Content auditing method and device, storage medium and electronic equipment

Abstract

The invention discloses a content auditing method and device, a storage medium and electronic equipment. The method comprises the steps of obtaining target interaction content in a content interaction platform; the target interaction content is input into a multi-mode semantic recognition model, a fused semantic vector is obtained, and the fused semantic vector is used for indicating emotion description information of the target interaction content; based on the fusion semantic vector and the regional culture description vector, evaluation parameters in the content auditing model are adjusted, an adjusted content auditing model is obtained, and the evaluation parameters represent the influence degree of part of content with special meanings on the auditing result; and generating a target auditing result based on the adjusted content auditing model. According to the method and the device, the technical problem of relatively low accuracy in a content auditing process in related technologies is solved.

Description

Technical Field

[0001] The present application relates to the field of computers, and more specifically, to a content review method and device, a storage medium, and an electronic device. Background Art

[0002] With the rapid development of the online live broadcast industry, content review has become an important part of platform security management and user experience. The content review of existing live broadcast platforms mainly includes two methods: manual review and automated review. However, the efficiency of identifying and filtering non-compliant content through manual review is low. Most of the existing automated review systems rely on keyword matching and simple rule engines, which are difficult to cope with complex and changing content scenarios.

[0003] For example, in an international live broadcast platform, the content to be reviewed may be generated by users from different countries and regions. The above-mentioned methods such as keyword matching lack the ability to effectively identify and process content with different languages ​​and cultural backgrounds, which means that the accuracy of the review results for multilingual content cannot be ensured. This leads to the technical problem of low accuracy of the review results of live broadcast content.

[0004] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0005] The embodiments of the present application provide a content review method and device, a storage medium and an electronic device to at least solve the technical problem of low efficiency that occurs during the content review process.

[0006] According to one aspect of an embodiment of the present application, a content review method is provided, including: obtaining target interactive content in a content interaction platform; obtaining a fused semantic vector by inputting the target interactive content into a multimodal semantic recognition model, wherein the fused semantic vector is used to indicate emotional description information of the target interactive content; based on the fused semantic vector and the regional cultural description vector, adjusting the evaluation parameters in the content review model to obtain an adjusted content review model, wherein the evaluation parameters represent the degree of influence of some content with special meanings on the review results, and the regional cultural description vector represents the deep meaning obtained by deeply mining the target interactive content in combination with regional cultural information; based on the adjusted content review model, generating a target review result.

[0007] Optionally, the above-mentioned fused semantic vector is obtained by inputting the target interaction content into a multimodal semantic recognition model, including: extracting features from video interaction data in the multimodal interaction data to obtain video interaction features, wherein the multimodal interaction data includes video interaction data, audio interaction data and text interaction data, and the target interaction content includes multimodal interaction data; extracting features from audio interaction data to obtain audio interaction features; extracting features from text interaction data to obtain text interaction features; and obtaining a fused semantic vector by fusing video interaction features, audio interaction features and text interaction features, and inputting the fused interaction features into the multimodal semantic recognition model.

[0008] Optionally, the above-mentioned adjustment of evaluation parameters in the content review model based on the fused semantic vector and the regional cultural description vector includes: obtaining the account behavior data corresponding to the account that publishes the target interactive content and the regional description data of the location of the client logged in by the account; obtaining the regional cultural description vector for representing the current state of the account by inputting the account behavior data and the regional description data into the target perception model; and adjusting the evaluation parameters in the content review model based on the emotional type of the target interactive content represented by the fused semantic vector and the deep meaning represented by the regional cultural description vector.

[0009] Optionally, the above-mentioned adjustment of the evaluation parameters in the content review model based on the emotion type of the target interactive content represented by the fused semantic vector and the deep meaning represented by the regional cultural description vector includes: when the fused semantic vector indicates that the emotion type of the target interactive content is the first type and the deep meaning indicates compliance with regional cultural description standards, increasing the value of the evaluation parameter, wherein the higher the value of the evaluation parameter, the less likely it is that some content with special meaning in the target interactive content will be reviewed as non-compliant content; when the fused semantic vector indicates that the emotion type of the target interactive content is the second type and the deep meaning indicates non-compliance with regional cultural description standards, reducing the value of the evaluation parameter; when the fused semantic vector indicates that the emotion type of the target interactive content is the first type and the deep meaning indicates non-compliance with regional cultural description standards, reducing the value of the evaluation parameter.

[0010] Optionally, the above-mentioned target audit result is generated based on the adjusted content audit model, including: using a generative adversarial network to detect the target interaction content to obtain a target detection result, wherein the generative adversarial network is built into the content audit model; when the target detection result indicates that the target interaction content does not contain non-compliant content, differential privacy processing is performed on the target interaction content by introducing random noise to obtain processed interaction content; and the target audit result is obtained by inputting the processed interaction content into the adjusted content audit model.

[0011] Optionally, the above-mentioned use of a generative adversarial network to detect the target interactive content and obtain a target detection result includes: performing word segmentation and encoding processing on the text data in the target interactive content to obtain processed data features; and obtaining the target detection result by inputting the processed data features into a discriminator in the generative adversarial network.

[0012] Optionally, after obtaining the target detection result by inputting the processed data features into the discriminator in the generative adversarial network, the above method also includes: when the target detection result indicates that there is an abnormality in the target interaction content, adding an abnormal mark to the part of the target interaction content that has the abnormality; optimizing the weight parameters in the discriminator to obtain an optimized generative adversarial network.

[0013] Optionally, after generating a target audit result based on the adjusted content audit model, the above method also includes: when the target audit result indicates that the target interactive content contains non-compliant content, processing the target interactive content includes at least one of the following: blocking the target interactive content, adjusting the target interactive content, and modifying the target interactive content.

[0014] According to another aspect of the embodiment of the present application, a content review device is also provided, including: a first acquisition unit, used to acquire target interactive content in a content interaction platform; a first processing unit, used to obtain a fused semantic vector by inputting the target interactive content into a multimodal semantic recognition model, wherein the fused semantic vector is used to indicate emotional description information of the interactive content; an adjustment unit, used to adjust evaluation parameters in the content review model based on the fused semantic vector and the regional cultural description vector to obtain an adjusted content review model, wherein the evaluation parameters represent the degree of influence of some content with special meaning on the review result, and the regional cultural description vector represents the deep meaning obtained by deeply mining the target interactive content in combination with regional cultural information; a second processing unit, used to generate a target review result based on the adjusted content review model.

[0015] According to another aspect of the embodiments of the present application, a computer-readable storage medium is provided, in which a computer program is stored, wherein the computer program is used to execute the above-mentioned content review method when run by an electronic device.

[0016] According to another aspect of the embodiments of the present application, a computer program product is provided, including a computer program that implements the steps of the above method when the computer program is executed by a processor.

[0017] According to another aspect of an embodiment of the present application, there is also provided an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to execute the content review method through the computer program.

[0018] By adopting the above-mentioned embodiment provided by the present application, the live broadcast content containing multiple language types is translated and processed, and the live broadcast content is converted into target interactive content of a unified language type. Then, based on the target interactive content, the multimodal live broadcast data is combined to obtain a fused semantic vector for representing the emotional information of the live broadcast content. Finally, by combining the fused semantic vector with the regional cultural description vector of the user's location, the model can fully understand the contextual meaning and cultural background knowledge of the content to be reviewed, thereby adjusting the evaluation parameters in the content review model in real time and generating the target review results. In other words, by performing translation processing, multimodal data fusion, context understanding and cultural background recognition on the live broadcast content in the multilingual live broadcast scene, it is possible to accurately identify the live broadcast content that does not meet the requirements, thereby achieving the technical effect of improving the accuracy of the content recognition results. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0020] Figure 1 is a schematic diagram of an application scenario of an optional content review method according to an embodiment of the present application;

[0021] Figure 2 is a flow chart of an optional content review method according to an embodiment of the present application;

[0022] Figure 3 is an overall flow chart of an optional multi-language support and cross-cultural review mechanism according to an embodiment of the present application;

[0023] Figure 4 is a flowchart of an optional multimodal sentiment and semantic analysis according to an embodiment of the present application;

[0024] Figure 5 is a flow chart of an optional context-aware and personalized audit strategy adjustment according to an embodiment of the present application;

[0025] Figure 6 is a flowchart of an optional dynamic content detection based on a generative adversarial network according to an embodiment of the present application;

[0026] Figure 7 is a flowchart of an optional multi-level federated learning and privacy protection method according to an embodiment of the present application;

[0027] Figure 8 is a schematic diagram of the structure of an optional content review device according to an embodiment of the present application;

[0028] Fig. 9 It is a schematic diagram of the structure of an optional electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0029] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0030] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0031] The technical solutions in the embodiments of the present application will comply with legal provisions during implementation. When performing operations according to the technical solutions in the embodiments, the data used will not involve user privacy. While ensuring that the operation process is compliant and legal, the security of the data is guaranteed.

[0032] In addition, when the above embodiments of the present application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data need to comply with relevant regulations and standards of the relevant country or region.

