Content analysis method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention discloses a content analysis method and device, electronic equipment and a storage medium, and the method comprises the steps: calling a first intelligent agent to obtain object information and historical operation information of a target object, and determining an object tag when the target object browses content according to the object information and the historical operation information; calling a second agent to obtain a plurality of content segments corresponding to the to-be-analyzed content and a content text corresponding to the to-be-analyzed content, and determining an abnormal risk level and a content risk description of the to-be-analyzed content according to the plurality of content segments and the content text; calling a third agent to determine an analysis result of the to-be-analyzed content according to the object label, the abnormal risk level and the content risk description, wherein the analysis result is used for indicating an abnormal type and an abnormal reason of the to-be-analyzed content; according to the embodiment of the invention, the exception type and the exception reason of the to-be-analyzed content can be accurately determined without manual intervention.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of content recommendation, and in particular to a content analysis method and device, an electronic device, and a storage medium. BACKGROUND

[0002] In a content recommendation scenario, the content recommended to a user is actually not liked by the user, which may generate various negative feedbacks, such as quickly swiping away, clicking a "dislike" control, and clicking a "report and complaint" control. After the user generates negative feedback, a business staff needs to view the content that the user has historically browsed and the content that triggers the user's negative feedback to determine the specific content that triggers the user's negative feedback (for example, the content conflicts with the user's values), so as to reduce the recommendation of related content for the user. However, this analysis method is inefficient, and in the case of a large number of negative feedbacks, it will bring a huge workload to the business staff. SUMMARY

[0003] The following is a summary of the subject matter of the detailed description of the present disclosure. This summary is not intended to limit the protection scope of the claims.

[0004] The embodiments of the present disclosure provide a content analysis method and device, an electronic device, and a storage medium, which can accurately determine the abnormal type and abnormal reason of the content to be analyzed without human intervention.

[0005] In one aspect, the embodiments of the present disclosure provide a content analysis method, comprising: calling a first intelligent agent to obtain object information and historical operation information of a target object, and determining an object label of the target object when browsing content according to the object information and the historical operation information, the historical operation information being operation information when the target object browses content; calling a second intelligent agent to obtain a plurality of content segments corresponding to the content to be analyzed and content text corresponding to the content to be analyzed, and determining an abnormal risk level and a content risk description of the content to be analyzed according to the plurality of content segments and the content text, the content text being text used to describe the content of the content to be analyzed; calling a third intelligent agent to determine an analysis result of the content to be analyzed according to the object label, the abnormal risk level, and the content risk description, the analysis result being used to indicate an abnormal type and an abnormal reason of the content to be analyzed, the first intelligent agent, the second intelligent agent, and the third intelligent agent being obtained through joint training.

[0006] In another aspect, the embodiments of the present disclosure also provide a content analysis device, comprising: The first obtaining module is configured to invoke a first agent to obtain object information and historical operation information of a target object, and determine an object label of the target object when browsing content according to the object information and the historical operation information, wherein the historical operation information is operation information when the target object browses the content. The second obtaining module is configured to invoke a second agent to obtain a plurality of content segments corresponding to the content to be analyzed and content text corresponding to the content to be analyzed, and determine an abnormal risk level and a content risk description of the content to be analyzed according to the plurality of content segments and the content text, wherein the content text is text used to describe content of the content to be analyzed. The determining module is configured to invoke a third agent to determine an analysis result of the content to be analyzed according to the object label, the abnormal risk level and the content risk description, wherein the analysis result is used to indicate an abnormal type and an abnormal reason of the content to be analyzed, and the first agent, the second agent and the third agent are obtained through joint training.

[0007] Further, the historical operation information includes a viewing rate of the target object for each content and feedback content of the target object for the content. The first obtaining module is specifically configured to: filter out first target content with a viewing rate lower than a first viewing threshold; perform semantic clustering on a content title of the first target content to obtain a plurality of negative title clusters; perform semantic clustering on content segment features of the first target content to obtain a plurality of negative visual feature clusters; analyze the negative title clusters, the negative visual feature clusters, feedback content of the target object for the first target content and the object information by using the first agent to obtain the object label of the target object when browsing the content.

[0008] Further, the first obtaining module is specifically configured to: for each first target content, perform semantic vector conversion on a content title of the first target content to obtain a plurality of title semantic vectors; perform vector clustering on the title semantic vectors by using a clustering algorithm to obtain a plurality of negative title clusters, and the semantics of the plurality of title semantic vectors in each negative title cluster are similar.

[0009] Further, the first obtaining module is specifically configured to: for each first target content, perform uniform sampling on the first target content to obtain a plurality of content segments corresponding to the first target content; extract content segment features in the content segments by using a visual encoder. The content segment features corresponding to each of the first target content are clustered into vectors by a clustering algorithm to obtain a plurality of negative visual feature clusters, and the semantics of the plurality of content segment features in each of the negative visual feature clusters are similar.

[0010] Further, the first acquisition module is specifically configured to: When the historical operation information is updated, a second target content with a viewing rate lower than a first viewing threshold is filtered out from the newly added historical operation information. The content titles of the second target content are semantically clustered to obtain a plurality of newly added negative title clusters. The content segment features of the second target content are semantically clustered to obtain a plurality of newly added negative visual feature clusters. The object label is updated based on the newly added negative title clusters and the newly added negative visual feature clusters.

[0011] Further, the second acquisition module is specifically configured to: Extract content elements in the content text, and identify abnormal elements in the content segment, the content elements indicating content with actual semantics, and the abnormal elements indicating visual content with risks; Determine a first abnormal risk content associated with the content elements, a risk level corresponding to the content elements, a second abnormal risk content associated with the abnormal elements, and a risk level corresponding to the abnormal elements, respectively, by using a scene knowledge graph, wherein the scene knowledge graph is a knowledge graph used to describe the association between abnormal risk content clusters and object clusters; Determine an abnormal risk level of the content to be analyzed based on the risk level corresponding to the content elements and the risk level corresponding to the abnormal elements, and determine the content risk description based on the first abnormal risk content and the second abnormal risk content.

[0012] Further, the second acquisition module is specifically configured to: Obtain multi-platform negative feedback text, the multi-platform negative feedback text being negative feedback content published by a plurality of objects on a platform; Extract content entities and object entities in the multi-platform negative feedback text; Analyze the association between the object entities and the content entities based on the multi-platform negative feedback text, and determine a risk level of the association between the object entities and the content entities; determine the content entity and the object entity as nodes, determine the association between the object entity and the content entity as node association information, determine the risk level of the association between the object entity and the content entity as node association attribute information, and construct a scene knowledge graph based on the nodes, the node association information, and the node association attribute information.

[0013] Further, the second obtaining module is specifically configured to: determine a first content entity corresponding to the content element and a second content entity corresponding to the abnormal element in the scene knowledge graph; determine, based on the scene knowledge graph, a first object entity associated with the first content entity, an association between the first content entity and the first object entity, and a risk level of the association between the first content entity and the first object entity, and determine the risk level of the association between the first content entity and the first object entity as a risk level corresponding to the content element, and determine the association between the first content entity and the first object entity as the first abnormal risk content; determine, based on the scene knowledge graph, a second object entity associated with the second content entity, an association between the second content entity and the second object entity, and a risk level of the association between the second content entity and the second object entity, and determine the risk level of the association between the second content entity and the second object entity as a risk level corresponding to the abnormal element, and determine the association between the second content entity and the second object entity as the second abnormal risk content.

[0014] Further, the second obtaining module is specifically configured to: extract a plurality of content segment features in each content segment; perform abnormal element identification processing on the content segment features to obtain the abnormal element.

[0015] Further, the determining module is specifically configured to: analyze the object label, the abnormal risk level, and the content risk description by a third intelligent agent to obtain a conflict judgment on the object label and the content risk description, an abnormal type of the content to be analyzed, and an abnormal reason; generate the analysis result based on the conflict judgment on the object label and the content risk description, the abnormal type of the content to be analyzed, and the abnormal reason.

[0016] Further, the determining module is specifically configured to: obtain a multi-platform negative feedback text, content text of multi-platform negative feedback content to which the multi-platform negative feedback text is directed, and an object label of an object publishing the multi-platform negative feedback text; analyze the multi-platform negative feedback text and the content text of the multi-platform negative feedback content by using a large language model to obtain a conflict judgment of the object label of the object publishing the multi-platform negative feedback text and the content text of the multi-platform negative feedback content, an abnormal type of the multi-platform negative feedback content, and an abnormal reason; use the content text of the multi-platform negative feedback content, the object label of the object publishing the multi-platform negative feedback text, the conflict judgment, the abnormal type of the multi-platform negative feedback content, and the abnormal reason as training data to train the third intelligent agent.

[0017] Further, the determination module is specifically configured to: split the training data into fine-tuning training data and reinforcement training data; perform supervised fine-tuning on the third intelligent agent based on the fine-tuning training data; perform reinforcement fine-tuning on the third intelligent agent after the supervised fine-tuning based on the reinforcement training data.

