Media asset content processing method and device

Abstract

The invention provides a media asset content processing method and equipment, and relates to the technical field of data processing. According to the method, a first pre-training language model is adopted, abstract generation is carried out on text content of each piece of media asset data in a preset media asset database, and content abstract information corresponding to each piece of media asset data is obtained; according to the media asset data and the corresponding content summary information, a second pre-training language model is adopted, and based on input target theme information, target media asset text content corresponding to the target theme information is generated; performing content auditing on the target media asset text content based on a preset rule base by adopting a preset rule engine; and if the content verification is passed, carrying out content distribution on the target media asset text content by adopting a preset content distribution strategy. Therefore, the overall working efficiency and accuracy are improved, and the user experience is improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a media content processing method and device. Background Art

[0002] In the current information age, with the growing demand for timeliness and diversity in news and sports event reporting, we are faced with the challenge of how to efficiently utilize these massive materials and generate personalized and accurate content in real time.

[0003] The existing content production and distribution model often relies on a large number of human editors, which not only increases operating costs but also limits the flexibility and speed of content production. With the development of big data and artificial intelligence technologies, some automated tools and platforms designed to improve content production efficiency have begun to appear on the market, but these tools still have significant shortcomings when dealing with professional content in specific fields.

[0004] Existing content production mainly relies on manual work at all stages. Starting from the collection of raw materials, editors need to manually obtain content such as live sports event video recordings, purchased copyrighted TV series and children's variety shows, information news collections, UGC bloggers' self-uploaded content, large-scale evening parties or leaders' meetings from sources such as CCTV's media resource system, iQiyi, Tencent Video, CCTV channel live streams, self-owned operating resources, and outsourced content producers, and select suitable materials from them.

[0005] Initial introduction of automation: Although automation tools have been introduced in recent years, they are mostly focused on simple tasks. For example, natural language processing technology is used to automatically generate text content. By learning from a large amount of text data, it attempts to generate relevant articles. However, when processing professional content, it has limited semantic understanding and logic grasp, and low accuracy. Summary of the invention

[0006] The purpose of the present invention is to address the deficiencies in the above-mentioned prior art and provide a method and device for locating problems in a front-end task node component to solve the problems of low content processing accuracy in the prior art.

[0007] To achieve the above purpose, the technical solution adopted in the embodiment of the present application is as follows:

[0008] In a first aspect, an embodiment of the present application provides a method for processing media content, the method comprising:

[0009] Using a first pre-trained language model, a summary is generated for the text content of each media asset data in a preset media asset database to obtain content summary information corresponding to each media asset data;

[0010] According to the media asset data and the corresponding content summary information, a second pre-trained language model is used to generate target media asset text content corresponding to the target theme information based on the input target theme information;

[0011] Using a preset rule engine to perform content review on the target media asset text content based on a preset rule library;

[0012] If the content is approved, the preset content distribution strategy is used to distribute the target media asset text content.

[0013] Optionally, the adopting a preset content distribution strategy to distribute the target media asset text content includes:

[0014] Using a third pre-trained language model, classify the target media asset text content to obtain a target content classification corresponding to the target media asset text content;

[0015] Using the target content classification and the target media asset text content, and the preset content distribution strategy, determine a target media channel and a target user group in the target media channel from a plurality of preset media channels;

[0016] According to the information of the target user group and the target media asset text content, the application program interface of the target media channel is called to distribute the target media asset text content to the target user group.

[0017] Optionally, the adopting the target content classification and the target media asset text content, adopting the preset content distribution strategy, and determining a target media channel and a target user group in the target media channel from a plurality of preset media channels includes:

[0018] According to the target media asset text content and the media asset features corresponding to the multiple preset media channels, the first matching strategy in the preset content distribution strategy is adopted to determine the target media channel from the multiple preset media channels;

[0019] According to the target content classification and the user group characteristics of the target media channel, the target user group is determined from the user group of the target media channel by adopting the second matching strategy in the preset content distribution strategy.

[0020] Optionally, the adopting a preset rule engine to perform content review on the target media asset text content based on a preset rule library includes:

[0021] Using the preset rule engine, based on the preset standard keyword library in the preset rule library, the target media asset text content is subjected to keyword screening to obtain a keyword screening result;

[0022] Using the preset rule engine, based on the preset pattern library in the preset rule library, pattern matching is performed on the target media asset text content to obtain a pattern matching result;

[0023] If the keyword screening result indicates that there is no matching target keyword in the target media asset text content, and the pattern matching result indicates that the pattern matching of the target media asset text content is successful, it is determined that the content review is passed.

[0024] Optionally, the method further comprises:

[0025] If the keyword screening result indicates that the target keyword exists in the target media asset text content, then outputting the review instruction information of the target keyword and marking the target keyword to indicate manual keyword content review; and / or,

[0026] If the pattern matching result indicates that the pattern matching of the target media asset text content fails, then indication information of pattern matching failure is output to instruct manual pattern content review.

[0027] Optionally, before distributing the target media asset text content, the method further includes:

[0028] Using a preset sentiment analysis model to perform sentiment analysis on the target media text content to obtain the tendency sentiment information corresponding to the target media text content; and / or,

[0029] A preset malicious content detection model is used to perform content detection on the target media asset text content to obtain a malicious content detection result corresponding to the target media asset text content, wherein the malicious content detection result is used to indicate whether the target media asset text content has a risk of violating regulations.