[0033] According to one aspect of the embodiments of the present application, a content audit method is provided. As an optional implementation, the content audit method can be applied to, but is not limited to, Figure 1 The application scenario shown in Figure 1In the application scenario shown, the target terminal 102 may, but is not limited to, communicate with the server 106 via the network 104, and the server 106 may, but is not limited to, perform operations on the database 108, such as write data operations or read data operations. The above-mentioned target terminal 102 may, but is not limited to, include a human-computer interaction screen, a processor, and a memory. The above-mentioned human-computer interaction screen may, but is not limited to, be used to display the target interaction content and target review results in the content interaction platform on the target terminal 102. The above-mentioned processor may, but is not limited to, be used to respond to the above-mentioned human-computer interaction operations, perform corresponding operations, or generate corresponding instructions, and send the generated instructions to the server 106. The above-mentioned memory is used to store relevant processing data, such as fused semantic vectors, evaluation parameters, and regional cultural description information.

[0034] Optionally, in this embodiment, the target terminal may be a terminal configured with a target client, which may include but is not limited to at least one of the following: a mobile phone (such as an Android phone, an iOS phone, etc.), a laptop, a tablet computer, a PDA, a MID (Mobile Internet Devices), a PAD, a desktop computer, a smart TV, etc. The target client may be a video client, an instant messaging client, a browser client, an education client, etc. The network may include but is not limited to: a wired network, a wireless network, wherein the wired network includes: a local area network, a metropolitan area network and a wide area network, and the wireless network includes: Bluetooth, WIFI and other networks that realize wireless communication. The server may be a single server, or a server cluster consisting of multiple servers, or a cloud server.

[0035] In order to solve the problem of low data processing efficiency, a content review method is proposed in the embodiment of the present application. Figure 2 It is a flowchart of the content review method according to an embodiment of the present application, and the process includes the following steps S202 to S208.

[0036] It should be noted that the content review method shown in steps S202 to S208 can be performed by, but is not limited to, an electronic device, wherein the electronic device can be, but is not limited to, Figure 1 The target terminal or server is shown.

[0037] Step S202, obtaining target interactive content in the content interactive platform;

[0038] It should be noted that the target interactive content may be content in the same language (for example, all comments in Chinese), or may be interactive content in multiple different languages, which is not limited in the present embodiment.

[0039] When it is identified that the language type of the interactive content includes a non-mainstream language, the target interactive content is translated, specifically, the interactive content in the non-mainstream language is translated into the interactive content in the mainstream language.

[0040] In this embodiment, multi-language translation and analysis are supported, wherein the system automatically identifies the language type of the user input, preparing for subsequent translation and sentiment analysis. The process of identifying the language of the target interactive content includes:

[0041] (1) Language detection: Use language detection algorithms (such as LangID) to automatically identify the language type of the input text. Supported languages ​​include Chinese, English, Japanese, Korean and other common languages;

[0042] (2) Language tagging: The identified language type is marked and attached to the text data to ensure that subsequent processing steps can identify and apply the correct language processing model.

[0043] When it is identified that the target interactive content contains content in a non-mainstream language, the identified content is translated. The content review system uses an integrated multi-language translation API to translate the text entered by the user into the platform's main review language (such as English). The basic process of translation processing is as follows:

[0044] (1) Sending a translation request: The system sends a translation request to a third-party translation API (such as Google Translate, Microsoft Translator) based on the language recognition results;

[0045] (2) Translation result processing: The system receives the results returned by the translation API and compares them with the original text to ensure the accuracy of the translation. If necessary, the system will fine-tune the translation results;

[0046] (3) Application of translation results: The translated text is passed to the sentiment analysis model or content review module for further processing.

[0047] Step S204, obtaining a fused semantic vector by inputting the target interaction content into a multimodal semantic recognition model, wherein the fused semantic vector is used to indicate the emotional description information of the target interaction content;

[0048] Step S206, based on the fused semantic vector and the regional cultural description vector, the evaluation parameters in the content audit model are adjusted to obtain an adjusted content audit model, wherein the evaluation parameters represent the degree of influence of the part of the content with special meaning on the audit result, and the regional cultural description vector represents the deep meaning obtained by deeply mining the target interactive content in combination with the regional cultural information;

[0049] Step S208, generating target audit results based on the adjusted content audit model.

[0050] The technical solutions in the embodiments of the present application can be applied to, but are not limited to, a global live broadcast platform that supports multiple languages, for example, an online live broadcast scene in which the live broadcast content includes both Chinese and English, wherein the live broadcast content includes, but is not limited to, video content, audio content, pictures, barrage, and comments from viewers who watch the live broadcast online.

[0051] For ease of understanding, the embodiments of the present application all take the live broadcast platform as the content interaction platform and the target interactive content as the live broadcast content as an example to explain the above content review method.

[0052] In the above scenario, users who publish live content in different languages ​​may be from the same region or from different regions around the world. In this regard, existing live content review platforms cannot accurately identify the true meaning of live content in non-mainstream languages, resulting in low accuracy of review results.

[0053] In order to solve the above problems, a multilingual support and cross-cultural review mechanism is proposed in the embodiment of the present application, which can adapt to the content review needs of the global live broadcast platform. At the same time, the content review system also integrates a multilingual translation API and a cultural background recognition module, which can automatically recognize and translate languages ​​from different countries, including Chinese, English, Japanese, Korean, etc. At the same time, the system dynamically adjusts the review strategy in combination with the cultural habits and review standards of various countries to ensure accurate review of content under different cultural backgrounds. For example, when processing some content with specific cultural metaphors, the system can understand its meaning and make corresponding review decisions.

[0054] like Figure 3 As shown, the content audit system in the embodiment of the present application includes but is not limited to a language recognition module, a translation module, a sentiment analysis module, a cultural background analysis module, an audit strategy module, an audit system, and all synchronization modules, etc. The functions of each module are briefly introduced below.

[0055] (1) Language identification module: used to identify whether the language type of the target interactive content to be reviewed is a mainstream language (e.g., Chinese);

[0056] (2) Translation module: used to translate some interactive content in non-mainstream languages ​​into content in mainstream languages, for example, translating English bullet comments into Chinese bullet comments, etc.

[0057] (3) Sentiment analysis module: used to preliminarily identify the sentiment type of the interactive content to be reviewed, such as positive (indicating that the content is optimistic), negative (negative content), or neutral (indicating objective expression);

[0058] (4) Cultural background analysis module: used to obtain the cultural background technology of the region or area based on the region where the user who posted the target interactive content in different languages ​​is located, so as to understand the true meaning of the content to be reviewed;

[0059] (5) Review decision module: used to use the analysis results of the aforementioned sentiment analysis module, cultural background analysis module and other modules to determine whether to adjust the evaluation parameters in the content review model;

[0060] (6) Audit system: used to output audit results in real time based on the adjusted content audit model;

[0061] (7) Global synchronization module: used to synchronize the audit results to each server of the distributed content audit system.

[0062] Combine the following Figure 3 , the specific steps of using the above modules to conduct content review are described in detail:

[0063] S302, submitting the target interactive content to be reviewed to the user platform;

[0064] Among them, the user platform includes but is not limited to the content review system of the content interaction platform, and the target interactive content includes but is not limited to the barrage, comments, and explanations of the anchor sent by the audience during the live broadcast. The target interactive content can be multimodal data, such as video, picture, and audio.

[0065] S304, detecting the language of the target interactive content through a language recognition module, and transmitting the recognition result to a translation module;

[0066] The language recognition result indicates that the language types of the target interactive content (which can also be understood as live broadcast content) are all mainstream languages ​​(for example, Chinese), or the languages ​​of the target interactive content include both Chinese and English.

[0067] S306, receiving the language recognition result, and calling the API to translate the text into the main review language;

[0068] It should be noted that when the language recognition results indicate that the language types of the target interactive content are all mainstream languages, there is no need to call the API for content translation. Only when the language recognition results indicate that the language types of the target interactive content include language types other than mainstream languages, the API will be called to translate the text content in the non-mainstream language into the mainstream language.

[0069] S308, analyzing the emotional tendency of the target interactive content processed by the translation module;

[0070] Among them, the interactive content input into the sentiment analysis module is multimodal data, for example, including video, audio, text and pictures, etc. The sentiment analysis module preliminarily determines whether the sentiment type of the content to be reviewed is positive, negative, neutral, etc.

[0071] The system performs sentiment analysis on the translated text to identify possible inappropriate content in the text. The basic content of sentiment analysis includes:

[0072] (1) Sentiment dictionary matching: Use a pre-built multilingual sentiment dictionary to match the translated text and extract words with sentiment tendencies and their sentiment scores;

[0073] (2) Sentiment classification: Use sentiment classification models (such as RNN or Transformer) to classify text and determine whether the text sentiment is positive, neutral, or negative, especially to detect offensive or sensitive content;

[0074] (3) Cultural background analysis: Based on the language and cultural background of the text, the cultural metaphors or puns that may exist in the text are analyzed to improve the accuracy of sentiment analysis.

[0075] S310, analyzing cultural metaphors or sensitive content in the text according to the user's geographical location and cultural background database;

[0076] Due to different cultural backgrounds, language expressions and habits are also different. Therefore, after translating the multilingual interactive content into interactive content in a unified mainstream language, in order to avoid inaccurate understanding of the translated text content by the sentiment analysis module, the embodiment of the present application will combine the user's geographical location and cultural background to further obtain the deep meaning of the interactive content.

