated by fusion;

[0121] Based on the multimodal feature vector, a violation database is constructed, which also includes:

[0122] Classification multimodal feature vector and

[0123] Identify a unique sample ID for each multimodal feature vector.

[0124] The generation of a preset vector P forms the basis for multimodal fusion feature vector representation. For example, in e-commerce live streaming, if the preset vector P includes the features of "red products + promotional language," the system can quickly identify illegal advertisements. The specific steps include categorizing and storing the fused multimodal feature vectors by violation type and assigning unique IDs. Real-time live streaming data can be quickly matched to known violation patterns in the database through similarity calculations. The database supports the continuous addition of new samples for model retraining to improve detection capabilities. Violations are then fine-grainedly classified (e.g., violence is divided into physical confrontation and weapon display), and each sample is uniquely identified. Different types of violations correspond to different handling strategies (such as interrupting the live stream or providing pop-up warnings). The IDs allow the sample source to be traced, assisting auditors in reviewing cases of misjudgment.

[0125] In one embodiment, the differentiation rule base is further used to store the classification threshold, based on the live broadcast risk type matching first threshold , and dynamically adjust based on real-time feedback.

[0126] In the above description, since the locations of violations corresponding to different live broadcast risk types are different, the locations of occurrence described here mainly refer to feature types, such as video features, audio features or text features, so different weight distributions need to be matched. In addition, due to the different live broadcast risk types, the risks in essence are also different. For example, the similarity distributions calculated for low-risk live broadcasts and high-risk live broadcasts are also different. The similarity distribution corresponding to low-risk live broadcasts should correspond to a smaller numerical distribution, while the similarity distribution corresponding to high-risk live broadcasts should correspond to a larger numerical distribution. If the same first threshold is used , high-risk live broadcasts are more likely to be penalized, while low-risk live broadcasts are more difficult to be penalized. Based on the above, in one embodiment of the present invention, a grading threshold is set based on the live broadcast risk type. The grading threshold is used to match different live broadcast risk types. That is, it can be understood that multiple first thresholds are set. , used to match different live broadcast risk types. When analyzing whether live broadcast data violates regulations, the first threshold is matched by identifying the live broadcast risk type. ;

[0127] Exemplary:

[0128] For high-risk live broadcast types, after identification, the first threshold of the match ;

[0129] For medium-risk live broadcast types, after identification, the first threshold of the match ;

[0130] For low-risk live broadcast types, after identification, the first threshold of the match ;

[0131] Based on the above, for high-risk live broadcast types, their live broadcast content may cause a high degree of similarity between them and the violations recorded in the violation database. Therefore, the first threshold for matching is relatively high. For corresponding low-risk live broadcast types, since their historical live broadcasts have fewer misjudgments, or in other words, in their live broadcast history, the live broadcast content is quite different from the violations. Therefore, the similarity is relatively low, so a lower similarity threshold is matched for them.

[0132] In one embodiment, the differentiation rule base further includes a feedback mechanism, wherein the feedback mechanism is used to dynamically adjust the classification threshold according to real-time feedback;

[0133] The dynamic adjustment method of the classification threshold is:

[0134] Real-time collection of audience feedback data during live broadcast;

[0135] Based on natural language technology, we differentiate audience feedback data and count the number of negative emotions in comments. , number of reports and interaction rate ;

[0136] Calculate the adjustment range of the rating threshold based on audience feedback data , according to the formula:

[0137]

[0138] in, is the emotional impact coefficient of the barrage, is the reporting impact coefficient, is the total number of barrages, is the interaction influence coefficient;

[0139] The calculated adjustment Update the first threshold , expressed as:

[0140]

[0141]

[0142] in, is the second threshold;

[0143] The calculated second threshold Applied to the updated , determine whether the preset point is triggered.

[0144] Based on the above, after identifying the live broadcast risk type, the differentiated rule library matches the first threshold for live broadcast data analysis Based on this, we should also consider the real-time feedback of the audience during the live broadcast, so as to adjust the first threshold Perform floating adjustment to pass the first threshold of floating adjustment To improve the accuracy of the penalty, in one embodiment of the present invention, by collecting audience feedback data in real time, the number of negative emotions in the barrage is counted based on natural language technology. , number of reports and interaction rate ; Based on the above data, the threshold adjustment range of the live data is calculated according to the formula , based on the calculated threshold adjustment amplitude , it is also necessary to determine the threshold adjustment range Whether the adjustment requirements are met, specifically, the settings as well as ,when When , it indicates the threshold adjustment range obtained by this calculation It is less than the minimum adjustment value, so no adjustment is made and the recovery threshold is the first threshold. , and when , then adjustment is required, and the second threshold is calculated according to the formula , using the second threshold As the first threshold after update Based on the above, the first threshold of the live broadcast data analyzed is adjusted by floating through the audience feedback data. , which is conducive to quickly responding to abnormal live broadcast situations and reducing the probability of imbalance in live broadcast penalties. For example: Assume that the first threshold , Negative barrage rate: , reporting rate: , interaction rate: , Barrage emotional impact coefficient: , Reporting impact coefficient: , interaction influence coefficient: ;

[0145] Calculation process:

[0146]

[0147]

[0148] Based on the above calculation, the threshold adjustment range , so the first threshold No update.

[0149] In one embodiment, the interaction rate The calculation method is:

[0150] Count the number of user likes per unit time , number of shares and the number of comments ;

[0151] According to the formula:

[0152]

[0153] in, The real-time number of people online in the live broadcast room.

[0154] The above interaction rate To correct the adjustment range , when the interaction rate When it is high, it can suppress the excessive lowering of the threshold due to the instantaneous reporting volume.