[0030] Optionally, the method further comprises:

[0031] Collecting feedback data on the target media asset text content;

[0032] According to the feedback data, parameters of the first pre-trained language model, the second pre-trained language model, and the third pre-trained language model are updated.

[0033] Optionally, before the method adopts the third pre-trained language model to classify the target media asset text content to obtain a target content classification corresponding to the target media asset text content, the method further includes:

[0034] Acquire training data for multiple tasks from the preset media resource database and the preset media material library;

[0035] Collaborative training is performed between tasks according to the training data of the multiple tasks to obtain pre-trained language models of the multiple tasks, wherein the pre-trained language models of the multiple tasks include at least: the first pre-trained language model, the second pre-trained language model and the third pre-trained language model.

[0036] Optionally, the method further comprises:

[0037] Collecting user interaction data for the target media asset text content;

[0038] An evaluation parameter of the preset content distribution strategy is generated according to the user interaction data.

[0039] In a second aspect, an embodiment of the present application provides an electronic device, comprising: a processor and a storage medium, wherein the processor and the storage medium are communicatively connected via a bus, the storage medium stores program instructions executable by the processor, and the processor calls the program stored in the storage medium to execute the steps of the media content processing method as described in any one of the first aspects.

[0040] Compared with the prior art, this application has the following beneficial effects:

[0041] The present application provides a method and device for processing media asset content. The method uses a first pre-trained language model to generate a summary of the text content of each media asset data in a preset media asset database to obtain content summary information corresponding to each media asset data; based on each media asset data and the corresponding content summary information, a second pre-trained language model is used to generate target media asset text content corresponding to the target subject information based on the input target subject information; a preset rule engine is used to perform content review on the target media asset text content based on a preset rule library; if the content review is passed, a preset content distribution strategy is used to distribute the target media asset text content. Thus, both internal and external media assets are efficiently processed through the pre-trained language model, and after review, the generated target media asset text content is more standardized and regulated, which greatly improves collaboration efficiency, reduces manual intervention, improves overall work efficiency and accuracy, and improves user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0043] Figure 1 A flowchart of a media content processing method provided in this application;

[0044] Figure 2 A flowchart of a method for distributing target media asset text content provided in an embodiment of the present application;

[0045] Figure 3 A flowchart of a method for determining a target media channel and a target user group provided in an embodiment of the present application;

[0046] Figure 4 A flowchart of a method for reviewing text content of a media asset provided in an embodiment of the present application;

[0047] Figure 5 A model updating method provided in an embodiment of the present application;

[0048] Figure 6 A model training method provided in an embodiment of the present application;

[0049] Figure 7 A method for evaluating a preset content distribution strategy provided in an embodiment of the present application;

[0050] Figure 8 A schematic diagram of a media content processing device provided in an embodiment of the present application;

[0051] Fig. 9 A schematic diagram of an electronic device provided in an embodiment of the present application.

[0052] Icons: 801 - first generation module, 802 - second generation module, 803 - review module, 804 - distribution module, 901 - processor, 902 - storage medium. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, 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 part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0054] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for protection, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.

[0055] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0056] In addition, the terms “first”, “second”, etc., if used, are merely used to distinguish between the descriptions and should not be understood as indicating or implying relative importance.

[0057] It should be noted that, in the absence of conflict, the features in the embodiments of the present invention may be combined with each other.

[0058] A media content processing method provided by the present application is explained below through a specific example. Figure 1 The present application provides a flow chart of a method for processing media content, wherein the execution subject of the method is an electronic device having a computing and processing function. Figure 1 As shown, the method includes:

[0059] S101: Use a first pre-trained language model to generate a summary of the text content of each media asset data in a preset media asset database to obtain content summary information corresponding to each media asset data.

[0060] The preset media resource database will collect various media resource data to facilitate content processing of these media resource data. It should be noted that after the preset media resource database collects new media resource data, this step will generate a summary and continue to store each media resource data and the corresponding content summary information in the preset media resource database for subsequent use.

[0061] For example, if the media data is a text file, the text content of the text file is directly obtained; if the media data is a picture / video file, the content of the picture / video file is identified as text content, or the title, introduction and other information of the picture / video file are used as text content.

[0062] The text content of the media data is input into the first pre-trained language model, which identifies the important content in the text content, removes redundant information, and generates concise and accurate content summary information. For example, for a long meeting minutes, the first pre-trained language model can extract the main topics of the meeting minutes, decision results and other key points, and generate a concise summary.

[0063] For example, before collecting various types of media data in the preset media database, data cleaning will be performed on various types of media data. The collected media data will be denoised to remove duplicate, erroneous or incomplete data to ensure the accuracy and consistency of the data. For example, garbled characters and redundant punctuation marks in the text will be removed. At the same time, text normalization operations will be performed, such as unifying uppercase and lowercase letters, processing abbreviations, etc., to unify the text format for subsequent processing.

[0064] S102: According to each media asset data and the corresponding content summary information, a second pre-trained language model is used to generate target media asset text content corresponding to the target subject information based on the input target subject information.