[0077] It should be noted that cultural background recognition is to identify cultural factors that may affect content review by combining the cultural background of the user's region. This cultural factor includes: (1) Geographic location association: Through the user's IP address or GPS data, the system determines the user's geographic location and associates it with the corresponding cultural background; (2) Cultural background database: The system uses a pre-built cultural background database, which contains information such as cultural habits, language characteristics, and religious beliefs in different regions; (3) Cultural feature extraction: The system extracts cultural features related to the user's region from the database and applies them to text review to avoid misjudgment.

[0078] S312, dynamically adjust content review strategies based on the results of sentiment analysis and cultural background analysis;

[0079] Based on the preliminary review suggestions output by the sentiment analysis module and the cultural analysis report output by the cultural background analysis module, determine whether it is necessary to adjust the values ​​of the evaluation parameters in the review strategy module. For example, lower the weight of parameter 1, which is used to determine whether the content is a sensitive word (or non-standard terminology, etc.), because the value of parameter 1 will directly affect the score of the final review system's review results.

[0080] Based on the identified cultural background, the system dynamically adjusts the content review strategy to ensure that the review standards meet the needs of users from different cultural backgrounds. The specific contents of the adjustment are as follows:

[0081] (1) Dynamic policy adjustment: The system adjusts the review policy parameters based on the cultural background recognition results, such as increasing or decreasing the review sensitivity for certain sensitive words;

[0082] (2) Real-time policy application: The system applies the adjusted review policy to the current content review process in real time to ensure that the review results meet the cultural and legal requirements of the user's region;

[0083] (3) Audit standard upgrade: The system regularly updates the cultural background database and audit strategies to cope with the ever-changing international cultural environment and laws and regulations.

[0084] S314, performing real-time content review according to the adjusted review strategy to obtain a target review result;

[0085] Among them, whether the target interactive content is compliant can be determined, but is not limited to, by comparing the value of the target audit result with a preset threshold.

[0086] For example, interactive content with a value less than 0.6 is determined as non-compliant content, and interactive content with a value less than 0.3 in the audit results is directly blocked.

[0087] It can be seen that the above embodiments can identify and filter inappropriate content in different languages ​​and cultural backgrounds, especially when reviewing internationalized bullet comments. According to different languages ​​and cultural backgrounds, the system can automatically adapt to different review standards and strategies to ensure the accuracy and efficiency of the review results.

[0088] S316, sending the content review result to the user end;

[0089] S318, synchronizing the content review results to the global synchronization module;

[0090] In this embodiment, the content review system can be built using, but is not limited to, a distributed system. Therefore, after the review is completed, the review results need to be synchronized to each server in real time, that is, step S320 is executed.

[0091] S320, synchronizes the content review results to review servers around the world to ensure the consistency of content review standards.

[0092] By adopting the above method, the live broadcast content containing multiple language types is translated and converted into target interactive content of unified language type. Then, based on the target interactive content, the multimodal live broadcast data is combined to obtain a fused semantic vector used to represent the emotional information of the live broadcast content. Finally, by combining the fused semantic vector with the regional cultural description vector of the user's location, the model can fully understand the contextual meaning and cultural background knowledge of the content to be reviewed, thereby adjusting the evaluation parameters in the content review model in real time and generating the target review results. In other words, by translating the live broadcast content in the multilingual live broadcast scene, fusing multimodal data, understanding the context and identifying the cultural background, it is possible to accurately identify the live broadcast content that does not meet the requirements, thereby achieving the technical effect of improving the accuracy of the content recognition results.

[0093] As an optional example, the above method obtains a fused semantic vector by inputting the target interaction content into a multimodal semantic recognition model, including:

[0094] Extracting features of video interaction data in the multimodal interaction data to obtain video interaction features, wherein the multimodal interaction data includes video interaction data, audio interaction data, and text interaction data, and the target interaction content includes the multimodal interaction data;

[0095] Extract features from audio interaction data to obtain audio interaction features;

[0096] Extract features from text interaction data to obtain text interaction features;

[0097] By fusing video interaction features, audio interaction features and text interaction features, and inputting the fused interaction features into a multimodal semantic recognition model, a fused semantic vector is obtained.

[0098] like Figure 4 As shown in the figure, the target interactive content after translation can be analyzed through a multimodal sentiment and semantic analysis model, which can simultaneously process and fuse multimodal data from video, audio and text. The model uses a multi-head attention mechanism and Transformer architecture to perform deep feature extraction on different modal data, and dynamically optimizes the contribution of each modality to sentiment recognition through an adaptive weight adjustment strategy.

[0099] The specific steps are as follows:

[0100] S402, data preprocessing;

[0101] Among them, the objects of data preprocessing include video interaction data, audio interaction data, text interaction data, etc. in multimodal interaction data.

[0102] S404, Processing of video interaction data;

[0103] In this embodiment, it is possible but not limited to denoising and frame extraction of video interaction data. Specifically, the OpenCV library is used to denoise the live video stream, extract key frames, and standardize the video frame size. The specific processing steps are as follows:

[0104] 1) Denoising processing: Remove the noise in the video through median filtering or Gaussian filtering methods to improve the picture clarity;

[0105] 2) Key frame extraction: Use a scene change detection algorithm (such as color histogram change) to extract key frames in the video;

[0106] 3) Standardization processing: Adjust all video frames to a unified resolution and ratio for subsequent feature extraction.

[0107] S406, Processing of audio interaction data;

[0108] It is possible but not limited to performing spectral analysis on audio data. Specifically, the Librosa library is used to perform spectral analysis on the audio stream and extract MFCC (Mel Frequency Cepstral Coefficients) features. The specific steps are as follows:

[0109] 1) Audio segmentation: Segment the audio stream according to a fixed time window to ensure that each segment of audio data is long enough for spectral analysis;

[0110] 2) Spectral conversion: Perform Fourier transform on the segmented audio signal to convert it into a spectrogram;

[0111] 3) MFCC feature extraction: Extract the spectral information by calculating the MFCC features of the audio signal for sentiment analysis of audio content.

[0112] S408, Processing of text interaction data;

[0113] Perform word segmentation and sentiment vocabulary extraction on the text data. Specifically, use the NLTK library to perform word segmentation on real-time bullet screens and comments, and extract sentiment vocabulary features. The specific steps are as follows:

[0114] 1) Word segmentation processing: Use the word segmentation tool of NLTK to perform word segmentation on bullet screen and comment texts to break down meaningful word units;

[0115] 2) Stop word filtering: Remove common stop words (such as "de", "shi", etc.) and retain words meaningful for sentiment analysis;

[0116] 3) Sentiment vocabulary extraction: Use sentiment dictionaries (such as SentiWordNet) to annotate the word segmentation results and extract sentiment-oriented words and their sentiment scores.

[0117] S410, standardized feature data;

[0118] It can be understood, but is not limited to, as processing each modality data in the multimodal data into standard unified feature data.

[0119] S412, extracting features from the standardized multimodal data;

[0120] (1) Feature Extraction of Video Interaction Data

[0121] In this embodiment, a convolutional neural network (CNN) may be used, but is not limited to, to extract visual features of a video frame, that is, a pre-trained ResNet model is used to extract visual features of a video frame and generate a feature vector. The specific steps are as follows:

[0122] (a) Loading the ResNet model: Load the pre-trained ResNet-50 model and input the extracted key frames into the model;

[0123] (b) Feature vector generation: The convolutional layer of the ResNet model is used to extract the high-dimensional feature vector of each frame image, which represents the visual features of the image.

[0124] (c) Feature pooling: Pooling operation is performed on the feature vector of each frame to generate a feature vector of fixed length for subsequent processing.

[0125] (2) Feature Extraction of Audio Interaction Data

[0126] In this embodiment, the spectral features of the audio signal can be extracted using, but not limited to, a long short-term memory network LSTM model to generate a feature vector. The specific steps are as follows:

[0127] (a) Input spectrum data: The preprocessed MFCC features are used as the input of the LSTM model;

[0128] (b) Time series feature extraction: The time series information of the audio data is processed using the recurrent unit of the LSTM model to extract the time series features of the audio signal;

[0129] (c) Feature vector generation: The output of the LSTM model is the feature vector of the audio signal, which represents the temporal pattern and spectral characteristics of the audio.

[0130] (3) Feature Extraction of Text Interaction Data

[0131] In this embodiment, the Transformer model can be used, but is not limited to, to extract sentiment and semantic features of text content and generate feature vectors. The specific steps are as follows:

[0132] (a) Word segmentation input: The text sequence after word segmentation is input into the pre-trained BERT model;

[0133] (b) Sentiment and semantic extraction: Through the multi-layer attention mechanism of the BERT model, the sentiment and semantic features of each word are extracted, and the context feature vector of the entire sentence is generated;

[0134] (c) Feature vector generation: The output vector of the BERT model is used as the feature vector of the text for subsequent multimodal fusion.

[0135] S414, inputting video, audio and text features into the multi-head attention mechanism for feature fusion to generate a comprehensive sentiment semantic vector;

[0136] The features of different modalities are input into the multi-head attention mechanism for fusion to generate a unified sentiment semantic vector. The influence of each modality in sentiment recognition is dynamically adjusted through an adaptive weight strategy. The specific steps are as follows:

[0137] (1) Multimodal input: visual, audio, and text feature vectors are simultaneously input into the multi-head attention mechanism;

[0138] (2) Weight calculation: The adaptive weight of each modality feature is calculated through the attention mechanism, and weighted according to the importance of different modalities;

[0139] (3) Feature fusion: The weighted features of each modality are fused to generate a unified comprehensive sentiment semantic vector.