[0018] Further, the determination module is specifically configured to: use the content text and the object label in the reinforcement training data as input data to input into the third intelligent agent to obtain an output result of the third intelligent agent, the output result including a conflict judgment of the object label of the object publishing the multi-platform negative feedback text and the content text, an abnormal type of the multi-platform negative feedback content, and an abnormal reason; when the conflict judgment is that the object label of the object publishing the multi-platform negative feedback text and the content text exist conflict, adjust model parameters of the third intelligent agent based on a first reward weight; when the conflict judgment is that the object label of the object publishing the multi-platform negative feedback text and the content text exist conflict, and the abnormal type of the multi-platform negative feedback content is consistent with the abnormal type in the reinforcement training data, adjust model parameters of the third intelligent agent based on a second reward weight, and the second reward weight is greater than the first reward weight; When the conflict is determined to exist between the object label of the object publishing the multi-platform negative feedback text and the content text, and the abnormal type of the multi-platform negative feedback content is consistent with the abnormal type in the reinforcement training data, and the abnormal reason of the multi-platform negative feedback content is consistent with the abnormal reason in the reinforcement training data, the model parameters of the third agent are adjusted based on a third reward weight, and the third reward weight is greater than the second reward weight. Return to the step of inputting the content text and the object label in the reinforcement training data as input data into the third agent until the third agent converges.

[0019] Further, the determination module is specifically configured to: input the content text and the object label in the reinforcement training data as input data into a reference model to obtain an output result of the reference model, the reference model being a third agent after supervised fine-tuning; determine a deviation penalty weight based on the difference between the output result of the reference model and the output result of the third agent; adjust the model parameters of the third agent based on the deviation penalty weight.

[0020] Further, the second acquisition module is specifically configured to: acquire a sampling time interval; sample the to-be-analyzed content based on the sampling time interval to obtain a plurality of content segments corresponding to the to-be-analyzed content.

[0021] In another aspect, the embodiments of the present disclosure also provide an electronic device, including a memory and a processor, the memory stores a computer program, and the processor implements the content analysis method described above when executing the computer program.

[0022] In another aspect, the embodiments of the present disclosure also provide a computer readable storage medium, the storage medium stores a computer program, and the computer program is executed by a processor to implement the content analysis method described above.

[0023] In another aspect, the embodiments of the present disclosure also provide a computer program product, which includes a computer program stored in a computer readable storage medium. The processor of a computer device reads the computer program from the computer readable storage medium, and the processor executes the computer program, so that the computer device executes the content analysis method described above.

[0024] The embodiments of the present disclosure at least have the following beneficial effects: by calling the first agent to obtain the object information and the historical operation information of the target object, and determining the object label of the target object when browsing the content according to the object information and the historical operation information, the operation information of the target object when watching the liked content and the operation information of the target object when watching the disliked content can be determined. Then, by calling the second agent to obtain the plurality of content segments corresponding to the content to be analyzed and the content text corresponding to the content to be analyzed, and determining the abnormal risk level and the content risk description of the content to be analyzed according to the plurality of content segments and the content text, the abnormal content existing in the content to be analyzed and the risk degree of the abnormal content can be comprehensively understood from the visual level and the text content level, so that the accuracy of the content analysis is effectively improved. Further, because different objects have different tolerance degrees for the abnormal content in the content to be analyzed, some objects can accept the abnormal content in the content to be analyzed, and some objects cannot accept the abnormal content in the content to be analyzed. Based on this, the third agent can be called to analyze the abnormal risk level and the content risk description in combination with the object label of the target object when browsing the content obtained previously, to judge the acceptance degree of the target object for the content to be analyzed, and obtain the final analysis result. If the target object cannot accept the content to be analyzed, the analysis result will detail the abnormal type and the abnormal reason of the content to be analyzed, to comprehensively analyze the reason why the target object cannot accept the content to be analyzed. Compared with the existing scheme of determining the specific content causing the negative feedback of the target object by artificial, the embodiments of the present disclosure jointly train the first agent, the second agent and the third agent, so that the content to be analyzed can be analyzed from the level of multiple modalities without human intervention, the efficiency of the content analysis is improved, and by introducing the processing mode of the multiple agents, different agents are responsible for different analysis and processing, so that more refined analysis can be realized under the premise of mutual cooperation, thereby improving the accuracy of the content analysis.

[0025] Other features and advantages of the present disclosure will be set forth in the following description, and in part will become apparent from the description, or can be learned by practice of the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0026] The accompanying drawings are included to provide a further understanding of the technical solutions of the present disclosure, and constitute a part of the specification, and are used to explain the technical solutions of the present disclosure together with the embodiments of the present disclosure, and do not constitute a limitation on the technical solutions of the present disclosure.

[0027] Figure 1 A schematic diagram of an optional implementation environment provided for the embodiments of the present disclosure; Figure 2 A schematic diagram of the content analysis method provided for the embodiments of the present disclosure applied to news content analysis; Figures 3A-3B A schematic diagram of the content analysis method provided by the embodiment of the present disclosure applied to short video analysis is shown in FIG. 1; Figure 4 An optional flowchart of the content analysis method provided by the embodiment of the present disclosure is shown in FIG. 2; Figure 5 A schematic diagram of determining the analysis result of the content to be analyzed according to the object label, the abnormal risk level and the content risk description provided by the embodiment of the present disclosure is shown in FIG. 3; Figure 6 An optional flowchart of calling the first intelligent agent to determine the object label of the target object when browsing the content according to the object information and the historical operation information provided by the embodiment of the present disclosure is shown in FIG. 4; Figure 7 An optional flowchart of calling the second intelligent agent to determine the abnormal risk level and the content risk description of the content to be analyzed according to the multiple content segments and the content text provided by the embodiment of the present disclosure is shown in FIG. 5; Figure 8 An optional flowchart of calling the third intelligent agent to determine the analysis result of the content to be analyzed according to the object label, the abnormal risk level and the content risk description provided by the embodiment of the present disclosure is shown in FIG. 6; Figure 9 A schematic diagram of training the third intelligent agent provided by the embodiment of the present disclosure is shown in FIG. 7; Figure 10 An optional architecture diagram of the content analysis method provided by the embodiment of the present disclosure is shown in FIG. 8; Figure 11 An optional structure diagram of the content analysis device provided by the embodiment of the present disclosure is shown in FIG. 9; Figure 12 A partial structure block diagram of the terminal provided by the embodiment of the present disclosure is shown in FIG. 10; Figure 13 A partial structure block diagram of the server provided by the embodiment of the present disclosure is shown in FIG. 11. DETAILED DESCRIPTION

[0028] In order to make the objects, technical solutions and advantages of the present disclosure clearer, the present disclosure will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present disclosure and not to limit the present disclosure.

[0029] It should be noted that in various specific embodiments of the present disclosure, when relevant processing needs to be performed on data related to the characteristics of the target object, such as target object attribute information or attribute information set, the permission or consent of the target object is obtained first, and the collection, use and processing of the data comply with relevant laws, regulations and standards. Among them, the target object can be a user. In addition, when the embodiments of the present disclosure need to obtain target object attribute information, the separate permission or separate consent of the target object is obtained through a pop-up window or by jumping to a confirmation page, and after obtaining the separate permission or separate consent of the target object, the necessary target object related data for enabling the embodiments of the present disclosure to normally run is obtained.

[0030] In the embodiments of the present disclosure, 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 target, and can be implemented entirely or partially by using software, hardware (such as a processing circuit or a memory) or a combination thereof. Similarly, one 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 a part of an overall module or unit that includes the functions of the module or unit.

[0031] At present, in the video recommendation scene, many times the content recommended to the user is actually not liked by the user, which will produce various negative feedbacks, such as quickly swiping away, clicking the "dislike" control, and clicking the "report complaint" control. After the user produces negative feedback, the business personnel needs to view the videos that the user has historically browsed and the videos that trigger the user's negative feedback to determine the specific content that triggers the user's negative feedback (for example: the video content conflicts with the user's values), so as to reduce the recommendation of related videos for the user. But this way is very time-consuming, and in the case of a large number of negative feedbacks, it will bring a huge workload to the business personnel.

[0032] Based on this, the embodiments of the present disclosure provide a content analysis method and device, electronic equipment and storage medium, which can accurately determine the abnormal type and abnormal reason of the content to be analyzed without human intervention.

[0033] Reference Figure 1 , Figure 1 An optional implementation environment provided by the embodiments of the present disclosure is shown in a schematic diagram, which includes a terminal 101 and a server 102, wherein the terminal 101 and the server 102 are connected through a communication network.

[0034] Exemplarily, the server 102 can acquire object information and historical operation information of the target object in response to the analysis request sent by the terminal 101, and determine the object label of the target object when browsing the content according to the object information and the historical operation information. Then, a plurality of content segments corresponding to the content to be analyzed and content text corresponding to the content to be analyzed are acquired, and the abnormal risk level and the content risk description of the content to be analyzed are determined according to the plurality of content segments and the content text. Finally, the analysis result of the content to be analyzed is determined according to the object label, the abnormal risk level and the content risk description.

[0035] The server 102 acquires the object information and the historical operation information of the target object by calling the first intelligent agent, and determines the object label of the target object when browsing the content according to the object information and the historical operation information. The operation information of the target object when watching the favorite content and the operation information of the target object when watching the disliked content can be determined. Then, the plurality of content segments corresponding to the content to be analyzed and the content text corresponding to the content to be analyzed are acquired by calling the second intelligent agent, and the abnormal risk level and the content risk description of the content to be analyzed are determined according to the plurality of content segments and the content text. The abnormal content existing in the content to be analyzed and the risk degree of the abnormal content can be comprehensively understood from the visual level and the text content level, so that the accuracy of the content analysis is effectively improved. Further, because different objects have different tolerance degrees for the abnormal content in the content to be analyzed, some objects can accept the abnormal content in the content to be analyzed, and some objects cannot accept the abnormal content in the content to be analyzed. Based on this, the third intelligent agent can be called to analyze the abnormal risk level and the content risk description in combination with the object label of the target object when browsing the content to determine the acceptance degree of the target object for the content to be analyzed, and obtain the final analysis result. If the target object cannot accept the content to be analyzed, the analysis result will detail the abnormal type and the abnormal reason of the content to be analyzed, to comprehensively analyze the reason why the target object cannot accept the content to be analyzed. Compared with the existing scheme of determining the specific content causing the negative feedback of the target object by artificial means, the embodiments of the present disclosure jointly train the first intelligent agent, the second intelligent agent and the third intelligent agent, which can not only analyze the content to be analyzed from the level of multiple modalities without human intervention, improve the accuracy of content analysis, but also determine the abnormal type and the abnormal reason of the content to be analyzed in combination with the object label of the target object when browsing the content, so that the analysis result is more in line with the preference of the target object when browsing the content. Moreover, by introducing the processing mode of multiple intelligent agents, different intelligent agents are responsible for different analysis and processing, which can realize more refined analysis under the premise of mutual cooperation, thereby improving the accuracy of content analysis.