[0065] When generating media asset text content, the user inputs target subject information, and the target subject information can be compared with multiple generated content summary information to determine the content summary information with the highest similarity to the target subject information as the target content summary information. Based on the media asset data corresponding to the target content summary information and the preset writing style, the target subject information is expanded to generate target media asset text content corresponding to the target subject information.

[0066] For example, the target subject information may be: subject keywords, user preference data, etc. For example, when the input target subject information is "Olympic Opening Ceremony", the content summary information closest to "Olympic Opening Ceremony" is first determined as the target content summary information, and based on the media data corresponding to the target content summary information and the preset writing style (for example, CCTV reporting style), "Olympic Opening Ceremony" is expanded. The second pre-trained language model may generate "The Opening Ceremony of the Olympic Games is a grand event that has attracted the attention of athletes and spectators from all over the world. At the opening ceremony, delegations of athletes from various countries entered the venue one by one. They were dressed in bright uniforms and were full of passion and vitality. The lights, music and performances on the scene complemented each other, creating a warm and solemn atmosphere." and other content.

[0067] It should be noted that the preset writing style has been embedded when training the second pre-trained language model, and the media text content with the preset writing style can be used to train the second pre-trained language model, so that the writing style of the generated text content is the same as the preset writing style. Thus, it is convenient to generate personalized content for users and improve the user experience.

[0068] S103: Using a preset rule engine, perform content review on the target media asset text content based on a preset rule library.

[0069] By conducting content review on the target media text content, the target media text content is made more standardized and accurate, avoiding the publication of illegal and irregular content.

[0070] S104: If the content is approved, a preset content distribution strategy is used to distribute the target media asset text content.

[0071] The target media asset text content is distributed through a preset content distribution strategy, thereby achieving accurate distribution of the target media asset text content and improving the user experience.

[0072] As a result, both internal and external media resources are efficiently processed through pre-trained language models, and after review, the generated target media text content is more standardized and regulated, which greatly improves collaboration efficiency, reduces manual intervention, improves overall work efficiency and accuracy, and enhances user experience.

[0073] In summary, in this embodiment, a first pre-trained language model is used to generate a summary of the text content of each media asset data in the preset media asset database, and obtain the content summary information corresponding to each media asset data; according to each media asset data and the corresponding content summary information, a second pre-trained language model is used to generate the target media asset text content corresponding to the target subject information based on the input target subject information; a preset rule engine is used to perform content review on the target media asset text content based on the preset rule library; if the content review is passed, the preset content distribution strategy is used to distribute the target media asset text content. Thus, both internal and external media assets are efficiently processed through the pre-trained language model, and after review, the generated target media asset text content is more standardized and normative, which greatly improves the collaboration efficiency, reduces manual intervention, improves the overall work efficiency and accuracy, and improves the user experience.

[0074] In the above Figure 1 On the basis of the corresponding embodiments, the embodiments of the present application also provide a method for distributing target media asset text content. Figure 2 A flow chart of a method for distributing target media asset text content provided in an embodiment of the present application. Figure 2 As shown, in S104, the preset content distribution strategy is adopted to distribute the target media asset text content, including:

[0075] S201. Use a third pre-trained language model to classify target media asset text content to obtain a target content classification corresponding to the target media asset text content.

[0076] For example, news is classified into political news, social news, sports event news, etc. based on the content. The model can accurately classify the target media text content by analyzing the keywords, grammatical structure and other features of the target media text content.

[0077] S202: Using the target content classification and the target media asset text content, and the preset content distribution strategy, determine the target media channel and the target user group in the target media channel from multiple preset media channels.

[0078] By determining the target media channels and the target user groups in the target media channels, the target media asset text content can be accurately distributed, thereby improving the user experience.

[0079] S203: Calling the application program interface of the target media channel according to the information of the target user group and the target media asset text content, so as to distribute the target media asset text content to the target user group.

[0080] For example, during actual distribution, the target media asset text content is distributed to devices of the target user group (eg, televisions, mobile phones, computers).

[0081] In summary, in this embodiment, the third pre-trained language model is used to classify the target media text content to obtain the target content classification corresponding to the target media text content; the target content classification and the target media text content are used, and the preset content distribution strategy is used to determine the target media channel and the target user group in the target media channel from multiple preset media channels; according to the information of the target user group and the target media text content, the application program interface of the target media channel is called to distribute the target media text content to the target user group. Thus, the target media text content is accurately distributed and the user experience is improved.

[0082] In the above Figure 2 On the basis of the corresponding embodiments, the embodiments of the present application also provide a method for determining a target media channel and a target user group. Figure 3 A flowchart of a method for determining a target media channel and a target user group provided in an embodiment of the present application. Figure 3 As shown, in S202, the target content classification and the target media asset text content are adopted, and the preset content distribution strategy is adopted to determine the target media channel and the target user group in the target media channel from multiple preset media channels, including:

[0083] S301 : According to the target media asset text content and the media asset features corresponding to the plurality of preset media channels, a first matching strategy in the preset content distribution strategy is adopted to determine the target media channel from the plurality of preset media channels.

[0084] For example, the first matching strategy is to distribute content according to the characteristics of media assets of different media channels. For example, target media assets in the form of short articles and pictures are suitable for push through social media, and target media assets in the form of in-depth reports are suitable for push through news websites, etc.).