[0140] S416, generating a sentiment semantic vector;

[0141] Among them, the emotional semantic vector can be understood as a fusion semantic vector, which mainly uses an adaptive weight strategy to dynamically adjust the influence of each modality in emotion recognition to generate the final emotional semantic vector, which does not include:

[0142] (1) Weight adaptive adjustment: According to the changes in real-time data, the weight of each modality is adaptively adjusted to optimize the accuracy of emotion recognition;

[0143] (2) Emotional semantic vector output: The fused emotional semantic vector is used for subsequent review decisions and represents the comprehensive emotional state of the content.

[0144] S418, deploying the trained multimodal sentiment and semantic analysis model to the cloud or edge server;

[0145] (1) Model optimization: Optimize the trained model to reduce computing overhead and adapt to the cloud or edge computing environment;

[0146] (2) Model deployment: Deploy the model to the server through a Docker container or Kubernetes cluster to support large-scale real-time processing of requests.

[0147] S420, performing real-time sentiment and semantic analysis on the live broadcast content to generate preliminary real-time review results.

[0148] (1) Real-time data input: input the live video, audio, and text data streams into the deployed model;

[0149] (2) Sentiment analysis: The model generates sentiment and semantic analysis results of live broadcast content in real time, including whether there is a risk of inappropriate content;

[0150] (3) Audit decision output: Based on the analysis results, audit decisions are generated in real time, and the system automatically takes corresponding audit measures, such as blocking inappropriate content or warning users.

[0151] By performing sentiment analysis on the translated interactive content, we can identify inappropriate content that may exist in the target interactive content. Sentiment analysis mainly includes the following aspects:

[0152] (1) Sentiment dictionary matching: Use a pre-built multilingual sentiment dictionary to match the translated text and extract words with sentiment tendencies and their sentiment scores;

[0153] (2) Sentiment classification: Use sentiment classification models (such as RNN or Transformer) to classify text and determine whether the text sentiment is positive, neutral, or negative, especially to detect offensive or sensitive content;

[0154] (3) Cultural background analysis: Based on the language and cultural background of the text, the cultural metaphors or puns that may exist in the text are analyzed to improve the accuracy of sentiment analysis.

[0155] It should be noted that the result of predicting real-time data using the multimodal sentiment and semantic analysis model is the sentiment analysis result, which is a preliminary audit result, mainly including positive, negative and neutral, etc. The model can be used to understand the comprehensive sentiment and semantics of multimodal data to ensure the accuracy of the basic direction of the audit results.

[0156] As an optional example, the above adjustment of the evaluation parameters in the content review model based on the fusion semantic vector and the regional cultural description vector includes:

[0157] Obtaining the account behavior data corresponding to the account that published the target interactive content and the regional description data of the location of the client logged in by the account;

[0158] By inputting the account behavior data and the regional description data into the target perception model, a regional cultural description vector for representing the current status of the account is obtained;

[0159] Based on the sentiment type of the target interactive content represented by the fused semantic vector and the deep meaning represented by the regional cultural description vector, the evaluation parameters in the content audit model are adjusted.

[0160] In the embodiment of the present application, more abundant data can be obtained by (1) collecting context data, device information and environmental factors of the target interactive content to be reviewed; (2) performing context analysis on the data in step (1) to obtain a context vector as input for subsequent generation of personalized review strategies; (3) through reinforcement learning, the content review system can generate a personalized review strategy based on the context vector and user behavior pattern. Each step in steps (1) to (3) is described below.

[0161] (1) Contextual Data Collection

[0162] S21, collection of user behavior data;

[0163] Through the data collection module, the system collects users' historical behavior data, including but not limited to viewing history, interaction records, likes and comments, etc.

[0164] 1) Behavior data recording: The system records all user operations on the live broadcast platform in real time, including clicks, views, sharing, etc., and stores these data in the user behavior database;

[0165] 2) Data storage and management: After preprocessing, behavioral data is stored in a distributed database. The system indexes and classifies the data according to user ID for subsequent analysis.

[0166] S22, collection of device information;

[0167] The system further improves the user portrait by collecting hardware and software information of the user's device.

[0168] 1) Device information collection: The system collects information such as the user's device hardware configuration (such as CPU, memory, screen resolution) and operating system version;

[0169] 2) Network environment detection: The system monitors the network connection status of the user's device in real time, including network type (Wi-Fi, 4G / 5G), signal strength and bandwidth, to evaluate the quality of the live broadcast experience;

[0170] 3) Device information storage: The collected device information is stored in the user database together with the user behavior data as the input of the context-aware model.

[0171] S23, collection of environmental factors;

[0172] The system collects information such as the user's geographic location, time, weather, and social environment through sensors and other external data sources.

[0173] 1) Geographic location acquisition: The system obtains the user's geographic location information through the GPS module or IP address to analyze the user's environmental background;

[0174] 2) Time and weather data: The system obtains real-time weather information and time information in the user's area through external APIs, and performs environmental analysis based on geographic location;

[0175] 3) Social environment data: The system obtains the user’s social environment data through social media platforms or other data sources, including the user’s friend dynamics, hot topics and other information.

[0176] (2) Contextual Analysis

[0177] S31, context vector generation;

[0178] The system uses a context-aware model to convert the collected data into context vectors for the generation of personalized audit strategies.

[0179] 1) Data cleaning and processing: The system cleans and processes the collected user behavior data, device information, and environmental factor data to remove noise data and redundant information;

[0180] 2) Feature extraction: The system extracts features from the processed data through a context-aware model to generate a context feature vector representing the user’s current state;

[0181] 3) Context vector generation: All features are integrated into a context vector as input for subsequent personalized review strategy generation.

[0182] S32, user portrait combination;

[0183] The system combines user portrait data to further analyze the context vector to generate more accurate personalized review strategies.

[0184] 1) User portrait acquisition: The system obtains user portrait information from the user database, including basic information, interest preferences, historical behavior patterns, etc.

[0185] 2) Portrait and context fusion: The system fuses user portrait information with the context vector generated in real time to form a real-time behavior model of the user.

[0186] 3) Behavioral pattern analysis: The system uses machine learning algorithms to analyze user behavior patterns, predict the user’s next action, and incorporate this information into the generation process of personalized review strategies.

[0187] Based on the generated context vector and combined with the user portrait, a reinforcement learning algorithm is used to generate personalized review strategies based on the context vector and user behavior patterns.

[0188] As an optional implementation method, the above-mentioned adjustment of the evaluation parameters in the content audit model based on the emotional type of the target interactive content represented by the fusion semantic vector and the deep meaning represented by the regional cultural description vector includes:

[0189] When the fused semantic vector indicates that the target interactive content has the first type of sentiment and the deep meaning is in compliance with the regional cultural description norm, the value of the evaluation parameter is increased, wherein the higher the value of the evaluation parameter, the less likely it is that the part of the target interactive content with special meaning will be reviewed as non-compliant content;

[0190] When the fused semantic vector indicates that the emotion type of the target interactive content is the second type and the deep meaning does not conform to the regional cultural description norm, the value of the evaluation parameter is reduced;

[0191] When the fused semantic vector indicates that the emotion type of the target interactive content is the first type and the deep meaning does not conform to the regional cultural description norm, the value of the evaluation parameter is reduced.

[0192] In an embodiment of the present application, a reinforcement learning algorithm (such as Q-learning) is used to adaptively adjust the review strategy based on the user's real-time feedback and behavior patterns.

[0193] The specific steps for generating an audit policy are as follows:

[0194] S41, state space definition: The system defines the user's behavior data, device information and environmental factors as the state space, which is used as the input of the reinforcement learning model;

[0195] S42, action selection: the system selects the optimal audit strategy action based on the current status and historical behavior, such as content filtering, prompt information display or interactive recommendation;

[0196] S43, Reward Mechanism: The system gives rewards or penalties based on the execution effect of the audit strategy (such as user satisfaction, content audit accuracy, etc.) to optimize the strategy generation model;

[0197] S44, Strategy Update: The reinforcement learning model updates its strategy based on the reward signal obtained to improve the accuracy and personalization of future review decisions.

[0198] The specific steps for adaptively adjusting the audit strategy based on real-time user feedback are as follows:

[0199] S51, real-time feedback collection: the system collects user feedback information on the review strategy in real time, such as the click rate of content recommendations, user stay time, etc.

[0200] S52, Strategy Adjustment: Based on user feedback, the system dynamically adjusts the review strategy through a reinforcement learning model, including adjusting content recommendation rules, filtering sensitive information, etc.

[0201] S53, Continuous Learning: The system continuously updates and optimizes the audit policy model through continuous learning to ensure the effectiveness of the model in long-term use.

[0202] The system dynamically adjusts the review strategy based on real-time data analysis results to optimize user experience and content security. The specific steps are as follows:

[0203] 1) Dynamic strategy generation: The system generates new audit strategies based on real-time data analysis results, including adjusting content filtering standards and optimizing recommendation algorithms;

[0204] 2) Strategy execution and feedback: The system will apply the generated strategy immediately and make fine adjustments based on user feedback to ensure the effectiveness of the audit strategy and improve user experience.