[0036] The server 102 can be a standalone physical server, a server cluster composed of multiple physical servers, or a distributed system, and can also be a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and basic cloud computing services such as big data and artificial intelligence platforms. In addition, the server 102 can also be a node server in a blockchain network.

[0037] The terminal 101 can be a mobile phone, a computer, a smart voice interaction device, a smart home appliance, a vehicle-mounted terminal, and the like, but is not limited thereto. The terminal 101 and the server 102 can be directly or indirectly connected through wired or wireless communication, and the present disclosure is not limited thereto.

[0038] The present disclosure can be applied in various scenarios, such as news content analysis. Referring to Figure 2 , Figure 2 is a schematic diagram of the content analysis method provided by the present disclosure applied to news content analysis. When using the news content analysis software to which the content analysis method of the present disclosure is applied, the object can enter the news analysis requirement (generally used to analyze whether the video contains controversial content) through the input box at the top of the interface, and the number of the news in the system or the news video file that needs to be analyzed. The question entered by the object can be “please analyze this news 20250916V05I0100, please briefly describe the video and analyze whether the video contains controversial content”. After receiving the question entered by the object and the news video, the system will perform deep understanding and analysis on the news video based on the question entered by the object, and output the analysis result of the news video. As shown in Figure 2 , after analyzing the news video with the number 20250916V05I0100, the system can obtain the analysis result as “the video describes …… There is no controversial content in the video, and the discussion is about a topic of personal health balance.” Through the analysis result, the object can quickly understand the content of the news video and clearly know whether the news video contains controversial content.

[0039] Further, the present disclosure can also be used to analyze a large number of short videos to analyze whether there is controversial content in the short videos. Referring to Figure 3A , the title of the content to be analyzed can be obtained as “xxxx”, the classification tags of the content to be analyzed in the system are “xx”, “xx”, “xxx” and “xx”, and the multiple video frames corresponding to the content to be analyzed are as shown in Figure 3AAs shown, the system can obtain the analysis result as "the video shows a cooking process, which contains xxxxx. This processing method may cause discomfort to some people, especially xxxxx, in addition, the video shows xxxxx, which may cause visual discomfort to some audiences. The video contains controversial elements, including uncomfortable visual pictures" after analyzing the video. The analysis result of whether the video contains negative energy (negative energy refers to the label of the objective attribute of the content, which judges whether the content contains negative elements such as "violation, bad guidance, controversial information", which is the qualitative evaluation of whether the video content itself has a problem) is xx. The analysis result of whether the video contains negative body sensation (negative body sensation refers to the evaluation of the negative feeling that the object may produce after watching the content) is xx. Through the analysis result of the system for the video Figure 3A , it can be determined that the video may cause visual discomfort.

[0040] Referring to Figure 3B , the title of the content to be analyzed can be obtained as "xxxxxxxx", the classification tags of the content to be analyzed in the system are "xxx", "xxx", "xxx", "xxxx" and "xxx", and the multiple video frames corresponding to the content to be analyzed are as shown in Figure 3B . The system can obtain the analysis result as "the content title mentions "xxxxx", which may involve the sensitive topics of xxx and xxx, and also mentions "xxxxx", which may involve the complex relationship between xx and xx. From these information, the video may contain some negative values, such as the controversial elements in the video, and the negative values." The analysis result of whether the video contains negative energy is xx. The analysis result of whether the video contains negative body sensation is xx. Through the analysis result of the system for the video Figure 3B , it can be determined that the video may contain negative values, which may cause discomfort to the audience.

[0041] Referring to Figure 4 , Figure 4 is an optional flowchart of a content analysis method provided by the embodiments of the present disclosure. The content analysis method can be executed by a server, or can be executed by a terminal, or can be executed by a server in cooperation with a terminal. The content analysis method includes but is not limited to the following steps 401 to 403.

[0042] Step 401: calling a first intelligent agent to obtain object information and historical operation information of a target object, and determining object labels of the target object when browsing the content according to the object information and the historical operation information.

[0043] The target object refers to a user who has a content analysis requirement. The historical operation information is operation information of the target object when browsing the content. The content can be a video, an article, or a webpage. For the convenience of description, the embodiments of the present disclosure are described in the case of content being a video. The historical operation information includes a watching rate of the target object for each content, a fast-forward behavior of the target object for each content, and feedback content of the target object for the content. Specifically, the watching rate refers to the watching completion degree of the target object for the content. For example, the playing time of a video is ten minutes in total, and the target object only watches the first three minutes of the video content. Then, the watching rate of the target object for the content is 0.3. Or, the playing time of a video is ten minutes in total, and the target object watches the first nine minutes of the video content. Then, the watching rate of the target object for the content is 0.9. The watching rate can clearly indicate the preference degree of the target object for various contents. Generally, the higher the watching rate, the higher the preference degree of the target object for the content. The fast-forward behavior can be whether the target object uses the fast-forward play function when watching the video content. For a single video content, if the target object has a fast-forward behavior when watching the video content, it indicates that the target object may not like some scenes in the video content or feel that the video content is too slow. If the target object does not have a fast-forward behavior when watching the video content, it indicates that the preference degree of the target object for the video content is high. The feedback content of the target object for the content refers to the view of the target object for the content. The feedback content includes positive feedback content and negative feedback content. The positive feedback content indicates the preference of the target object for the content, and the negative feedback content indicates the aversion of the target object for the content. For example, the video content is the processing and roasting process of a certain meat, and the feedback content of the target object for the video content is “too disgusting, I feel uncomfortable in my heart after watching it”. It can be determined that the feedback content is negative feedback content. Or, the video content is the blooming process of a cactus, and the feedback content of the target object for the video content is “too beautiful”. It can be determined that the feedback content is positive feedback content.

[0044] It can be understood that the object information of the target object refers to basic attribute information of the target object. The object information of the target object can include the work of the target object and the interest and hobby of the target object, and the like. For example, referring to Figure 5 , the object information of the target object is “work is XX, and interest is sports events and movie clips”. Through the work and interest of the target object, the preference of the target object for various contents and the content type that may make the target object feel uncomfortable can be preliminarily inferred.

[0045] It should be noted that the object label refers to the tendency of the target object to various types of content, for example, the type of content that the target object likes, and the type of content that the target object dislikes. Since the object information indicates the basic attribute information of the target object, the historical operation information indicates various operation information when the target object browses the content, by calling the first intelligent agent to analyze the object information and the historical operation information, the tendency of the target object when watching various types of content can be clearly determined. Among them, the first intelligent agent refers to a deep learning model used for in-depth understanding and analysis of object information and historical operation information, so as to analyze the tendency of the target object to various types of content.

[0046] Exemplarily, with reference to Figure 5 If the object information of the target object is "work is XX, interest is sports events and movie clips", the historical operation information of the target object is that the fast-forward action is performed on the content corresponding to content title 2, the content corresponding to content title 2 is a scene of a sports match, and the content implies that team A wins the game through the improper judgment of the referee, there is controversial content. Based on this, by calling the first intelligent agent to analyze the historical operation information and the object information, the object label can be obtained as "the object is more inclined to 'neutral and objective, focus on competition' to sports event related content, and the tolerance of controversial and biased sports content is low".

[0047] Step 402: calling the second intelligent agent to obtain a plurality of content segments corresponding to the content to be analyzed and content text corresponding to the content to be analyzed, and determining the abnormal risk level and content risk description of the content to be analyzed according to the plurality of content segments and the content text.

[0048] Among them, the content to be analyzed can be an article, a video or a webpage that the target object needs the system to analyze, or an article, a video or a webpage that the target object has not browsed and will be pushed to the target object. The source, format and encoding format of the content to be analyzed are not limited in the embodiment of the disclosure, for example, in the case of video content to be analyzed, the content to be analyzed can be from a camera device, screen recording, software production or model generation, etc. The video format of the content to be analyzed can be QuickTime Movie (MOV), Audio Video Interleave (AVI) or Matroska Video (MKV), etc. The encoding format of the content to be analyzed can be H.264 or H.265, etc.

[0049] It can be understood that the content text is a text for describing the content of the content to be analyzed, and the content text can include a natural language description, a keyword description, or a structured description, etc. For example, if a cooking process of a certain meat is shot in the content to be analyzed, the content text of the content to be analyzed can be "a video shows a cooking process, which contains the processing of a certain meat, the certain meat is cut and marinated with spices, then it is threaded on a bamboo stick, and finally it is roasted on a stove." Through the content text corresponding to the content to be analyzed, the core theme, key pictures and viewpoint tendency of the content to be analyzed can be directly summarized, and the accuracy of content understanding is improved.