[0085] S302: According to the target content classification and the user group characteristics of the target media channel, a second matching strategy in the preset content distribution strategy is adopted to determine the target user group from the user group of the target media channel.

[0086] For example, the second matching strategy is to distribute content according to user group characteristics (such as age, region, interest preference, etc.) of different user groups. For example, the user group characteristics with the highest similarity to the target content classification are determined as the target group characteristics, and the user group corresponding to the target group characteristics are determined as the target user group.

[0087] For example, for social media channels with a large number of young user groups, priority should be given to pushing content related to fashion, entertainment, etc., and more attractive titles and picture formats should be used.

[0088] In summary, in this embodiment, according to the target media asset text content and the media asset characteristics corresponding to multiple preset media channels, the first matching strategy in the preset content distribution strategy is adopted to determine the target media channel from multiple preset media channels; according to the target content classification and the user group characteristics of the target media channel, the second matching strategy in the preset content distribution strategy is adopted to determine the target user group from the user group of the target media channel. Thus, accurate distribution is achieved and user experience is improved.

[0089] In the above Figure 1 On the basis of the corresponding embodiments, the embodiments of the present application also provide a method for reviewing the text content of media assets. Figure 4 The following is a flow chart of a method for reviewing the content of a media asset text provided in an embodiment of the present application. Figure 4 As shown, in S103, a preset rule engine is used to perform content review on the target media asset text content based on a preset rule library, including:

[0090] S401: Using a preset rule engine, based on a preset standard keyword library in a preset rule library, perform keyword screening on the target media asset text content to obtain a keyword screening result.

[0091] The preset rule engine is based on a custom preset rule library to perform keyword screening on generated or imported content.

[0092] For example, the preset rule library includes but is not limited to legal and regulatory keywords (such as words related to copyright, privacy, advertising law, etc.) and content standard words (such as vulgar and violent words that are prohibited from use).

[0093] Perform keyword screening on the target media asset text content. For example, if there are copyright-related words such as "All Rights Reserved" in the preset rule library, when the word appears in the target media asset text content, the keyword screening result will show the presence of the keyword.

[0094] S402: Using a preset rule engine, based on a preset pattern library in a preset rule library, pattern matching is performed on the target media asset text content to obtain a pattern matching result.

[0095] On the basis of keyword screening, pattern matching is added to improve the accuracy of review.

[0096] For example, in terms of pattern matching, if there are specific pattern requirements for the title format of a news article (such as "[Event Body] + [Event Overview]"), the title will be checked for pattern matching to determine whether the title meets this specific pattern requirement. This makes the target media asset text content more standardized and more accurate.

[0097] S403: If the keyword screening result indicates that there is no matching target keyword in the target media asset text content, and the pattern matching result indicates that the pattern matching of the target media asset text content is successful, it is determined that the content review has passed.

[0098] If the keyword screening result indicates that there is no matching target keyword in the target media asset text content, it means that the target media asset text content does not violate the rules. If the pattern matching result indicates that the pattern matching of the target media asset text content is successful, it means that the target media asset text content complies with the pattern regulations. Then it is determined that the content review has passed.

[0099] In summary, in this embodiment, a preset rule engine is used to perform keyword screening on the target media asset text content based on the preset standard keyword library in the preset rule library to obtain a keyword screening result; a preset rule engine is used to perform pattern matching on the target media asset text content based on the preset pattern library in the preset rule library to obtain a pattern matching result; if the keyword screening result indicates that there is no matching target keyword in the target media asset text content, and the pattern matching result indicates that the pattern matching of the target media asset text content is successful, it is determined that the content review has passed. Thus, the review accuracy is improved, making the target media asset text content more standardized and more accurate.

[0100] In the above Figure 4 On the basis of the corresponding embodiment, in another embodiment of the present application, the method further includes:

[0101] If the keyword screening result indicates that there is a matching target keyword in the target media asset text content, then the review instruction information of the target keyword is output, and the target keyword is marked to instruct manual keyword content review.

[0102] If the keyword screening results indicate that there are matching target keywords in the target media asset text content, it means that the target media asset text content is not in violation of the rules. It is necessary to instruct manual keyword content review. Manual can modify the target keywords in the target media asset text content until the review is passed.

[0103] And / or, if the pattern matching result indicates that the pattern matching of the target media asset text content fails, then indication information of pattern matching failure is output to instruct manual pattern content review.

[0104] If the pattern matching result indicates that the pattern matching of the target media asset text content fails, it means that the target media asset text content does not meet the pattern requirements. Then manual pattern content review is required to modify the model in the target media asset text content until the pattern matching succeeds.

[0105] In summary, in this embodiment, if the keyword screening result indicates that there is a matching target keyword in the target media asset text content, then the target keyword review indication information is output, and the target keyword is marked to indicate manual keyword content review; and / or, if the pattern matching result indicates that the pattern matching of the target media asset text content fails, then the pattern matching failure indication information is output to indicate manual pattern content review. Thus, the review accuracy is improved, making the target media asset text content more standardized and more accurate.

[0106] In the above Figure 4 On the basis of the corresponding embodiment, in another embodiment of the present application, before distributing the target media asset text content in S104, the method further includes:

[0107] The preset sentiment analysis model is used to perform sentiment analysis on the target media text content to obtain the tendency sentiment information corresponding to the target media text content.