[0205] In order to more clearly understand the overall implementation process of the above context-aware and personalized audit strategy, Figure 5 The flowchart shown further describes this.

[0206] S502, collecting device information;

[0207] The system further improves the user portrait by collecting the hardware and software information of the user's device. The collected device information includes but is not limited to the hardware configuration of the user's device (such as CPU, memory, screen resolution) and operating system version.

[0208] S504, collecting user behavior data;

[0209] The system records all user operations on the live broadcast platform in real time, including clicks, views, sharing, etc., and stores this data in the user behavior database.

[0210] S506, collecting environmental data;

[0211] The system collects information such as the user's geographic location, time, weather, and social environment through sensors and other external data sources.

[0212] S508, contextual analysis;

[0213] It mainly includes two processes: context vector generation and system collection of user portrait data. Among them, the system uses a context-aware model to convert the collected data into a context vector for the generation of personalized audit strategies, refer to step S510; the system combines the user portrait data to further analyze the context vector to generate a more accurate personalized audit strategy, refer to step S512.

[0214] S510, using a context-aware model to convert the collected data into a context vector for generating a personalized audit strategy;

[0215] The specific steps of generating the context vector can refer to the description in the above embodiment and will not be repeated here.

[0216] S512, the system further analyzes the context vector in combination with the user portrait data to generate a more accurate personalized review strategy;

[0217] S514, using a reinforcement learning algorithm to generate a personalized audit strategy based on the context vector and the user behavior pattern;

[0218] It mainly includes processes such as state space definition, action selection, reward mechanism and strategy update. For details, please refer to the description in the above embodiment, which will not be repeated here.

[0219] S516, dynamically adjust the audit strategy based on user feedback.

[0220] The method of dynamically adjusting the audit strategy includes two parts: data flow analysis and strategy adjustment. The data flow analysis content audit system analyzes the user's data flow in real time and dynamically adjusts the audit strategy according to user behavior and environmental changes. The specific steps are as follows:

[0221] 1) Data flow monitoring: The system monitors the user's real-time behavior data and identifies key nodes and change trends in the data flow;

[0222] 2) Behavior prediction: The system predicts the user's next action through machine learning algorithms and adjusts the review strategy in advance based on the prediction results;

[0223] 3) Real-time policy application: The system will immediately apply the adjusted audit policy to the current data stream to ensure the real-time and accuracy of content audit.

[0224] Policy adjustment means that the system dynamically adjusts the audit strategy based on real-time data analysis results to optimize user experience and content security. The specific steps are as follows:

[0225] 1) Dynamic strategy generation: The system generates new audit strategies based on real-time data analysis results, including adjusting content filtering standards and optimizing recommendation algorithms;

[0226] 2) Strategy execution and feedback: The system will apply the generated strategy immediately and make fine adjustments based on user feedback to ensure the effectiveness of the audit strategy and improve user experience.

[0227] In a specific embodiment, it is assumed that the evaluation parameters in the content review system are adjusted according to the preliminary review suggestions and cultural analysis report output by the cultural background analysis module. The preliminary review results include blocking the target interactive content, adjusting the target interactive content, and requiring the user to modify the target interactive content; the cultural analysis report includes but is not limited to a report formed by combining the allusions, idioms or other content understanding in the regional cultural background to obtain the deep meaning of the target interactive content.

[0228] It should be noted that the input of the cultural background analysis module includes: (1) text data, pictures, video data, etc.; (2) user information: region, ethnicity, etc.; (3) cultural norms: for example, sensitive words, grammatical information, etc.

[0229] When the fused semantic vector indicates that the emotion type of the target interactive content is positive, and the regional cultural description vector indicates that the deep meaning of the target interactive content conforms to the regional cultural description norms, the value of the evaluation parameter in the content review system is increased, for example, the value of evaluation parameter 1 is increased, or the weight of evaluation parameter 1 is increased, so that the final review result of the content review system is adjusted to tend to a high score.

[0230] For another example, when the fused semantic vector indicates that the emotion type of the target interactive content is negative, and the regional cultural description vector indicates that the deep meaning of the target interactive content does not conform to the regional cultural description standards, the value of the evaluation parameter 2 in the content review system is reduced, or the weight of the evaluation parameter 1 is reduced, so that the final review result of the content review system is adjusted to tend towards a low score.

[0231] For example, when the fused semantic vector indicates that the target interactive content has a positive sentiment type, but the regional cultural description vector indicates that the deep meaning of the target interactive content does not conform to the regional cultural description norms, it means that the literal meaning of the target interactive content meets the audit requirements, but after in-depth understanding based on the regional cultural background knowledge, the possibility of determining that the target interactive content contains sensitive words or is non-compliant has increased. Therefore, the value of the evaluation parameter is reduced, and through this adjustment, the final audit result of the content audit system tends to be scored in the direction of a low score.

[0232] Through the above method, a multi-level review mechanism is established, and under the action of the layered review mechanism, the content review results can be made more objective and more in line with the true meaning of the target interactive content itself, thereby improving the accuracy of the review results.

[0233] As an optional example, the above-mentioned target audit results are generated based on the adjusted content audit model, including:

[0234] Generate adversarial networks to detect target interactive content and obtain target detection results, where the generative adversarial network is built into the content review model;

[0235] When the target detection result indicates that the target interaction content does not contain non-compliant content, the target interaction content is differentially privately processed by introducing random noise to obtain the processed interaction content;

[0236] By inputting the processed interactive content into the adjusted content audit model, the target audit result is obtained.

[0237] like Figure 6 As shown in the figure, during the entire process of content review, the generative adversarial network will be used to dynamically detect the live content, so as to filter out more obvious sensitive words or non-compliant content.

[0238] S602, generating realistic forged content;

[0239] In this embodiment, a specific GAN structure is designed, the generator generates realistic forged content, and the discriminator identifies real and forged content. Through continuous adversarial training, the detection ability of the discriminator is improved.

[0240] Among them, the generator network of GAN is used to generate adversarial samples to train the discriminator network and improve the robustness and accuracy of the detection model. The specific steps are as follows:

[0241] S61, Random Noise Input: The generator network accepts a random noise vector as input to generate fake content;

[0242] Here, the noise vector usually comes from a standard normal distribution.

[0243] S62, Content Generation: The generator uses a series of convolutional and deconvolutional layers to transform the input noise vector into a realistic fake content sample;

[0244] For example, forging video frames, audio clips, or text snippets.

[0245] S63, content output: The generated forged content samples are used to train the discriminator to improve the discriminator's ability to identify real content and forged content.

[0246] S604, identifying real and fake content;

[0247] The discriminator is used to judge the authenticity of the input content and improve its recognition ability through continuous learning. The specific steps include:

[0248] S71, content input: the discriminator receives real content and forged content generated by the generator as input;

[0249] The input content includes video frames, audio signals and text data.

[0250] S72, discrimination process: The discriminator extracts the features of the input content through the convolution layer, classifies it through the fully connected layer, and outputs the true or false discrimination result.

[0251] S73, discriminant optimization: The discriminator continuously adjusts weight parameters through the back-propagation algorithm to improve the ability to identify forged content.

[0252] S606, cyclic adversarial training;

[0253] Among them, cyclic adversarial training includes two parts: sample generation and detection, and model optimization. These parts are described below respectively.

[0254] It should be noted that in the GAN model, the generator and the discriminator optimize each other through adversarial training. The specific implementation of the optimization is as follows:

[0255] S81, Adversarial Optimization: Adversarial training of the generator and the discriminator is performed by alternately optimizing the loss functions of the two models. The generator tries to generate more realistic fake content to deceive the discriminator, while the discriminator continuously improves its ability to distinguish between real and fake content.

[0256] S82, loss function calculation: the loss function of the generator is calculated according to the discrimination result of the discriminator on the forged content, and the loss function of the discriminator is calculated according to the discrimination accuracy of the real and forged content;

[0257] S83, Model training: Through multiple rounds of adversarial training, the generator and discriminator are improved together, and finally the generator is able to generate highly realistic forged content, so that the discriminator can accurately identify these forged contents.

[0258] S608, continuously generate new adversarial samples and optimize the detection model;

[0259] The generator continuously generates adversarial samples and the discriminator continuously improves its detection ability. Sample generation is to generate new adversarial samples based on the best weight parameters obtained through training. These samples are used to further train the discriminator. Sample detection is that the discriminator uses the latest generated adversarial samples for detection. By comparing the discrimination results with the actual labels, the weights of the discriminator are further optimized.

[0260] S610, the generator and discriminator are jointly optimized to improve detection capabilities;

[0261] Through training, the performance of the generator and discriminator is continuously optimized to improve the detection accuracy of the overall model, including: (1) Model iteration includes multiple rounds of cyclic training, the content generated by the generator becomes more and more realistic, and the detection accuracy of the discriminator is gradually improved; (2) Performance evaluation: After each round of training, the model performance is evaluated and the validation set data is used to detect the effects of the generator and discriminator; (3) Parameter adjustment: According to the results of the performance evaluation, the hyperparameters of the generator and discriminator are adjusted, such as the learning rate, regularization parameter, etc., to optimize the final performance of the model.

[0262] S612, real-time detection;

[0263] During the live broadcast, video streams and audio signals are captured in real time, and potential false content and manipulation behaviors are detected through the trained GAN model.