[0050] It can be understood that in the case of the content to be analyzed being a web page, the content segment refers to a picture or a video in the web page. In the case of the content to be analyzed being a video, the content segment refers to a video frame. Here, the case of the content being a video is expanded: the content to be analyzed is composed of a plurality of video frames arranged in chronological order, and a video frame is used to represent the picture of the video at a certain time. Moreover, each video frame contains rich visual information, which can include objects, scenes, colors, actions, etc., and semantic information can include object relationships, action meanings, event backgrounds, etc. Exemplarily, the plurality of video frames corresponding to the content to be analyzed can be as shown in Figure 3A , and the plurality of video frames corresponding to the content to be analyzed can also be as shown in Figure 3B . Since the content to be analyzed itself is unstructured multi-modal data, many key visual information of the content to be analyzed cannot be accurately described by the content text, or is easily ignored. Through the plurality of content segments corresponding to the content to be analyzed, details and features not mentioned in the content text can be captured, and the content to be analyzed can be understood in depth. Moreover, the content text of part of the content to be analyzed can have a description bias, and through the plurality of content segments of the content to be analyzed, this description bias can be corrected, and the authenticity of the understanding of the content to be analyzed is ensured.

[0051] It should be noted that the abnormal risk level of the content to be analyzed refers to the degree of influence of the content exhibited by the content to be analyzed on various specific groups of people (groups of people who feel uncomfortable about the exhibited content). For example, the abnormal risk level of the content to be analyzed that shoots a certain meat barbecue for animal lovers is high risk. The abnormal risk level of the content to be analyzed with an exaggerated title for rational objects is medium risk. Through the abnormal risk level of the content to be analyzed, the influence degree and controversy degree of the content to be analyzed on different groups of people can be determined. The content risk description of the content to be analyzed refers to a structured description of the events occurring in the content to be analyzed and the content that may exist in the content to be analyzed. Through the content risk description, the controversy and risk existing in the content to be analyzed can be understood.

[0052] Further, the second intelligent agent refers to a deep learning model for deep understanding and analysis of the content segments and the content text, so as to analyze the risk content possibly existing in the content to be analyzed, and the influence degree and the controversy degree of the content to be analyzed on different groups of people. After the visual information (i.e., the content segments corresponding to the content to be analyzed) and the content description (i.e., the content text of the content to be analyzed) of the content to be analyzed are determined, the second intelligent agent can be called to analyze the multiple content segments and the content text, so as to determine the abnormal risk level and the content risk description of the content to be analyzed. Through such processing, the abnormal content existing in the content to be analyzed and the risk degree of the abnormal content can be comprehensively understood from the visual level and the text content level, and the accuracy of the content analysis is effectively improved.

[0053] Exemplarily, if there is a content to be analyzed, the content segments corresponding to the content to be analyzed contain pictures of certain meat, cutting with a knife, marinating certain meat, etc., and the content text corresponding to the content to be analyzed is "this is an outdoor barbecue tutorial of certain meat, which contains the processing of certain meat, certain meat is cut into pieces and marinated with seasonings, then is threaded on bamboo sticks, and finally is roasted on a stove". By calling the second intelligent agent to analyze the multiple content segments and the content text, it can be determined that the abnormal risk level of the content to be analyzed is high risk, and the content risk description of the content to be analyzed is "this is an outdoor barbecue tutorial of certain meat, which contains the processing of certain meat, certain meat is cut into pieces and marinated with seasonings, then is threaded on bamboo sticks, and finally is roasted on a stove. This video involves animal slaughtering and eating, and the picture of slaughtering animals is relatively bloody, which may cause discomfort of animal protectors".

[0054] Step 403: calling the third intelligent agent to determine the analysis result of the content to be analyzed according to the object label, the abnormal risk level and the content risk description.

[0055] It can be understood that because different objects have different tolerance levels for abnormal content in the content to be analyzed, some objects can accept abnormal content in the content to be analyzed, and some objects cannot accept abnormal content in the content to be analyzed. Based on this, a third intelligent agent can be called to analyze the abnormal risk level and the content risk description in combination with the object label obtained previously when the target object browses the content, to judge the acceptance degree of the target object for the content to be analyzed, and obtain the final analysis result. The third intelligent agent refers to a deep learning model for in-depth understanding and analysis of the object label, the abnormal risk level and the content risk description, so as to analyze the abnormal type and the abnormal reason of the content to be analyzed. Moreover, the first intelligent agent, the second intelligent agent and the third intelligent agent are obtained by joint training. Through such setting, the three hierarchical intelligent agents can work together to realize the accurate prediction and explanation of negative feedback. The analysis result is used to indicate the abnormal type and the abnormal reason of the content to be analyzed, and the analysis result of the content to be analyzed includes positive explanation and negative explanation. If the analysis result of the content to be analyzed is positive explanation, it means that there is no conflict between the object label and the content risk description, and the target object can accept the abnormality in the content to be analyzed. At this time, the analysis result does not include the abnormal type and the abnormal reason, but only indicates the specific content of the content to be analyzed, and explains that there is no controversial content in the content to be analyzed. If the analysis result of the content to be analyzed is negative explanation, it means that there is a conflict between the object label and the content risk description, and the target object cannot accept the abnormality in the content to be analyzed. At this time, the analysis result includes the abnormal type and the abnormal reason of the content to be analyzed.

[0056] Further, the abnormal type of the content to be analyzed refers to the negative category system to which the controversial risk of the content to be analyzed belongs. These negative categories include visual interference types (such as bloody pictures, thrilling shots, blurred pictures), idea conflict types (such as animal protectors seeing animals eating, environmentalists seeing pollution behavior), content mismatch types (such as objects who like hardcore technology seeing popular introductory content), misleading title types (such as the title writing "advanced baking tutorial", which is actually a zero-base content), etc. Through the abnormal type of the content to be analyzed, the reason why the content to be analyzed causes negative feedback of the target object can be structured.

[0057] The abnormal reason of the content to be analyzed refers to a natural language explanation of a reason why the target object cannot accept the content to be analyzed, and the abnormal reason can comprehensively analyze the reason why the target object cannot accept the content to be analyzed by integrating the visual level of the content to be analyzed and the semantic content level of the content to be analyzed. For example, as shown above, the content risk description of the content to be analyzed is "this is an outdoor barbecue tutorial of a certain meat, which contains the processing of a certain meat, the meat is cut into pieces and marinated with seasonings, then it is threaded on bamboo sticks, and finally it is roasted on the stove. This video involves animal slaughter and consumption, and the picture of slaughtering animals is relatively bloody, which may cause discomfort to animal protectors". In the case where the object label of the target object is an animal protector, it can be determined that the abnormal reason of the content to be analyzed is "the target object is an animal protector, and the video contains the picture of roasting a certain meat, which may cause psychological discomfort to the target object".

[0058] Based on this, the object information and the historical operation information of the target object are obtained by calling the first agent, and the object label when the target object browses the content is determined according to the object information and the historical operation information, so that the operation information of the target object when watching the liked content and the operation information of the target object when watching the disliked content can be clearly determined. Then, the plurality of content segments corresponding to the content to be analyzed and the content text corresponding to the content to be analyzed are obtained by calling the second agent, and the abnormal risk level and the content risk description of the content to be analyzed are determined according to the plurality of content segments and the content text, so that the abnormal content existing in the content to be analyzed and the risk degree of the abnormal content can be comprehensively understood from the visual level and the text content level, and the accuracy of the content analysis is effectively improved. Further, because different objects have different tolerance degrees for the abnormal content in the content to be analyzed, some objects can accept the abnormal content in the content to be analyzed, and some objects cannot accept the abnormal content in the content to be analyzed. Based on this, the third agent can be called to analyze the abnormal risk level and the content risk description in combination with the object label when the target object browses the content to determine the acceptance degree of the target object for the content to be analyzed, and obtain the final analysis result. If the target object cannot accept the content to be analyzed, the analysis result will detail the abnormal type and the abnormal reason of the content to be analyzed to comprehensively analyze the reason why the target object cannot accept the content to be analyzed. Compared with the existing scheme of determining the specific content causing the negative feedback of the target object by artificial means, the embodiments of the present disclosure jointly train the first agent, the second agent and the third agent, which can not only analyze the content to be analyzed from the level of multiple modalities without human intervention, improve the accuracy of content analysis, but also determine the abnormal type and the abnormal reason of the content to be analyzed in combination with the object label when the target object browses the content, so that the analysis result is more in line with the preference of the target object when browsing the content. Moreover, by introducing the processing mode of multiple agents, different agents are responsible for different analysis and processing, which can realize more refined analysis under the premise of mutual cooperation, thereby improving the accuracy of content analysis.

[0059] It is worth noting that the historical operation information includes the watching rate of the target object for each content and the feedback content of the target object for the content. The watching rate of the target object for each content reflects the preference degree of the target object for different types of content, and the feedback content of the target object for the content reflects the most intuitive view of the target object on each content. Based on this, the object label when the target object browses the content can be accurately determined according to the object information and the historical operation information, and the process of determining the object label when the target object browses the content is described in detail as follows: In a possible implementation, the object label when the target object browses the content is determined according to the object information and the historical operation information, specifically, the first target content with a viewing rate lower than a first viewing threshold can be screened out; the content titles of the first target content are subjected to semantic clustering to obtain a plurality of negative title clusters; the content segment features of the first target content are subjected to semantic clustering to obtain a plurality of negative visual feature clusters; the first intelligent agent is used to analyze the negative title clusters, the negative visual feature clusters, the feedback content of the target object on the first target content, and the object information to obtain the object label when the target object browses the content.