[0108] By using a preset sentiment analysis model, the target media text content is analyzed for features such as vocabulary and sentence structure to determine the sentiment tendency information (e.g., positive, negative, or neutral). By obtaining the sentiment tendency information of the target media text content, it is convenient to review and classify the target media text content according to the sentiment tendency information.

[0109] For example, for a text like "This game was so exciting and the athletes' performance was amazing!", the preset sentiment analysis model would judge it as a positive sentiment.

[0110] And / or, a preset malicious content detection model is used to perform content detection on the target media text content to obtain a malicious content detection result corresponding to the target media text content, and the malicious content detection result is used to indicate whether the target media text content has a violation risk.

[0111] The preset malicious content detection model detects whether the target media asset text content has violation risks by identifying malicious intent vocabulary, grammatical structure and other features in the text.

[0112] In summary, in this embodiment, a preset sentiment analysis model is used to perform sentiment analysis on the target media text content to obtain the tendency sentiment information corresponding to the target media text content; and / or a preset malicious content detection model is used to perform content detection on the target media text content to obtain the malicious content detection result corresponding to the target media text content, and the malicious content detection result is used to indicate whether the target media text content has a risk of violation. Thus, the audit accuracy is improved, making the target media text content more standardized and more accurate.

[0113] Furthermore, in another embodiment of the present application, a preset black and white list engine is used to perform black and white list screening on the target media asset text content based on a preset black and white name word library in a preset black and white list library to obtain a black and white list screening result.

[0114] For example, the blacklist contains information such as words, IP addresses, and user accounts that are explicitly prohibited from being published, and the whitelist contains information about sources that are allowed to be published first or trusted. When the target media asset text content involves words in the blacklist, it is directly prohibited from being published. If the target media asset text content comes from a reliable source in the whitelist and there are no obvious problems with the content, it can be quickly reviewed.

[0115] This application can ensure that the content is legal and compliant while retaining necessary information to make the content more comprehensive and objective. The review process specifically for political content combines manual review with machine learning technology to achieve accurate identification and efficient processing of political information, ensuring that all published news and information meet the highest review standards.

[0116] In the above Figure 2 Based on the corresponding embodiments, the embodiments of the present application also provide a model updating method. Figure 5 A model updating method is provided in an embodiment of the present application. Figure 5 As shown, the method also includes:

[0117] S501: Collect feedback data on target media asset text content.

[0118] Acquire feedback data on the target media asset text content in real time. The feedback data includes direct feedback from users on the target media asset text content (such as ratings, keywords in comments, etc.) and indirect behavior data (such as click-through rate, reading time, likes, comments and sharing, etc.).

[0119] For example, if an article about the opening ceremony of the Olympics has a high click-through rate, a long reading time, and positive user comments, the model will use this data as a basis for optimization.

[0120] S502: Update parameters of the first pre-trained language model, the second pre-trained language model, and the third pre-trained language model according to the feedback data.

[0121] Model parameters are optimized through online learning or regular offline training to continuously improve content quality and user experience.

[0122] Furthermore, a reinforcement learning algorithm is introduced to guide the model to generate content that is consistent with the characteristics of the domain and meets user needs by designing a reasonable reward function.

[0123] For example, the reward function is designed based on multiple factors such as content accuracy (whether it conforms to facts, expertise, etc.), relevance (the degree of relevance to the topic and user needs), and novelty (whether it provides new perspectives or information). For example, if the generated content accurately describes the key aspects of the Olympic Games opening ceremony and is attractive, the reward function will give a higher reward value, otherwise it will give a lower reward value. The model continuously iterates and optimizes model parameters based on reward feedback to improve content quality.

[0124] In summary, in this embodiment, feedback data on the target media asset text content is collected; and parameters of the first pre-trained language model, the second pre-trained language model, and the third pre-trained language model are updated according to the feedback data. Thus, the model parameters are iteratively optimized to continuously improve the content quality and user experience.

[0125] In the above Figure 2 Based on the corresponding embodiments, the embodiments of the present application also provide a model training method. Figure 6 A model training method is provided in the embodiment of the present application. Figure 6 As shown, in S201, before the third pre-trained language model is used to classify the target media asset text content and obtain the target content classification corresponding to the target media asset text content, the method further includes:

[0126] S601: Acquire training data for multiple tasks from a preset media resource database and a preset media material library.

[0127] In order to improve the training accuracy, the training data is preprocessed as follows:

[0128] Data cleaning: Use text cleaning tools to pre-process the acquired training data, including but not limited to removing noise data (such as garbled characters, special symbols, etc.) and performing text cleaning (unified format, case conversion, etc.) to make the text more standardized.

[0129] Lexical analysis: Using the common part-of-speech distribution and grammatical rules of words in a large-scale corpus, the text is broken down into basic language units such as words and punctuation marks, and part-of-speech annotations are performed (such as nouns, verbs, etc.). For example, for the sentence "Company A is a content operation company", "Company A" will be identified as a noun phrase, and "is" will be identified as a verb, etc.

[0130] Syntactic analysis: Use syntactic analysis algorithms to construct a sentence's syntactic structure tree and determine the grammatical relationship between words. For example, the syntactic structure of the above sentence may be "[Company A (subject)][is (predicate)][a content operation company (object)]".