[0264] The detection process includes content capture and GAN detection. Content capture includes but is not limited to real-time capture of video streams, audio signals and text data for detection during live broadcast, such as step S612-1; GAN detection includes but is not limited to using a trained GAN model to perform real-time detection on the captured content to identify forged and abnormal content, such as step S612-2.

[0265] S612-1, real-time capture of video streams, audio signals, and text data for detection;

[0266] S612-2, detect potential virtual content through the GAN model.

[0267] S614, strategy adjustment, automatically adjusting the detection strategy when abnormal content is detected;

[0268] When abnormal content is detected, the system automatically adjusts the detection strategy to deal with new counterfeiting methods.

[0269] S616, triggering content filtering and management notification.

[0270] After identifying high-risk or fake content, the system will automatically adjust the detection strategy and trigger content filtering and management notifications to ensure the authenticity and compliance of live content, including:

[0271] (1) Content filtering: The system automatically blocks or deletes identified fake or inappropriate content to prevent it from being further spread on the live streaming platform;

[0272] (2) Notification management: The system sends an alert notification to the administrator of the live broadcast platform, detailing the type of fake content identified, the time of detection, and the treatment measures taken;

[0273] (3) Audit record archiving: The system archives all identified and processed records for subsequent audits and compliance checks.

[0274] In an embodiment of the present application, a generative adversarial network (GAN) is innovatively applied to dynamic detection of live content, and the robustness and adaptability of the detection model are improved by generating adversarial samples.

[0275] It should be noted that in the content review system in the embodiment of the present application, a multi-level federated learning mechanism is introduced (this mechanism acts on the entire processing process of the content review system), and combined with a differential privacy algorithm, so as to improve the review efficiency and security while ensuring user privacy.

[0276] The basic process of multi-level federated learning privacy protection is as follows:

[0277] (1) Local model training: Perform local model training on the user device and update model parameters to ensure that data does not leave the device;

[0278] (2) Model aggregation: Collect local model updates, perform global model aggregation through a federated learning framework, and apply differential privacy technology to protect user data;

[0279] (3) Privacy risk assessment: real-time monitoring and management of data leakage risks, and optimization of the security of data transmission and processing;

[0280] (4) Global model update: Deploy the updated global model to user devices to improve audit capabilities and user experience.

[0281] The detailed process of multi-level federated learning and privacy protection is as follows: Figure 7 As shown, including:

[0282] S702, the user turns on the device;

[0283] S704-1, collecting data through user equipment;

[0284] Collect data locally on the user's device for model training to ensure data privacy is not leaked. Data collection can be, but is not limited to, local collection of user behavior data and device information for training personalized models. Data includes user viewing history, device usage habits, etc.

[0285] S704-2, data preprocessing;

[0286] The collected data is denoised and normalized to ensure that the data quality is suitable for model training.

[0287] In addition, after preprocessing, the data also needs to be stored, specifically, the processed data is saved on the user device and the data update information is synchronized with the central server.

[0288] S706, training a local model using the preprocessed data;

[0289] (1) Model initialization: Download the initial parameters of the pre-trained model from the central server and load the model on the local device;

[0290] (2) Local training: Use the user's local data to fine-tune the model to meet the user's personalized needs. Model training uses a small batch gradient descent method to reduce the occupation of device resources;

[0291] (3) Model update: After training is completed, the updated model parameters are saved and wait for aggregation with the central server.

[0292] S708, during the iterative training of the model, adjusting the structural parameters of the model to optimize the model;

[0293] S710, start event reporting, and transform and protect differential privacy of data through the model data information exchanger;

[0294] Differential privacy technology is used in the model aggregation process to ensure the privacy and security of user data. The specific steps are as follows:

[0295] a) Noise injection: Before model aggregation, noise is injected into local model parameters to mask the user’s real data characteristics, thereby ensuring data privacy;

[0296] b) Privacy budget management: Use the privacy budget to control the intensity of injected noise to ensure a balance between privacy protection and model accuracy;

[0297] c) Data security assessment: Before the aggregation model is released, the privacy protection measures are evaluated to ensure that the user's sensitive information is not leaked.

[0298] It should be noted that the differential privacy protection process runs through all aspects of the content review system, and its purpose is to ensure the security of user data and user privacy during the content review process.

[0299] S712, model aggregation;

[0300] Model aggregation is the process of the central server receiving model update data uploaded by each user device through the federated learning framework and performing global model aggregation. The specific steps are as follows:

[0301] a) Model parameter collection: The central server receives model parameter updates after local training from each user device, ensuring that data is not uploaded and only parameter information is transmitted;

[0302] b) Model aggregation: Use a weighted average algorithm to aggregate the collected model parameters to generate updated parameters for the global model. The weights are dynamically adjusted based on the amount of data and computing power of each user device;

[0303] c) Aggregate model validation: An independent validation dataset is used on a central server to evaluate the aggregated global model to ensure improved model performance.

[0304] S714, merging each sub-model built into the content review system to obtain a global model;

[0305] Among them, the models built into the content review system include but are not limited to context-aware models, generative pair networks, multimodal sentiment and semantic analysis models, etc.

[0306] S716, determine whether the privacy policy is passed;

[0307] If passed, execute step S718; otherwise, execute step S708.

[0308] S718, global model deployment and update;

[0309] The aggregated global model is deployed to each user device and updated in real time. The process is as follows:

[0310] a) Model publishing: The central server publishes the global model protected by differential privacy to the user's device. The model is transmitted through a secure connection (such as HTTPS) to ensure the security of the transmission process;

[0311] b) Local model update: After receiving the new global model, the user device replaces the old model parameters to ensure that the audit system always uses the latest optimized model;

[0312] c) Deployment monitoring: The central server continuously monitors the deployment of the global model and records the update status of the device and the model usage effect.

[0313] S720, transfers data from the AI ​​big model to the big data platform through a preset data collection interface;

[0314] Among them, the big data platform can be used, but is not limited to, for privacy risk monitoring and management of user data.

[0315] Specifically, the system ensures the security of user data during federated learning by real-time monitoring and managing privacy risks. The processing process is as follows:

[0316] a) Real-time monitoring: The system monitors abnormal behaviors during model parameter transmission in real time and records potential security threats;

[0317] b) Risk warning: The system issues warnings for detected privacy risks and notifies platform administrators to take corresponding measures, such as increasing encryption strength or reallocating privacy budgets;

[0318] c) Update of data protection measures: Based on the latest security research results, the system regularly updates privacy protection measures to ensure that it always stays at the forefront in data security.

[0319] S722, Flume collects data;

[0320] Flume is a highly available, highly reliable, distributed system for collecting, aggregating, and transmitting massive amounts of logs. It supports the customization of various types of data senders in the log system to collect data. At the same time, Flume provides the ability to perform simple data processing and write data to various data recipients.

[0321] S724, real-time monitoring of the collected data;

[0322] S726, during the real-time monitoring process of the system, abnormal behaviors during the model parameter transmission process are obtained and potential security threats are recorded;

[0323] S728, determining whether to increase encryption strength or reallocate privacy budget based on the monitored abnormal behavior;

[0324] S730: Issue a security warning message based on the monitored abnormal behavior.

[0325] By introducing an adaptive data stream processing algorithm, the processing strategy is dynamically adjusted according to data flow and system load. Specifically, the machine learning model is used to predict data flow, and the processing strategy is dynamically adjusted based on the prediction results.

[0326] The embodiment of the present application introduces a multi-level federated learning mechanism for the first time in the content review system, combined with a differential privacy algorithm, to ensure user privacy while improving review efficiency and security.

[0327] As an optional implementation method, the above-mentioned use of a generative adversarial network to detect the target interactive content to obtain a target detection result includes:

[0328] Perform word segmentation and encoding processing on the text data in the target interactive content to obtain processed data features;

[0329] By inputting the processed data features into the discriminator in the generative adversarial network, the target detection result is obtained.

[0330] When the target interactive content is audio data or video data, it needs to be recognized as text data first.

[0331] For the above step S612-1, the specific steps of real-time capturing of video streams, audio signals and text data for detection are as follows:

[0332] (1) Real-time stream capture: Use streaming media processing technology to capture video frames, audio clips, and text data such as bullet comments in real time during live broadcasts;

[0333] (2) Data preprocessing: Standardization of real-time captured data, including video frame resizing, audio signal denoising, and text data segmentation and encoding;

[0334] (3) Data stream allocation: The preprocessed data stream is passed to the GAN detection model for real-time detection.

[0335] For the above step S612-2, potential virtual content is detected through the GAN model, and the specific implementation process is as follows:

[0336] (1) Data input: The preprocessed video, audio, and text data are input into the discriminator of the GAN model for analysis;

[0337] (2) Forgery detection: The GAN discriminator analyzes the input data and outputs the detection results to determine whether the content is forged or abnormal;

[0338] (3) Dynamic response: Based on the detection results, the system will dynamically adjust the subsequent content review process to ensure the authenticity and legality of the live broadcast content.

[0339] As an optional implementation, after obtaining the target detection result by inputting the processed data features into the discriminator in the generative adversarial network, the method further includes:

[0340] When the target detection result indicates that the target interactive content has an abnormal phenomenon, an abnormal mark is added to the part of the target interactive content that has the abnormal phenomenon;

[0341] The weight parameters in the discriminator are optimized to obtain the optimized generative adversarial network.