[0060] The first viewing threshold is the maximum value of the viewing rate of the content that the target object dislikes. The specific value of the first viewing threshold can be determined according to the average value of the viewing rate of the target object on each content. In general, the first viewing threshold can be set to 0.3. For a single content, if the viewing rate of the target object on this content is lower than the first viewing threshold, it means that the target object does not like this content. In order to determine the content information and visual information that the target object specifically dislikes, the first target content with a viewing rate lower than the first viewing threshold can be screened out. Figure 6 Figure 6 The first intelligent agent is used to process the object information and the historical operation information. By invoking the first intelligent agent, the content with a viewing rate lower than the first viewing threshold can be screened out as the content whose content information and visual information need to be analyzed, that is, the first target content.

[0061] It can be understood that semantic clustering is an unsupervised learning method that automatically classifies texts or contents with similar meanings into a category based on the semantic similarity of the texts or contents. In general, the content title of the first target content can describe the event of the first target content in a short language. By subjecting the content titles of the first target contents to semantic clustering, the content titles with similar semantics can be classified into a category to obtain a plurality of negative title clusters. The content titles in different negative title clusters have different semantics, and the content titles in the same negative title cluster have similar semantics. For example, the core semantics of the negative title cluster 1 is animal slaughtering related content, and the content titles in the negative title cluster 1 include “Wild slaughtering chicken skills”, “Live fish present killing method”, etc. The core semantics of the negative title cluster 2 is exaggerated title, and the content titles in the negative title cluster 2 include “100% learn C language”, “Zero foundation monthly income over ten thousand” and the like.

[0062] ​It can be understood that the content segment feature of the first target content refers to a key visual feature in each content segment of the first target content, such as a bloody picture of roasting a certain meat, or a blurred picture, or a picture in which a person is fighting. The content segment feature can be a video frame feature, and the content segment feature of the first target content reflects the key visual elements in the first target content. Since the content segment feature itself is a vector feature, the content segment features of the first target content can be directly clustered in general cases, which can classify content segment features with similar semantics into a category to obtain multiple negative visual feature clusters. The semantics of the content segment features in different negative visual feature clusters are different, and the semantics of the multiple content segment features in the same negative visual feature cluster are similar. For example, the core semantics of the negative visual feature cluster 1 is animal injury visual content, and the content segment features in the negative visual feature cluster 1 include a bloody picture of roasting a certain meat, a live fish killing picture, etc. The core semantics of the negative visual feature cluster 2 is visual content contrary to public order and good customs, and the content segment features in the negative visual feature cluster 2 include a picture of throwing garbage, a picture of damaging public property, a picture of insulting others, etc.

[0063] Further, after determining the multiple negative title clusters and the negative visual feature clusters, the content information and visual information that the target object dislikes can be determined. Through the content information and visual information that the target object dislikes, the target object's tendency for various types of content can be more accurately determined. Based on this, the first intelligent agent can analyze the negative title clusters, the negative visual feature clusters, the feedback content of the target object for the first target content, and the object information to obtain the object label of the target object when browsing content. Based on this information, the first intelligent agent will deeply think about the target object's tendency for various types of content from multiple dimensions such as interest, plot, visual, and negativity, and finally output the object label.

[0064] It should be noted that by performing semantic clustering on the content title of the first target content to obtain multiple negative title clusters, and performing semantic clustering on the content segment feature of the first target content to obtain multiple negative visual feature clusters, the content information and visual information that the target object dislikes can be determined. Further, by using the first intelligent agent to analyze the negative title clusters, the negative visual feature clusters, the feedback content of the target object for the first target content, and the object information to obtain the object label of the target object when browsing content, the target object's tendency for various types of content can be deeply thought from multiple dimensions such as interest, plot, visual, and negativity, and the accuracy of the finally output object label is improved.

[0065] In a possible implementation, the content titles of the first target content are subjected to semantic clustering to obtain a plurality of negative title clusters. Specifically, for each first target content, the content title of the first target content is subjected to semantic vector conversion to obtain a plurality of title semantic vectors; and the title semantic vectors are subjected to vector clustering by a clustering algorithm to obtain the plurality of negative title clusters.

[0066] It can be understood that a prerequisite for semantic clustering is to convert unstructured text into vectors that can be understood by a computer. Therefore, the content title of each first target content needs to be subjected to semantic vector conversion to convert the content title of the first target content into a numerical form that is easier for a computer to process, i.e., a title semantic vector. Moreover, the title semantic vector retains important features and semantic messages of the content title. By converting the content title of the first target content into a title semantic vector, it is easier to implement more complex clustering logic subsequently.

[0067] Further, the more similar the semantics of the content titles are, the more similar the distances of the corresponding title semantic vectors are. Based on this, the clustering algorithm (for example, the K-Means algorithm or the DBSCAN algorithm) can be used to classify the title semantic vectors with similar distances into a class, which can also be understood as classifying the plurality of title semantic vectors with similar distances as a negative title cluster. The plurality of title semantic vectors in each negative title cluster are similar in semantics. In this way, the plurality of negative title clusters can be quickly obtained. This clustering algorithm does not need to define various class labels in advance. Instead, the algorithm automatically classifies the classes according to the semantic similarity of the title semantic vectors, which is suitable for processing a large amount of unannotated content titles of the first target content. Moreover, the content titles of the first target content are unstructured and scattered text collections. Direct analysis of these content titles can only obtain a superficial conclusion that the target object does not like these first target content. By converting the content titles of the first target content into title semantic vectors and performing vector clustering on the title semantic vectors by the clustering algorithm, the title semantic vectors can be automatically grouped according to the semantic similarity to output structured negative title clusters. That is, by processing in this way, the content that can cause discomfort to the target object can be accurately extracted from the scattered negative historical operation information, providing a structured basis for subsequent determination of the object label.

[0068] It should be noted that by converting the content title of the first target content into a title semantic vector, unstructured text can be converted into a numerical form that is easier for a computer to process. By performing vector clustering on the title semantic vectors by the clustering algorithm, the content that can cause discomfort to the target object can be accurately extracted from the scattered negative historical operation information, providing a structured basis for subsequent determination of the object label.

[0069] In a possible implementation, the content segment features of the first target content are semantically clustered to obtain a plurality of negative visual feature clusters. Specifically, for each first target content, the first target content can be uniformly sampled to obtain a plurality of content segments corresponding to the first target content; the content segment features in the content segments are extracted by using a visual encoder; and the content segment features corresponding to each first target content are vector clustered by using a clustering algorithm to obtain a plurality of negative visual feature clusters.

[0070] It can be understood that, in a case where the first target content is a video and the content segments are video frames, uniformly sampling the first target content means that a fixed number of video frames are extracted from an original video frame sequence of the first target content at fixed time intervals to form a video frame sequence that is uniformly distributed in time and has a uniform length, that is, a plurality of continuous video frames sorted in time, that is, a plurality of continuous content segments sorted in time. By uniformly sampling each first target content to obtain a plurality of content segments corresponding to each first target content, first target content of different lengths and different frame rates can be standardized into frame-level data in a uniform format, facilitating subsequent batch processing of the content segments.

[0071] Further, as described above, the premise of semantic clustering is to convert unstructured data into vectors that can be understood by a computer. Based on this, a visual encoder can be used to extract content segment features in each content segment to convert unstructured visual information into structured vector features that can be understood by a computer. For example, a content segment of the first target content is a bloody scene. After feature extraction of the content segment by using the visual encoder, the obtained content segment features include animal carcasses, knives, and red bloodstains. Specifically, a content segment contains a large number of implicit and fine-grained visual details (such as picture color, object shape, motion features, and scene atmosphere). These information can either not be accurately described in text or be easily ignored by humans. By using the visual encoder, not only can core visual entities in the content segment (for example, “animal slaughtering tool”, “player foul action”, “color distribution of bloody scene”, and other key entities) be identified, but also the degree of visual impact in the content segment (for example, the bloodiness, violence, and density of the picture) can be quantified. Through such processing, the limitations of content title description can be compensated for, and more fine-grained visual basis can be provided for content analysis and abnormal risk judgment.

[0072] It can be understood that, since the content segment features are already standardized numerical representations, the higher the semantic similarity, the closer the distance in the vector space, so the content segment features corresponding to each first target content can be directly vector clustered by a clustering algorithm (such as a K-Means algorithm, a DBSCAN algorithm), and the content segment features with close distances are classified into a class as a negative visual feature cluster. The semantic of the plurality of content segment features in each negative visual feature cluster is similar. In this way, a plurality of negative visual feature clusters can be quickly obtained. Such a clustering algorithm does not need to define various category labels in advance, and the algorithm will automatically classify categories according to the semantic similarity of the content segment features, which is suitable for processing the corresponding content segment features of a large amount of unlabeled first target content such as the embodiments of the present disclosure. Moreover, the content segment features corresponding to the first target content are unordered high-dimensional vector sets, and direct analysis cannot derive rules. By vector clustering the content segment features corresponding to the first target content by a clustering algorithm, these visual features can be automatically divided into structured categories, and structured negative visual feature clusters are output.

[0073] Specifically, since the content segment features are quantifiable numerical representations, not only can categories be classified during clustering, but also the impact degree of the same category of content segment features can be judged through the distance or distribution of the content segment features. For example: in a negative visual feature cluster with a core semantic of “bloody picture”, the more similar the distance of the content segment features, the more similar the degree of bloodiness. Compared to semantic clustering of content titles, semantic clustering of content segment features can further classify the abnormality degree of content segment features to provide a more fine-grained visual basis.