[0155] In one embodiment, the time window The length is dynamically adjusted according to the live broadcast type, as follows:

[0156] Identify the risk type of live broadcast and define the time window based on the risk type of live broadcast Length, including:

[0157] First time window, s , the live broadcast type is high risk;

[0158] The second time window, s , the live broadcast type meets the medium risk;

[0159] The third time window, s , the live broadcast type is low risk;

[0160] The time window that will be matched Apply to The update calculation is in progress.

[0161] After identifying the live broadcast type, you can also match time windows of different lengths based on the live broadcast risk type. , thereby reducing the imbalance in penalties for live broadcast violations. In addition, linking the time window with the risk level of the live broadcast type can improve the robustness of multimodal fusion.

[0162] The differentiation rule base also stores a risk level-first threshold mapping table, and the classification threshold is bound to the live broadcast risk type. The specific rule is to use the risk level-first threshold mapping table stored in the differentiation rule base to match the live broadcast risk type with the first threshold. Automatic adaptation.

[0163] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for triggering a preset value point of a platform algorithm, characterized in that: The following steps are involved: S1: Setting a differentiated rule base according to the live broadcast risk type, wherein the differentiated rule base is used to store weights corresponding to different live broadcast risk types; S2: Data collection and preprocessing, including video data, audio data, and text data collection and preprocessing; S3: Feature extraction, extracting video features, audio features, and text features from the preprocessed video data, audio data, and text data respectively; S4: Weight assignment, assigning dynamic weights to video, audio, and text features based on the weights stored in the differentiation rule base; S5: Multimodal fusion, representing the extracted video, audio, and text features as feature vectors 、 、 , and perform weighted fusion based on weights; S6: Algorithm preset point trigger judgment, including: Represent the preset point as a preset feature vector ; Calculate the fused feature vector With the preset feature vector Similarity ; Set the first threshold ,like , then the preset point is triggered, otherwise it is not triggered; The differentiation rule base is also used to store the classification threshold, based on the live risk type matching first threshold , and dynamically adjust based on real-time feedback; The differentiation rule base further includes a feedback mechanism for dynamically adjusting the classification threshold based on real-time feedback; The dynamic adjustment method of the classification threshold is: Real-time collection of audience feedback data during live broadcast; Based on natural language technology, we differentiate audience feedback data and count the number of negative emotions in comments. , number of reports and interaction rate ; Calculate the adjustment range of the rating threshold based on audience feedback data , according to the formula: in, is the emotional impact coefficient of the barrage, is the reporting impact coefficient, is the total number of barrages, is the interaction influence coefficient; The calculated adjustment Update the first threshold , expressed as: in, is the second threshold; The calculated second threshold Applied to the updated , determine whether the preset point is triggered; Said interaction rate The calculation method is: Count the number of user likes per unit time , number of shares and the number of comments ; According to the formula: in, The real-time number of people online in the live broadcast room.

2. A method for triggering a platform algorithm preset value point according to claim 1, characterized in that: The video data acquisition and preprocessing method is: real-time acquisition of live video stream, extraction of video frames at a fixed frame rate, and obtaining a video frame sequence. ; Adjust the resolution of the video frames to unify, and then perform denoising on the image; The method of audio data acquisition and preprocessing is as follows: synchronously record the live audio stream to obtain a continuous audio signal; denoise the audio signal, and then perform frame and window processing on the audio signal to obtain an audio signal sequence. ; The text data collection and preprocessing method is as follows: capture live subtitles and bullet screen text information, perform word segmentation on the text to obtain word units, and mark the part of speech of each word to obtain a text information sequence .

3. The method for triggering a preset value point of a platform algorithm according to claim 2, characterized in that: The method for weighted fusion based on the weights is: Identify live broadcast risk types and obtain differentiated weight distribution based on differentiated rule base 、 、 ; According to the formula: in, ; Any live broadcast type; According to the formula, the fusion feature vector is calculated ; Calculate the fused feature vector With the preset feature vector Similarity The method is: According to the formula, the fusion feature vector is calculated With the preset feature vector Similarity .

4. A method for triggering a platform algorithm preset value point according to claim 3, characterized in that: Calculate the similarity The method also includes: Set time window ; Calculation time window The cumulative eigenvector within , expressed as: in, express The fused feature vector at the moment; Update fusion feature vector With the preset feature vector The similarity calculation formula is expressed as: Based on the calculation, the updated .

5. The method for triggering a preset value point of a platform algorithm according to claim 4, characterized in that: The preset feature vector The generation method is: Collect multimodal data on violations and extract unimodal features; Based on single-modal features, multi-modal feature vectors are generated by fusion; Based on the multimodal feature vector, a violation database is constructed, which also includes: Classification multimodal feature vector and Identify a unique sample ID for each multimodal feature vector.

6. The method for triggering a preset value point of a platform algorithm according to claim 4, characterized in that: The time window The length is dynamically adjusted according to the live broadcast type, as follows: Identify the risk type of live broadcast and define the time window based on the risk type of live broadcast Length, including: First time window, s , the live broadcast type is high risk; The second time window, s , the live broadcast type meets the medium risk; The third time window, s , the live broadcast type is low risk; The time window that will be matched Apply to The update calculation is in progress.

7. The method for triggering a preset value point of a platform algorithm according to claim 6, characterized in that: The differentiation rule base also stores a risk level-first threshold mapping table, and the classification threshold is bound to the live broadcast risk type. The specific rule is to use the risk level-first threshold mapping table stored in the differentiation rule base to match the live broadcast risk type with the first threshold. Automatic adaptation.

Images