[0342] Regarding the above step S614, when abnormal content is detected, the system automatically adjusts the detection strategy to deal with new counterfeiting methods. The specific implementation method is as follows:

[0343] (1) Abnormal content identification: When the discriminator identifies abnormal phenomena or signs of forgery in the content, the system will mark the content as high-risk content;

[0344] (2) Detection strategy adjustment: The system dynamically adjusts the detection strategy based on the marked high-risk content types, including optimizing the weight parameters of the discriminator and adding new detection rules.

[0345] From the description of the above embodiments, it can be seen that in the embodiments of the present application, a generative adversarial network (GAN) is innovatively applied to dynamic detection of live content, and the robustness and adaptability of the detection model are improved by generating adversarial samples.

[0346] As an optional example, when the target audit result indicates that the target interaction content contains non-compliant content, the processing of the target interaction content includes at least one of the following: blocking the target interaction content, adjusting the target interaction content, and modifying the target interaction content.

[0347] Among them, the adjustment of the target interactive content includes but is not limited to fine-tuning of the position, state, quantity, etc., and the modification may be but is not limited to changing the content quantity, content structure, etc.

[0348] It can be seen from the description of the above embodiments that the technical solution in the embodiments of the present application is used to process real-time data, which has at least the following beneficial effects:

[0349] (1) The embodiments of the present application provide a multi-language support and cross-cultural review mechanism that can meet the content review requirements of global live broadcast platforms and can identify and filter non-compliant content in different languages ​​and cultural backgrounds;

[0350] (2) By using a multi-head attention mechanism and combining it with the Transformer architecture, we can perform deep feature extraction and sentiment fusion on different modal data, and dynamically optimize the contribution of each modality to sentiment recognition by introducing an adaptive weight adjustment strategy, thereby achieving accurate recognition of complex emotions and implicit semantics. This solves the shortcomings of existing systems in terms of sentiment and semantic understanding, and improves the ability to recognize complex emotional expressions and potential threats in live content.

[0351] (3) By introducing a generative adversarial network, live content is dynamically detected, and a specific GAN structure is designed to generate adversarial samples for training and optimizing the detection model, thereby improving the detection accuracy and real-time response capability of deep fake content. At the same time, the generation of adversarial samples can simulate possible fake content, making the model more robust and adaptable in actual references, so that high-risk content can be identified and filtered in real time during live broadcasts, ensuring the authenticity and compliance of the content. Youxia Oh solves the shortcomings of existing systems in terms of real-time performance and accuracy;

[0352] (4) The embodiment of the present application introduces a context-aware mechanism, combines user behavior data and environmental information, and formulates a personalized review strategy. The system uses a reinforcement learning algorithm to adaptively update the review strategy based on the user's real-time feedback and behavior patterns, and achieves accurate judgment and processing of different content contexts. By dynamically adjusting the review rules, the system can adapt to the needs of different users and scenarios, improving the flexibility of the review and user satisfaction. This makes up for the shortcomings of the existing system in terms of personalization and dynamic adjustment capabilities;

[0353] (5) In the content review system of the embodiment of the present application, a multi-level federated learning mechanism is introduced for the first time, combined with a differential privacy algorithm, to ensure that user privacy is fully protected while achieving distributed data processing. The system designs a multi-level federated learning framework that reduces the risk of privacy leakage and enhances the security and efficiency of data processing by processing user data and model updates in a layered manner. The framework can flexibly switch between user devices and servers to optimize data transmission and privacy protection. It solves the shortcomings of existing systems in terms of privacy and security issues.

[0354] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0355] According to another aspect of the embodiment of the present application, there is also provided Figure 8 A content review device is shown, the device comprising:

[0356] A first acquisition unit 802 is used to acquire target interactive content in the content interaction platform;

[0357] A first processing unit 804 is configured to obtain a fused semantic vector by inputting the target interaction content into a multimodal semantic recognition model, wherein the fused semantic vector is used to indicate emotional description information of the target interaction content;

[0358] An adjustment unit 806 is used to adjust the evaluation parameters in the content audit model based on the fused semantic vector and the regional cultural description vector to obtain an adjusted content audit model, wherein the evaluation parameters represent the degree of influence of the part of the content with special meaning on the audit result, and the regional cultural description vector represents the deep meaning obtained by deeply mining the target interactive content in combination with the regional cultural information;

[0359] The second processing unit 808 is used to generate a target audit result based on the adjusted content audit model.

[0360] Optionally, the first processing unit 804 includes:

[0361] A first processing module is used to extract features from video interaction data in the multimodal interaction data to obtain video interaction features, wherein the multimodal interaction data includes video interaction data, audio interaction data and text interaction data, and the target interaction content includes the multimodal interaction data;

[0362] The second processing module is used to extract features from the audio interaction data to obtain audio interaction features;

[0363] The third processing module is used to extract features from the text interaction data to obtain text interaction features;

[0364] The fourth processing module is used to obtain a fused semantic vector by fusing the video interaction features, the audio interaction features and the text interaction features, and inputting the fused interaction features into a multimodal semantic recognition model.

[0365] Optionally, the adjusting unit 806 includes:

[0366] A first acquisition module is used to acquire account behavior data corresponding to the account that published the target interactive content and regional description data of the location of the client logged in by the account;

[0367] A fifth processing module, configured to obtain a regional cultural description vector for representing a current state of an account by inputting the account behavior data and the regional description data into a target perception model;

[0368] The adjustment module is used to adjust the evaluation parameters in the content audit model based on the emotional type of the target interactive content represented by the fusion semantic vector and the deep meaning represented by the regional cultural description vector.

[0369] Optionally, the adjustment module includes:

[0370] The first processing submodule is used to increase the value of the evaluation parameter when the fused semantic vector indicates that the emotion type of the target interactive content is the first type and the deep meaning indicates that it conforms to the regional cultural description specification, wherein the higher the value of the evaluation parameter, the less likely it is that the part of the target interactive content with special meaning will be reviewed as non-compliant content;

[0371] The second processing submodule is used to reduce the value of the evaluation parameter when the fused semantic vector indicates that the emotion type of the target interactive content is the second type and the deep meaning does not conform to the regional cultural description norm;

[0372] The third processing submodule is used to reduce the value of the evaluation parameter when the fused semantic vector indicates that the emotion type of the target interactive content is the first type and the deep meaning does not conform to the regional cultural description standard.

[0373] Optionally, the second processing unit 808 includes:

[0374] A detection module is used to detect the target interactive content using a generative adversarial network to obtain a target detection result, wherein the generative adversarial network is built into the content review model;

[0375] A sixth processing module, configured to perform differential privacy processing on the target interaction content by introducing random noise to obtain processed interaction content when the target detection result indicates that the target interaction content does not contain non-compliant content;

[0376] The seventh processing module is used to obtain a target audit result by inputting the processed interactive content into the adjusted content audit model.

[0377] Optionally, the detection module includes:

[0378] The fourth processing submodule is used to perform word segmentation and encoding processing on the text data in the target interactive content to obtain processed data features;

[0379] The fifth processing submodule is used to obtain the target detection result by inputting the processed data features into the discriminator in the generative adversarial network.

[0380] Optionally, the above device further includes:

[0381] A third processing unit is used to add an abnormality mark to the part of the target interactive content that has the abnormality when the target detection result indicates that the target interactive content has an abnormality;

[0382] The optimization unit is used to optimize the weight parameters in the discriminator to obtain an optimized generative adversarial network.

[0383] Optionally, the above-mentioned device is also used for processing the target interactive content when the target audit result indicates that the target interactive content contains non-compliant content, including at least one of the following: blocking the target interactive content, adjusting the target interactive content, and modifying the target interactive content.

[0384] By applying the above-mentioned device to translate live broadcast content containing multiple language types, the live broadcast content is converted into target interactive content with a unified language type, and then based on the target interactive content, the multimodal live broadcast data is combined to obtain a fused semantic vector for representing the emotional information of the live broadcast content. Finally, by combining the fused semantic vector with the regional cultural description vector of the user's location, the model can fully understand the contextual meaning and cultural background knowledge of the content to be reviewed, thereby adjusting the evaluation parameters in the content review model in real time and generating the target review results. In other words, by performing translation processing, multimodal data fusion, context understanding and cultural background recognition on the live broadcast content in the multilingual live broadcast scene, it is possible to accurately identify the live broadcast content that does not meet the requirements, thereby achieving the technical effect of improving the accuracy of the content recognition results.

[0385] It should be noted that the embodiment of the content review device here can refer to the embodiment of the above-mentioned content review method, which will not be repeated here.

[0386] According to another aspect of the embodiment of the present application, an electronic device for implementing the above-mentioned content review method is also provided. The electronic device may be Figure 1 The target terminal or server shown in FIG. This embodiment is described by taking the electronic device as the target terminal as an example. Fig. 9 As shown, the electronic device includes a memory 902 and a processor 904. The memory 902 stores a computer program, and the processor 904 is configured to execute the steps in any of the above method embodiments through the computer program.

[0387] Optionally, in this embodiment, the electronic device may be located in at least one network device among a plurality of network devices of a computer network.