[0074] It should be noted that, in the case of the first target content being a video, by uniformly sampling the first target content, a plurality of content segments corresponding to the first target content can be obtained, and the first target content with different lengths and different frame rates can be standardized into frame-level data of a unified format to facilitate subsequent batch processing of the content segments. Further, the content segment features in the content segments are extracted by using a visual encoder, and the content segment features corresponding to each first target content are vector clustered by a clustering algorithm, which can mine content that may trigger discomfort of the target object from a visual level, make up for the limitations of semantic clustering of content titles, and provide a fine-grained visual basis for subsequent accurate determination of object labels.

[0075] In a possible implementation, after obtaining the object label when the target object browses the content, specifically, when the historical operation information is updated, the second target content with a viewing rate lower than the first viewing threshold is filtered out from the newly added historical operation information; the content titles of the second target content are subjected to semantic clustering to obtain a plurality of newly added negative title clusters; the object label is updated based on the newly added negative title clusters and the newly added negative visual feature clusters.

[0076] It can be understood that the preference of the target object for the content is not always constant, but changes over time. The short-term behavior of the target object can better reflect the current true preference of the target object, and the long-term behavior of the target object can contain outdated interests, for example, the target object previously likes swimming content, but recently starts to reject due to a change in interest. Based on this, when the historical operation information is updated, the second target content with a viewing rate lower than the first viewing threshold can be filtered out from the newly added historical operation information (which can be the historical operation information in the last 7 days). The second target content refers to the content that causes the target object to be repelled in the newly added historical operation information. The second target content can most clearly reflect the content that the target object does not like at the moment, for example, the target object has quickly swiped away sports event related video content multiple times in the last 7 days, indicating that the target object currently dislikes sports event related content.

[0077] Further, the content titles of the second target content can be subjected to semantic clustering to classify content titles with similar semantics into a category, to obtain a plurality of newly added negative title clusters, i.e., the newly added negative title clusters. Moreover, the semantics of the content titles in different newly added negative title clusters are different, and the semantics of the multiple content titles in the same newly added negative title cluster are similar. By updating the negative title clusters, the content information that the target object recently dislikes can be determined, so that the tendency of the target object for various types of content in the recent period can be more accurately determined.

[0078] It can be understood that the content segment features of the second target content reflect the key visual elements in the second target content. Since the content segment features are vector features per se, in general, the content segment features of the second target content can be directly clustered, which can classify content segment features with similar semantics into a category to obtain a plurality of newly added negative visual feature clusters, i.e., the newly added negative visual feature clusters. Among them, the semantics of the content segment features in different newly added negative visual feature clusters are different, and the semantics of the multiple content segment features in the same newly added negative visual feature cluster are similar. By updating the negative visual feature clusters, the visual information that the target object recently dislikes can be determined, so that the tendency of the target object for various types of content in the recent period can be more accurately determined.

[0079] After the newly added negative title cluster and the newly added negative visual feature cluster are determined, the newly added negative title cluster and the newly added negative visual feature cluster can be compared with the original object label. Based on the newly added information, the first intelligent agent will deeply think about the target object's tendency for various types of content from multiple dimensions such as interest, plot, visual, and negativity, thereby completing the update of the object label. Specifically, if a new category (for example, a sports event) appears in the newly added negative title cluster or the newly added negative visual feature cluster, the object label is updated according to the newly added category. If the occurrence frequency of a category in the newly added negative title cluster or the newly added negative visual feature cluster increases, the category in the object label is upgraded from weak preference to strong preference. If a category that triggers the target object's aversion is not present in the newly added negative title cluster or the newly added negative visual feature cluster, the description of "the target object's aversion to this category" is deleted from the object label. Through such processing, it can be ensured that the object label can quickly reflect the target object's current tendency for various types of content.

[0080] It should be noted that by screening the second target content with a viewing rate lower than the first viewing threshold from the newly added historical operation information, the content that the target object currently dislikes can be determined. By performing semantic clustering on the content title of the second target content, a plurality of newly added negative title clusters are obtained, and by performing semantic clustering on the content segment features of the second target content, a plurality of newly added negative visual feature clusters are obtained, the content information and visual information that the target object recently dislikes can be clearly obtained. Further, the object label is updated based on the newly added negative title cluster and the newly added negative visual feature cluster, and such processing can ensure that the object label can quickly reflect the target object's current tendency for various types of content.

[0081] In a possible implementation, the plurality of content segments corresponding to the content to be analyzed are obtained, specifically, a sampling time interval can be obtained; the content to be analyzed is sampled based on the sampling time interval to obtain the plurality of content segments corresponding to the content to be analyzed.

[0082] It can be understood that, in the case that the content to be analyzed is a video and the content segment is a video frame, the sampling time interval refers to the time interval between each content segment when the content to be analyzed is sampled. The specific value of the sampling time interval can be determined according to the average length of each content to be analyzed. If the average length of the content to be analyzed is relatively long, in order to avoid too many content segments to be processed, the sampling time interval can be appropriately set to a relatively large value. Correspondingly, if the average length of the content to be analyzed is relatively short, in order to avoid missing important visual information in the content to be analyzed, the sampling time interval can be appropriately set to a relatively small value. Generally, a 16-frame uniform sampling manner can be used to sample the content to be analyzed.

[0083] Further, after determining the sampling time interval, a sequence of uniformly distributed content segments in time can be formed from the original content segment of the content to be analyzed according to the sampling time interval, and the sequence of content segments includes a plurality of continuous content segments. By uniformly sampling each content to be analyzed according to the sampling time interval to obtain a plurality of content segments corresponding to the content to be analyzed, content to be analyzed of different lengths and different frame rates can be standardized into frame-level data of a unified format, so as to facilitate subsequent batch analysis of a plurality of content segments to determine abnormal visual elements in the content segments.

[0084] In a possible implementation, the abnormal risk level and the content risk description of the content to be analyzed are determined according to the plurality of content segments and the content text, specifically, content elements in the content text are extracted, and abnormal elements in the content segments are identified; the first abnormal risk content associated with the content elements, the risk level corresponding to the content elements, the second abnormal risk content associated with the abnormal elements, and the risk level corresponding to the abnormal elements are determined respectively by using the scene knowledge graph; the abnormal risk level of the content to be analyzed is determined based on the risk level corresponding to the content elements and the risk level corresponding to the abnormal elements, and the content risk description is determined based on the first abnormal risk content and the second abnormal risk content.

[0085] In the content text, the content elements refer to the content with actual semantics in the content text. For example, the content text corresponding to the content to be analyzed is "this is an outdoor barbecue tutorial of a certain meat, which includes processing of a certain meat, the certain meat is cut into pieces and marinated with seasonings, then it is threaded on bamboo sticks, and finally it is roasted on the stove." By extracting the content text, it can be determined that the content elements in the content text include "outdoor barbecue", "certain meat", "processing", "cutting" and other word groups with actual semantics. The abnormal elements in the content segments refer to the visual content with risks. For example, the visual content in a content segment of the content to be analyzed is cutting a certain meat, and after identifying the content segment of the content to be analyzed, it can be determined that the abnormal elements in the content segment include "animal carcass", "bloody scene", "knife" and other visual content with risks.

[0086] It can be understood that the scene knowledge graph is a knowledge graph for describing the association between the abnormal risk content cluster and the object cluster. The abnormal risk content cluster refers to a plurality of text contents and visual contents that may cause the target object to be disgusted, such as bloody scenes, knives, and the like. The object cluster refers to a group that may be affected by the abnormal risk content cluster, such as environmentalists, animal protectionists, and the like. The knowledge graph is a graph structure composed of nodes and edges. The nodes in the knowledge graph are used to represent specific entities. The entity refers to a thing or object that has distinguishability and exists independently. For example, the entity can be a person, a place, an event, and the like. In the embodiments of the present disclosure, the entity is an abnormal risk content or an object cluster. The edges between different nodes in the knowledge graph are used to represent the association (for example, “belongs to”, “occurs in”, “is related to”, and the like) between entities and the degree of association (for example, mild, moderate, severe, and the like) between entities. In the embodiments of the present disclosure, the edges between different nodes are used to represent the influence of the abnormal risk content on a certain object cluster and the degree of influence. For example, the abnormal risk content is “bloody scenes”, and the object cluster is “a certain group”. The influence corresponding to the edge between the node corresponding to the abnormal risk content and the node corresponding to the object cluster can be “bloody scenes may cause the certain group to be uncomfortable”, and the corresponding degree of influence can be severe.

[0087] Further, the first abnormal risk content refers to the influence of the content element on a certain object cluster. The risk level corresponding to the content element refers to the degree of influence of the content element on a certain object cluster. As shown in Figure 7 Figure 7 ​For the processing procedure of the second intelligent agent for the plurality of content segments of the content to be analyzed, and the content text corresponding to the content to be analyzed, after the content element and the abnormal element are determined, the second intelligent agent needs to determine the first abnormal risk content associated with the content element and the risk level corresponding to the content element by using the scene knowledge graph. Determining the first abnormal risk content associated with the content element and the risk level corresponding to the content element by using the scene knowledge graph means that the nodes connected with the node corresponding to the content element are found in the scene knowledge graph, the object cluster corresponding to the node is determined, the influence mapped by the edge between the node corresponding to the content element and the node corresponding to the object cluster is determined as the first abnormal risk content, and the degree of the influence mapped by the edge between the node corresponding to the content element and the node corresponding to the object cluster is determined as the risk level corresponding to the content element. For example, the content element is "grilling a certain meat", the node connected with the node "grilling a certain meat" can be determined as "animal lovers" by using the scene knowledge graph, and further, the influence mapped by the edge between the node corresponding to "grilling a certain meat" and the node corresponding to "animal lovers" can be determined as "grilling a certain meat may cause discomfort to animal lovers", and the degree of the influence mapped by the edge between the node corresponding to "grilling a certain meat" and the node corresponding to "animal lovers" is severe. Based on this, the first abnormal risk content can be determined as "grilling a certain meat may cause discomfort to animal lovers", and the risk level corresponding to the content element can be determined as severe.