[0131] Named Entity Recognition (NER): Use named entity recognition technology based on the conditional random field (CRF) model to identify specific entities in the text, such as names of people, organizations, places, etc. The output structured content is the text after the recognized entities are classified and labeled, for example, "[Company A (organization name)][is (verb)][a content operation company (organization name)]". The final output structured content is the text with part-of-speech tagging, syntactic structure, and named entity classification and labeling information after the above processing, so that subsequent modules can use important elements in the text more accurately for operations.

[0132] Knowledge graph construction: The captured key information is used as nodes, such as "Chinese athletes", "Olympic Games", "gold medals", "China", etc. as nodes respectively. Connections are established based on the semantic relationships between nodes. For example, "Chinese athletes" and "China" are connected through the "country of origin" relationship, "Chinese athletes" and "Olympic Games" are connected through the "participation" relationship, "Chinese athletes" and "gold medals" are connected through the "obtained" relationship, and "Olympic Games" and "gold medals" are connected through the "awarding" relationship. In this way, a semantic network is constructed to present the important elements in the text in a structured form, and subsequent operations such as content generation, recommendation, and classification can be performed more accurately based on this knowledge graph. For example, when generating relevant reports about the athlete, information such as his country of origin and participating events can be obtained from the knowledge graph, thereby generating more comprehensive, accurate, and personalized content. At the same time, when reviewing the content, the rationality and accuracy of the information can also be judged based on the relationships in the knowledge graph. For example, the event extraction model can more accurately capture key information and construct a knowledge graph based on such structured content.

[0133] Specifically, the event extraction model is applied to capture key information, such as event subject, event time, event location, event type, etc. For example:

[0134] Event subject: For example, in news information, the main people, organizations or institutions involved will be identified. For example, in a report "Chinese athletes won gold medals at the Olympics", "Chinese athletes" are the event subjects.

[0135] Event time: Extract the specific time or time period when the event occurred, such as "during the Olympic Games" in the above news (if there is a specific time, it will be extracted more accurately).

[0136] Event location: Specify the location where the event occurred, such as "Olympic Games Stadium" (if there is a detailed location, it will be further extracted).

[0137] Event Type: Determines the nature of the event, such as winning a sports event belongs to the "sports event result" type.

[0138] Related relations: Identify the relations between event subjects and between events and other elements. For example, the relation between athletes and their countries (“Chinese athletes” belong to “China”), the relation between athletes and events (“Chinese athletes” participated in “an Olympic event”), etc.

[0139] By preprocessing the training data, the training data contains structured content and knowledge graphs, which facilitates subsequent model training and improves the accuracy of model training.

[0140] S602: Perform collaborative training between tasks based on the training data of multiple tasks to obtain pre-trained language models of multiple tasks.

[0141] The pre-trained language models for multiple tasks include at least: a first pre-trained language model, a second pre-trained language model, and a third pre-trained language model.

[0142] For example, you can choose advanced pre-trained language models such as BERT or GPT series. These models have been pre-trained on large-scale corpora and have learned general language knowledge and semantic representations, which can provide a good foundation for subsequent specific field training. For example, BERT performs well in natural language understanding tasks and can better capture semantic information in texts; the GPT series has advantages in generation tasks and can generate coherent and logically reasonable texts.

[0143] The training data of multiple tasks are input into the pre-trained language model, and the model parameters are continuously adjusted until the training results output by the pre-trained language model meet the preset training stop conditions, so that it can adapt to the language characteristics and task requirements in this field.

[0144] For example, the first pre-trained language model, the second pre-trained language model, and the third pre-trained language model may be three different models obtained through training, or may be a model including three tasks. Through training on these different tasks, the model can learn the features and semantic relationships of the text from multiple perspectives, thereby improving its comprehensive understanding and generation capabilities.

[0145] The first pre-trained language model is used to generate content summary information: the training model can automatically generate a summary of the text and extract key information. The model learns how to identify important content in the text, remove redundant information, and generate a concise and accurate summary.

[0146] The second pre-trained language model is used for the question-answering system: constructing question-answer pairs so that the model can learn how to extract and generate accurate answers from the text based on the questions, and can also generate media text content corresponding to the subject information based on the input subject information. For example, for event reports in CCTV news, relevant questions are raised (such as "When did the event happen?", "Who are the main participants in the event?", etc.), allowing the model to find answers in the text and train its ability to accurately understand and extract text information.

[0147] The third pre-trained language model is used for text classification: the model learns to distinguish different types of text, such as classifying news into political news, social news, sports event news, etc. The model analyzes a large number of texts with labeled types and learns the keywords, grammatical structures and other features of different types of news, so that it can accurately classify new news texts.

[0148] Through collaborative learning between tasks, there is a mutually reinforcing relationship between different downstream tasks. The model can comprehensively improve its understanding and generation capabilities in specific fields, enabling it to better handle various content-related tasks and generate higher-quality content that meets the requirements.

[0149] In summary, in this embodiment, training data of multiple tasks are obtained from a preset media resource database and a preset media material library; collaborative training between tasks is performed based on the training data of multiple tasks to obtain pre-trained language models of multiple tasks, and the pre-trained language models of multiple tasks at least include: a first pre-trained language model, a second pre-trained language model, and a third pre-trained language model. Thus, the model can better handle various content-related tasks and generate high-quality content that meets the requirements.