[0388] Optionally, in this embodiment, the processor may be configured to perform the following steps through a computer program:

[0389] S1, obtaining target interactive content in the content interactive platform;

[0390] S2, by inputting the target interaction content into a multimodal semantic recognition model, a fused semantic vector is obtained, wherein the fused semantic vector is used to indicate the emotional description information of the target interaction content;

[0391] S3, based on the fusion semantic vector and the regional cultural description vector, the evaluation parameters in the content audit model are adjusted to obtain the adjusted content audit model, where the evaluation parameters represent the degree of influence of some content with special meanings on the audit results, and the regional cultural description vector represents the deep meaning obtained by deeply mining the target interactive content in combination with regional cultural information;

[0392] S4, generates target audit results based on the adjusted content audit model.

[0393] Alternatively, a person skilled in the art may understand that: Fig. 9 The structure shown is for illustration only. Fig. 9 The electronic device and the electronic equipment described above are not limited in structure. Fig. 9 More or fewer components (such as network interfaces, etc.) as shown in, or with Fig. 9 Different configurations shown.

[0394] Among them, the memory 902 can be used to store software programs and modules, such as the program instructions / modules corresponding to the content review method and device in the embodiments of the present application. The processor 904 executes various functional applications and data processing by running the software programs and modules stored in the memory 902, that is, to implement the above-mentioned content review method. The memory 902 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 902 may further include a memory remotely located relative to the processor 904, and these remote memories may be connected to the terminal via a network. Examples of the above-mentioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Among them, the memory 902 can be specifically, but not limited to, used to store target interactive content, fused semantic vectors, and regional cultural information, etc. As an example, such as Fig. 9 As shown, the memory 902 may include, but is not limited to, the first acquisition unit 802, the first processing unit 804, the adjustment unit 806, and the second processing unit 808 in the content audit device. In addition, it may also include, but is not limited to, other module units in the content audit device, which will not be repeated in this example.

[0395] Optionally, the transmission device 906 is used to receive or send data via a network. Specific examples of the above-mentioned network may include a wired network and a wireless network. In one example, the transmission device 906 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices and routers via a network cable so as to communicate with the Internet or a local area network. In one example, the transmission device 906 is a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0396] In addition, the above-mentioned electronic device also includes: a display 908, which is used to display the screen of the above-mentioned content interaction platform and the target review results; and a connection bus 910, which is used to connect the various module components in the above-mentioned electronic device.

[0397] In other embodiments, the target terminal or server may be a node in a distributed system, wherein the distributed system may be a blockchain system, and the blockchain system may be a distributed system formed by connecting the multiple nodes through network communication. The nodes may form a point-to-point network, and any form of computing device, such as a server, a target terminal or other electronic device, may become a node in the blockchain system by joining the point-to-point network.

[0398] According to another aspect of the present application, a computer program product or computer program is provided, the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the content audit method provided in various optional implementations of the above-mentioned server verification processing and other aspects, wherein the computer program is configured to perform the steps of any of the above-mentioned method embodiments when running.

[0399] Optionally, in this embodiment, the computer-readable storage medium may be configured to store a computer program for performing the following steps:

[0400] S1, obtaining target interactive content in the content interactive platform;

[0401] S2, by inputting the target interaction content into a multimodal semantic recognition model, a fused semantic vector is obtained, wherein the fused semantic vector is used to indicate the emotional description information of the target interaction content;

[0402] S3, based on the fusion semantic vector and the regional cultural description vector, the evaluation parameters in the content audit model are adjusted to obtain the adjusted content audit model, where the evaluation parameters represent the degree of influence of some content with special meanings on the audit results, and the regional cultural description vector represents the deep meaning obtained by deeply mining the target interactive content in combination with regional cultural information;

[0403] S4, generates target audit results based on the adjusted content audit model.

[0404] Optionally, in the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories), or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0405] Optionally, in this embodiment, a person of ordinary skill in the art may understand that all or part of the steps in the various methods of the above embodiments may be completed by instructing the hardware related to the target terminal through a program, and the program may be stored in a computer-readable storage medium, and the storage medium may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.

[0406] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0407] If the integrated units in the above embodiments are implemented in the form of software functional units and sold or used as independent products, they can be stored in the above computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling one or more computer devices (which can be personal computers, servers or network devices, etc.) to execute all or part of the steps of the methods of each embodiment of the present application.

[0408] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0409] In the several embodiments provided in the present application, it should be understood that the disclosed client can be implemented in other ways. Among them, the device embodiments described above are only schematic, for example, the division of units is only a logical function division, and there may be other division methods in actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0410] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0411] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0412] The above are only preferred implementations of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A content review method, characterized in that: include: Obtain target interactive content in the content interactive platform; By inputting the target interaction content into a multimodal semantic recognition model, a fused semantic vector is obtained, wherein the fused semantic vector is used to indicate the emotional description information of the target interaction content; Based on the fused semantic vector and the regional cultural description vector, the evaluation parameters in the content audit model are adjusted to obtain an adjusted content audit model, wherein the evaluation parameters represent the degree of influence of some content with special meanings on the audit results, and the regional cultural description vector represents the deep meaning obtained by deeply mining the target interactive content in combination with regional cultural information; Based on the adjusted content review model, a target review result is generated.

2. The method according to claim 1, characterized in that The step of inputting the target interaction content into a multimodal semantic recognition model to obtain a fused semantic vector includes: Extracting features of video interaction data in the multimodal interaction data to obtain video interaction features, wherein the multimodal interaction data includes the video interaction data, audio interaction data, and text interaction data, and the target interaction content includes the multimodal interaction data; Extracting features from the audio interaction data to obtain audio interaction features; Performing feature extraction on the text interaction data to obtain text interaction features; The fused semantic vector is obtained by fusing the video interaction features, the audio interaction features and the text interaction features, and inputting the fused interaction features into the multimodal semantic recognition model.

3. The method according to claim 1, characterized in that The step of adjusting the evaluation parameters in the content audit model based on the fused semantic vector and the regional culture description vector includes: Acquire account behavior data corresponding to the account that published the target interactive content and regional description data of the location of the client logged in by the account; By inputting the account behavior data and the regional description data into a target perception model, the regional culture description vector for representing the current state of the account is obtained; Based on the emotion type of the target interactive content represented by the fused semantic vector and the deep meaning represented by the regional cultural description vector, the evaluation parameters in the content audit model are adjusted.

4. The method according to claim 3, characterized in that The step of adjusting the evaluation parameters in the content audit model based on the emotion type of the target interactive content represented by the fused semantic vector and the deep meaning represented by the regional cultural description vector includes: When the fused semantic vector indicates that the emotion type of the target interactive content is the first type, and the deep meaning indicates compliance with the regional cultural description specification, the value of the evaluation parameter is increased, wherein the higher the value of the evaluation parameter, the less likely it is that the part of the target interactive content with special meaning will be reviewed as non-compliant content; When the fused semantic vector indicates that the emotion type of the target interactive content is the second type, and the deep meaning does not conform to the regional cultural description standard, reducing the value of the evaluation parameter; When the fused semantic vector indicates that the emotion type of the target interactive content is the first type and the deep meaning does not conform to the regional cultural description standard, the value of the evaluation parameter is reduced.

5. The method according to claim 1, characterized in that The generating a target audit result based on the adjusted content audit model includes: Using a generative adversarial network to detect the target interactive content and obtain a target detection result, wherein the generative adversarial network is built into the content audit model; When the target detection result indicates that the target interaction content does not contain non-compliant content, performing differential privacy processing on the target interaction content by introducing random noise to obtain processed interaction content; The target audit result is obtained by inputting the processed interactive content into the adjusted content audit model.

6. The method according to claim 5, characterized in that The using of the generative adversarial network to detect the target interaction content to obtain a target detection result includes: Performing word segmentation and encoding processing on the text data in the target interactive content to obtain processed data features; The target detection result is obtained by inputting the processed data features into the discriminator in the generative adversarial network.

7. The method according to claim 6, characterized in that After obtaining the target detection result by inputting the processed data features into the discriminator in the generative adversarial network, the method further includes: When the target detection result indicates that the target interactive content has an abnormal phenomenon, adding an abnormal mark to the part of the target interactive content that has the abnormal phenomenon; The weight parameters in the discriminator are optimized to obtain an optimized generative adversarial network.

8. The method according to any one of claims 1 to 7, characterized in that After generating a target audit result based on the adjusted content audit model, the method further includes: In the case where the target audit result indicates that the target interactive content contains non-compliant content, processing the target interactive content includes at least one of the following: shielding the target interactive content, adjusting the target interactive content, and modifying the target interactive content.

9. A content review device, characterized in that: include: A first acquisition unit, used to acquire target interactive content in the content interaction platform; A first processing unit is used to obtain a fused semantic vector by inputting the target interaction content into a multimodal semantic recognition model, wherein the fused semantic vector is used to indicate emotional description information of the target interaction content; An adjustment unit, configured to adjust the evaluation parameters in the content audit model based on the fused semantic vector and the regional cultural description vector to obtain an adjusted content audit model, wherein the evaluation parameters represent the degree of influence of the partial content with special meaning on the audit result, and the regional cultural description vector represents the deep meaning obtained by deeply mining the target interactive content in combination with regional cultural information; The second processing unit is used to generate a target review result based on the adjusted content review model.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein the program can be executed by a terminal device or a computer to execute the method as claimed in any one of claims 1 to 8.

11. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to execute the method according to any one of claims 1 to 8 through the computer program.

Images