[0088] Correspondingly, the second abnormal risk content refers to the influence of the abnormal element on a certain object cluster. The risk level corresponding to the abnormal element refers to the degree of the influence of the abnormal element on a certain object cluster. For example, Figure 7As shown, after the content element and the abnormal element are determined, the second abnormal risk content associated with the abnormal element and the risk level corresponding to the abnormal element need to be determined by using the scene knowledge graph. Determining the second abnormal risk content associated with the abnormal element and the risk level corresponding to the abnormal element by using the scene knowledge graph refers to: finding the node connected to the node corresponding to the abnormal element in the scene knowledge graph, determining the object cluster corresponding to the node, determining the influence mapped by the edge between the node corresponding to the abnormal element and the node corresponding to the object cluster as the second abnormal risk content, and determining the degree of the influence mapped by the edge between the node corresponding to the abnormal element and the node corresponding to the object cluster as the risk level corresponding to the abnormal element. For example, the abnormal element is "bloody picture", and the node connected to the node "bloody picture" can be determined as "certain group" by using the scene knowledge graph. Further, the influence mapped by the edge between the node corresponding to "bloody picture" and the node corresponding to "certain group" can be determined as "bloody picture may cause discomfort of certain group", and the degree of the influence mapped by the edge between the node corresponding to "bloody picture" and the node corresponding to "certain group" can be determined as severe. Based on this, the second abnormal risk content can be determined as "bloody picture may cause discomfort of certain group", and the risk level corresponding to the abnormal element can be determined as severe.

[0089] It can be understood that the risk level corresponding to the content element and the risk level corresponding to the abnormal element can be comprehensively judged to determine the abnormal risk level of the content to be analyzed. For example, if the risk level corresponding to the content element and the risk level corresponding to the abnormal element are both light, the abnormal risk level of the content to be analyzed can also be determined as light. If the risk level corresponding to the content element and the risk level corresponding to the abnormal element are both severe, the abnormal risk level of the content to be analyzed can also be determined as severe. If the risk level corresponding to the content element is light and the risk level corresponding to the abnormal element is severe, the abnormal risk level of the content to be analyzed needs to be comprehensively judged in combination with the specific content of the content element and the abnormal element. Further, the first abnormal risk content and the second abnormal risk content can be analyzed by the second intelligent agent to summarize the first abnormal risk content and the second abnormal risk content, and generate the content risk description. By such processing, the abnormal content existing in the content to be analyzed and the risk degree of the abnormal content can be more comprehensively understood from the visual level and the text content level, and the accuracy of content analysis is effectively improved.

[0090] It should be noted that by utilizing the scene knowledge graph, the first abnormal risk content associated with the content element, the risk level corresponding to the content element, the second abnormal risk content associated with the abnormal element, and the risk level corresponding to the abnormal element are determined respectively, the influence caused by the abnormality at the level of text content and the influence caused by the abnormality at the visual level can be quickly determined. Further, by determining the abnormal risk level of the content to be analyzed based on the risk level corresponding to the content element and the risk level corresponding to the abnormal element, by calling the second intelligent agent to analyze the first abnormal risk content and the second abnormal risk content, and generating the content risk description, the abnormal content existing in the content to be analyzed and the risk degree of the abnormal content can be comprehensively understood at the visual level and the text content level.

[0091] In a possible implementation, identifying the abnormal element in the content segment, specifically, for each content segment, a plurality of content segment features in the content segment can be extracted; the content segment features are subjected to abnormal element identification processing to obtain the abnormal element.

[0092] It can be understood that in the case where the content segment is a video frame and the content segment feature is a video frame feature, the content segment is a static visual image, and a single content segment feature cannot fully describe the picture content, nor can it support the judgment of the abnormal element. Based on this, for each content segment, a plurality of content segment features in the content segment can be extracted, and through these multiple dimensional features, the visual attributes of the content segment can be completely described. For example, for a single content segment, object features, scene features, action features, and color features can be extracted simultaneously. In the case where the content segment depicts "outdoor barbecue, using a kitchen knife to cut chicken", the object features are kitchen knife and chicken, the scene features are outdoor, the action features are cutting, and the color features are local red bloodstains. Through such processing, important visual information in the content segment can be avoided.

[0093] In an embodiment, in general, the scene knowledge graph can be utilized to perform abnormal element identification processing on the content segment features. If the content segment feature is consistent with a certain abnormal risk content in the scene knowledge graph, the content segment feature can be determined as an abnormal element.

[0094] In another embodiment, the content segment features can also be subjected to abnormal element identification processing by a large language model. If the large language model confirms that a certain content segment feature may cause discomfort to some people, the content segment feature can be determined as an abnormal element.

[0095] It should be noted that by extracting multiple content fragment features from the content segment and performing abnormal element identification processing on the content fragment features, visual information that may pose a risk can be accurately located, which facilitates the subsequent determination of the content risk description and abnormal risk level of the content to be analyzed.

[0096] In one possible implementation, the scene knowledge graph is constructed as follows: acquiring negative feedback text from multiple platforms; extracting content entities and object entities from the negative feedback text; analyzing the associations between object entities and content entities based on the negative feedback text, and determining the risk level of the associations between object entities and content entities; identifying content entities and object entities as nodes, identifying the associations between object entities and content entities as node association information, identifying the risk level of the associations between object entities and content entities as node association attribute information, and constructing the scene knowledge graph based on each node, node association information, and node association attribute information.

[0097] Multi-platform negative feedback text refers to negative feedback content posted by multiple individuals on various platforms. It can also be understood as negative evaluations of various types of content on the platform. For example, regarding content showing images of grilling meat, multi-platform negative feedback text could be, "The image of cutting this meat makes me feel disgusted." Regarding content showing images of littering, multi-platform negative feedback text could be, "Your littering behavior will set a bad example for children."

[0098] Furthermore, the object entity in multi-platform negative feedback text refers to the group affected by the content. The content entity in multi-platform negative feedback text refers to the textual or visual content that evokes aversion in the target audience. In one embodiment, the multi-platform negative feedback text can be input into a second intelligent agent, which analyzes the text and extracts the content and object entities. This process allows for the rapid extraction of textual and visual content that is averse to different target groups.

[0099] Understandably, the association between an object entity and a content entity refers to the impact of the content entity on the object entity. The risk level of the association between an object entity and a content entity refers to the degree of impact of the content entity on the object entity. For example, the content entity is a scene of cutting meat, and the object entity is an animal rights activist. If the scene of cutting meat causes severe discomfort to the animal rights activist, the association between the object entity and the content entity can be described as the content entity causing discomfort and aversion to the object entity, and the risk level of the association between the object entity and the content entity can be considered severe. Generally, negative feedback text on multiple platforms will directly or indirectly indicate the impact of the content entity on the object entity, and the degree of that impact. Therefore, by analyzing negative feedback text on multiple platforms, the association between the object entity and the content entity, as well as the risk level of that association, can be determined.

[0100] Furthermore, as mentioned above, a scene knowledge graph is a graph structure composed of nodes and edges. Therefore, to construct a scene knowledge graph, it is first necessary to determine the nodes and edges within it. Since nodes in a scene knowledge graph represent specific entities, content entities and object entities can be identified as nodes. Since edges between different nodes in a scene knowledge graph represent the relationships between entities, and the attributes of these edges represent the degree of association between entities, the relationship between object entities and content entities can be determined as node association information (i.e., the edges between nodes), and the risk level of the relationship between object entities and content entities can be determined as node association attribute information (i.e., the attributes of the edges between nodes). After determining the nodes, node association attributes, and node association attribute information, the scene knowledge graph can be quickly constructed based on this information.

[0101] It should be noted that, since the scene knowledge graph pre-stores information such as content entities / object entities, the people affected by the content entities / object entities, and the degree of influence of the content entities / object entities on various groups, the scene knowledge graph facilitates the subsequent accurate mapping of content elements / abnormal elements to the first abnormal risk content / second abnormal risk content, thereby improving the analysis efficiency of the content to be analyzed.

[0102] In one possible implementation, a scene knowledge graph is used to determine the first abnormal risk content associated with the content element, the risk level corresponding to the content element, the second abnormal risk content associated with the abnormal element, and the risk level corresponding to the abnormal element. Specifically, this can involve determining the first content entity corresponding to the content element and the second content entity corresponding to the abnormal element in the scene knowledge graph; determining the first object entity associated with the first content entity, the association between the first content entity and the first object entity, and the risk level of the association between the first content entity and the first object entity based on the scene knowledge graph, and determining the risk level of the association between the first content entity and the first object entity as the risk level corresponding to the content element, and determining the association between the first content entity and the first object entity as the first abnormal risk content; and determining the second object entity associated with the second content entity, the association between the second content entity and the second object entity, and the risk level of the association between the second content entity and the second object entity based on the scene knowledge graph, and determining the risk level of the association between the second content entity and the second object entity as the risk level corresponding to the abnormal element, and determining the association between the second content entity and the second object entity as the second abnormal risk content.

[0103] Understandably, the first content entity refers to the content entity in the scene knowledge graph that is identical to the content element. By searching through each content entity in the scene knowledge graph, the first content entity corresponding to the content element can be determined. The first object entity refers to the cluster of objects in the scene knowledge graph that is associated with the first content entity. Specifically, if object cluster A and content entity B are associated, then the node corresponding to object cluster A and the node corresponding to content entity B in the scene knowledge graph are connected. Therefore, the scene knowledge graph can be used to find the nodes connected to the nodes corresponding to the first content entity, and the object cluster mapped by the nodes connected to the nodes corresponding to the first content entity is the first object entity.