[0150] In the above Figure 1 On the basis of the corresponding embodiments, the embodiments of the present application also provide a method for evaluating a preset content distribution strategy. Figure 7 A method for evaluating a preset content distribution strategy is provided in an embodiment of the present application. Figure 7 As shown, the method also includes:

[0151] S701: Collect user interaction data for target media asset text content.

[0152] For example, a comprehensive data tracking and analysis system is established to collect user interaction data on the target media text content, such as click-through rate, reading time, likes, comments, sharing and other behavioral indicators.

[0153] S702: Generate evaluation parameters of a preset content distribution strategy according to user interaction data.

[0154] Based on user interaction data, the content strategy can be comprehensively evaluated. The evaluation parameters can be: excellent, good, or poor. This makes it easier for users to preset the evaluation parameters of the content distribution strategy and optimize the preset content distribution strategy.

[0155] For example, A / B testing and effect evaluation can be performed on a data analysis platform based on user interaction data. For example, two articles with the same theme but slightly different titles and content structures (version A and version B) are published at the same time. By comparing the click-through rate, reading time and other data of the two in different time periods and different user groups, it is evaluated which version is more popular with users. Effect evaluation comprehensively evaluates content strategies based on set goals (such as increasing user engagement, increasing content dissemination, etc.).

[0156] In summary, in this embodiment, user interaction data for target media asset text content is collected, and evaluation parameters of a preset content distribution strategy are generated based on the user interaction data, so as to facilitate the user to preset evaluation parameters of the content distribution strategy and optimize and adjust the preset content distribution strategy.

[0157] The following is a description of a media content processing device, equipment, storage medium, etc. provided by the present application for execution. The specific implementation process and technical effects are as mentioned above and will not be repeated below.

[0158] Figure 8 A schematic diagram of a media content processing device provided in an embodiment of the present application is shown in FIG. Figure 8 As shown, the device comprises:

[0159] The first generating module 801 is used to generate a summary of the text content of each media asset data in the preset media asset database by using the first pre-trained language model to obtain content summary information corresponding to each media asset data.

[0160] The second generating module 802 is used to generate target media asset text content corresponding to the target subject information based on the input target subject information by using a second pre-trained language model according to each media asset data and the corresponding content summary information.

[0161] The review module 803 is used to use a preset rule engine to perform content review on the target media asset text content based on a preset rule library.

[0162] The distribution module 804 is used to distribute the target media asset text content using a preset content distribution strategy if the content is approved.

[0163] Furthermore, the distribution module 804 is specifically used to use the third pre-trained language model to perform content classification on the target media text content to obtain a target content classification corresponding to the target media text content; use the target content classification and the target media text content, and use a preset content distribution strategy to determine the target media channel and the target user group in the target media channel from multiple preset media channels; and call the application program interface of the target media channel according to the information of the target user group and the target media text content to distribute the target media text content to the target user group.

[0164] Furthermore, the distribution module 804 is specifically used to determine the target media channel from multiple preset media channels by adopting a first matching strategy in the preset content distribution strategy based on the target media asset text content and the media asset characteristics corresponding to multiple preset media channels; and to determine the target user group from the user group of the target media channel by adopting a second matching strategy in the preset content distribution strategy based on the target content classification and the user group characteristics of the target media channel.

[0165] Furthermore, the review module 803 is specifically used to adopt a preset rule engine to perform keyword screening on the target media asset text content based on a preset standard keyword library in the preset rule library to obtain a keyword screening result; adopt a preset rule engine to perform pattern matching on the target media asset text content based on a preset pattern library in the preset rule library to obtain a pattern matching result; if the keyword screening result indicates that there is no matching target keyword in the target media asset text content, and the pattern matching result indicates that the pattern matching of the target media asset text content is successful, it is determined that the content review has passed.

[0166] Furthermore, the review module 803 is specifically used to output review indication information of the target keyword and mark the target keyword to instruct manual keyword content review if the keyword screening result indicates that there is a matching target keyword in the target media asset text content; and / or, if the pattern matching result indicates that the pattern matching of the target media asset text content fails, output indication information of pattern matching failure to instruct manual pattern content review.

[0167] Furthermore, the distribution module 804 is also used to use a preset sentiment analysis model to perform sentiment analysis on the target media text content to obtain the tendency sentiment information corresponding to the target media text content; and / or use a preset malicious content detection model to perform content detection on the target media text content to obtain the malicious content detection result corresponding to the target media text content, and the malicious content detection result is used to indicate whether the target media text content has a violation risk.

[0168] Furthermore, the distribution module 804 is also used to update parameters of the first pre-trained language model, the second pre-trained language model and the third pre-trained language model according to the feedback data.

[0169] Furthermore, the distribution module 804 is also used to obtain training data for multiple tasks from a preset media resource database and a preset media material library; perform collaborative training between tasks based on the training data of multiple tasks to obtain pre-trained language models for multiple tasks, and the pre-trained language models for multiple tasks include at least: a first pre-trained language model, a second pre-trained language model and a third pre-trained language model.

[0170] Furthermore, the distribution module 804 is also used to collect user interaction data for the target media asset text content; and generate evaluation parameters of a preset content distribution strategy based on the user interaction data.

[0171] Fig. 9 A schematic diagram of an electronic device provided in an embodiment of the present application, which electronic device may be a device with computing and processing functions.

[0172] The electronic device includes: a processor 901 and a storage medium 902. The processor 901 and the storage medium 902 are connected via a bus.

[0173] The storage medium 902 is used to store programs, and the processor 901 calls the programs stored in the storage medium 902 to execute the above method embodiment. The specific implementation method and technical effect are similar and will not be repeated here.

[0174] Optionally, the present invention also provides a storage medium, including a program, which is used to execute the above-mentioned method embodiment when executed by a processor. In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as 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 devices or units, which can be electrical, mechanical or other forms.

[0175] 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.

[0176] In addition, each functional unit in each embodiment of the present invention 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 hardware plus software functional units.

[0177] The above-mentioned integrated unit implemented in the form of a software functional unit can be stored in a storage medium. The above-mentioned software functional unit is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor (English: processor) to perform some steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (English: Read-Only Memory, abbreviated: ROM), random access memory (English: Random Access Memory, abbreviated: RAM), disk or optical disk and other media that can store program codes.

Claims

1. A method for processing media content, characterized in that: The method comprises: Using a first pre-trained language model, a summary is generated for the text content of each media asset data in a preset media asset database to obtain content summary information corresponding to each media asset data; According to the media asset data and the corresponding content summary information, a second pre-trained language model is used to generate target media asset text content corresponding to the target theme information based on the input target theme information; Using a preset rule engine to perform content review on the target media asset text content based on a preset rule library; If the content is approved, the preset content distribution strategy is used to distribute the target media asset text content.

2. The method according to claim 1, characterized in that The adopting a preset content distribution strategy to distribute the target media asset text content includes: Using a third pre-trained language model, classify the target media asset text content to obtain a target content classification corresponding to the target media asset text content; Using the target content classification and the target media asset text content, and the preset content distribution strategy, determine a target media channel and a target user group in the target media channel from a plurality of preset media channels; According to the information of the target user group and the target media asset text content, the application program interface of the target media channel is called to distribute the target media asset text content to the target user group.

3. The method according to claim 2, characterized in that The method of adopting the target content classification and the target media asset text content, and the preset content distribution strategy, to determine a target media channel and a target user group in the target media channel from a plurality of preset media channels includes: According to the target media asset text content and the media asset features corresponding to the multiple preset media channels, the first matching strategy in the preset content distribution strategy is adopted to determine the target media channel from the multiple preset media channels; According to the target content classification and the user group characteristics of the target media channel, the target user group is determined from the user group of the target media channel by adopting the second matching strategy in the preset content distribution strategy.

4. The method according to claim 1, characterized in that The method of using a preset rule engine to perform content review on the target media asset text content based on a preset rule library includes: Using the preset rule engine, based on the preset standard keyword library in the preset rule library, the target media asset text content is subjected to keyword screening to obtain a keyword screening result; Using the preset rule engine, based on the preset pattern library in the preset rule library, pattern matching is performed on the target media asset text content to obtain a pattern matching result; If the keyword screening result indicates that there is no matching target keyword in the target media asset text content, and the pattern matching result indicates that the pattern matching of the target media asset text content is successful, it is determined that the content review is passed.

5. The method according to claim 4, characterized in that The method further comprises: If the keyword screening result indicates that the target keyword exists in the target media asset text content, then outputting the review instruction information of the target keyword and marking the target keyword to indicate manual keyword content review; and / or, If the pattern matching result indicates that the pattern matching of the target media asset text content fails, then indication information of pattern matching failure is output to instruct manual pattern content review.

6. The method according to claim 1, characterized in that Before distributing the target media asset text content, the method further includes: Using a preset sentiment analysis model to perform sentiment analysis on the target media text content to obtain the tendency sentiment information corresponding to the target media text content; and / or, A preset malicious content detection model is used to perform content detection on the target media asset text content to obtain a malicious content detection result corresponding to the target media asset text content, wherein the malicious content detection result is used to indicate whether the target media asset text content has a risk of violating regulations.

7. The method according to claim 2, characterized in that The method further comprises: Collecting feedback data on the target media asset text content; According to the feedback data, parameters of the first pre-trained language model, the second pre-trained language model, and the third pre-trained language model are updated.

8. The method according to claim 2, characterized in that: Before the third pre-trained language model is used to classify the target media asset text content to obtain the target content classification corresponding to the target media asset text content, the method further includes: Acquire training data for multiple tasks from the preset media resource database and the preset media material library; Collaborative training is performed between tasks according to the training data of the multiple tasks to obtain pre-trained language models of the multiple tasks, wherein the pre-trained language models of the multiple tasks include at least: the first pre-trained language model, the second pre-trained language model and the third pre-trained language model.

9. The method according to claim 1, characterized in that: The method further comprises: Collecting user interaction data for the target media asset text content; An evaluation parameter of the preset content distribution strategy is generated according to the user interaction data.

10. An electronic device, characterized in that: include: A processor and a storage medium, wherein the processor and the storage medium are communicatively connected via a bus, the storage medium stores program instructions executable by the processor, and the processor calls the program stored in the storage medium to execute the steps of the media asset content processing method as described in any one of claims 1 to 9.

Images