[0104] Furthermore, the association between the first content entity and the first object entity can be identified as the first abnormal risk content. This association can be determined as follows: the node association information between the node corresponding to the first content entity and the node corresponding to the first object entity in the scene knowledge graph is used as the association between the first content entity and the first object entity. For example, if the first content entity is "barbecuing a certain type of meat" and the first object entity is "animal lover," the scene knowledge graph can determine that the node association information between the node corresponding to "barbecuing a certain type of meat" and the node corresponding to "animal lover" is "barbecuing a certain type of meat may cause discomfort to animal lovers." Based on this, "barbecuing a certain type of meat may cause discomfort to animal lovers" can be considered the association between the first content entity and the first object entity, which is the first abnormal risk content.

[0105] It is worth noting that the risk level of the association between the first content entity and the first object entity can also be determined as the risk level corresponding to the content element. The risk level of the association between the first content entity and the first object entity can be determined as follows: the node association attribute information between the node corresponding to the first content entity and the node corresponding to the first object entity in the scene knowledge graph is used as the risk level of the association between the first content entity and the first object entity. For example, as mentioned above, the first content entity is "barbecue a certain type of meat," and the first object entity is "animal lover." Using the scene knowledge graph, it can be determined that the node association attribute information between the node corresponding to "barbecue a certain type of meat" and the node corresponding to "animal lover" is severe. Based on this, severe association can be determined as the risk level of the association between the first content entity and the first object entity, which is also the risk level corresponding to the content element.

[0106] Understandably, the second content entity refers to the content entity in the scene knowledge graph that is identical to the abnormal element. By searching through the various content entities in the scene knowledge graph, the second content entity corresponding to the abnormal element can be identified. The second object entity refers to the cluster of objects in the scene knowledge graph that are associated with the second content entity. Specifically, the scene knowledge graph can be used to find the nodes connected to the nodes corresponding to the second content entity; the cluster of objects mapped by the nodes connected to the nodes corresponding to the second content entity is the second object entity.

[0107] Furthermore, the association between the second content entity and the second object entity can be identified as the second abnormal risk content. This association can be determined by using the node association information between the nodes corresponding to the second content entity and the nodes corresponding to the second object entity in the scene knowledge graph. For example, if the second content entity is "bloody scene" and the second object entity is "a certain group," the scene knowledge graph can determine that the node association information between the node corresponding to "bloody scene" and the node corresponding to "a certain group" is "the bloody scene may cause discomfort and fear in a certain group." Based on this, "the bloody scene may cause discomfort and fear in a certain group" can be considered the association between the second content entity and the second object entity, which is the second abnormal risk content.

[0108] It is worth noting that the risk level of the association between the second content entity and the second object entity can also be determined as the risk level corresponding to the anomalous element. This risk level can be determined by using the node association attribute information between the node corresponding to the second content entity and the node corresponding to the second object entity in the scene knowledge graph. For example, as mentioned above, if the second content entity is "bloody scene" and the second object entity is "a certain group," the node association attribute information between the node corresponding to "bloody scene" and the node corresponding to "a certain group" can be determined using the scene knowledge graph to be of a high degree. Based on this, the high degree can be determined as the risk level of the association between the second content entity and the second object entity, which is also the risk level corresponding to the anomalous element.

[0109] It should be noted that by defining the association between the first content entity and the first object entity as the first abnormal risk content, and the association between the second content entity and the second object entity as the second abnormal risk content, a more comprehensive understanding of the abnormal content within the analyzed content can be achieved by integrating both visual and textual content perspectives. Similarly, by defining the risk level of the association between the first content entity and the first object entity as the risk level corresponding to the content element, and the risk level of the association between the second content entity and the second object entity as the risk level corresponding to the abnormal element, a more comprehensive understanding of the risk level of the abnormal content can be achieved by integrating both visual and textual content perspectives.

[0110] In one possible implementation, a third-party intelligent agent is invoked to determine the analysis results of the content to be analyzed based on object tags, anomaly risk levels, and content risk descriptions. Specifically, the third-party intelligent agent analyzes the object tags, anomaly risk levels, and content risk descriptions to obtain a conflict judgment on the object tags and content risk descriptions, the anomaly type of the content to be analyzed, and the cause of the anomaly. The analysis results are then generated based on the conflict judgment on the object tags and content risk descriptions, the anomaly type of the content to be analyzed, and the cause of the anomaly.

[0111] Understandably, the third agent refers to a deep learning model used to determine the target object's acceptance level of the content to be analyzed, and to analyze why the target object cannot accept the content. The input to the third agent is object labels, anomaly risk levels, and content risk descriptions. The output is a conflict judgment between the object labels and content risk descriptions, the anomaly type of the content to be analyzed, and the reason for the anomaly. The anomaly types and reasons for the content to be analyzed have already been explained above and will not be repeated here. Given that object labels indicate the target object's inclination towards various types of content, and content risk descriptions indicate the structured description of events occurring in the content to be analyzed, as well as potentially risky content, the conflict judgment between object labels and content risk descriptions determines whether the target object has a tendency to dislike potentially risky content in the content to be analyzed.

[0112] For example, if the object label is "the object is more inclined to be 'neutral and objective, focusing on competition' when it comes to sports-related content, and has a low tolerance for controversial or biased sports content," and the content risk description is "the video shows a scene from a sports game... mentioning minor fouls and incorrect whistles, implying that Team A won the game through improper refereeing, which may cause discomfort to the target object," then the conflict judgment between the object label and the content risk description output by the third agent is: there is a conflict between the object label and the content risk description.

[0113] Understandably, the analysis result refers to the structured explanation output after a third-party agent integrates the conflict judgment, the anomaly type of the content to be analyzed, and the cause of the anomaly. For example, as mentioned above, if the object label is "the target object has feelings of fear and disgust towards cold-blooded animals," and the content risk description is "the video shows a cooking process that includes the processing of a certain type of meat, which is cut up, marinated with seasonings, skewered on bamboo sticks, and finally grilled on a stove. This processing method may cause discomfort to some people, especially those with special feelings towards certain types of animals," then the conflict judgment of the object label and the content risk description output by the third-party agent is: there is a conflict between the object label and the content risk description; the output anomaly cause is "the target object has feelings of fear and disgust towards a certain type of meat, and the processing method of the meat in the video causes discomfort to the target object"; and the output anomaly type is "conceptual conflict type." After integrating the conflict judgment, the anomaly type of the content to be analyzed, and the cause of the anomaly, the structured explanation output can be: "The video contains the processing of a certain type of meat. The meat is cut up, marinated with seasonings, skewered on bamboo sticks, and finally grilled on an oven. The target audience has feelings of fear and aversion towards this type of meat. Therefore, the way the meat is processed in the video conflicts with the target audience's perception and may cause discomfort to the target audience. In addition, the video shows the entire process of processing the meat, which may cause visual discomfort to the target audience."

[0114] In another embodiment, refer to Figure 8 If, after analyzing the object label, anomaly risk level, and content risk description through a third-party intelligent agent, the resulting conflict judgment on the object label and content risk description is that there is no conflict between the object label and the content risk description, then the analysis result is not a negative-format structured explanation generated by integrating the conflict judgment, the anomaly type of the content to be analyzed, and the anomaly cause, but rather a positive-format structured explanation generated based on the content itself of the content to be analyzed.

[0115] It should be noted that by integrating the conflict judgments of object tags and content risk descriptions output by the third-party intelligent agent, the anomaly types of the content to be analyzed, and the causes of anomalies, a structured analysis result can be output. This can transform fuzzy reasoning conclusions into structured, interpretable, and reusable decision information without human intervention, thereby improving the accuracy of content analysis and making subsequent recommendation strategies more targeted.

[0116] In one possible implementation, the third agent is trained by: acquiring negative feedback text from multiple platforms, the content text of negative feedback content from multiple platforms, and the object tags of the objects that publish the negative feedback text from multiple platforms; analyzing the negative feedback text from multiple platforms and the content text of negative feedback content from multiple platforms using a large language model to obtain the object tags of the objects that publish the negative feedback text from multiple platforms, the conflict judgment of the content text of the negative feedback content from multiple platforms, the anomaly type of the negative feedback content from multiple platforms, and the anomaly cause; and using the content text of the negative feedback content from multiple platforms, the object tags of the objects that publish the negative feedback text from multiple platforms, the conflict judgment, the anomaly type of the negative feedback content from multiple platforms, and the anomaly cause of the negative feedback content as training data to train the third agent.

[0117] Multi-platform negative feedback text refers to negative reviews of various types of content on a platform (which can be a video platform or a webpage). Multi-platform negative feedback content refers to the content targeted by the multi-platform negative feedback text (which can be video content, text, or a webpage), i.e., the content targeted by the negative reviews on the platform. The content text of multi-platform negative feedback content refers to the text used to describe the content of multi-platform negative feedback content. The object tags of the object that published the multi-platform negative feedback text refer to the object's tendency towards various types of content. Generally, the multi-platform negative feedback text, the content text of the multi-platform negative feedback content, and the object tags of the object that published the multi-platform negative feedback text can be obtained by pulling the published content, the corresponding comment section of the published content, and the browsing history of the object using the platform.

[0118] It is understandable that, based on the multi-platform negative feedback text, it has been determined that the recipient of the multi-platform negative feedback text is averse to the multi-platform negative feedback content. In other words, there must be a conflict between the object tags of the recipient who posted the multi-platform negative feedback text and the content text of the multi-platform negative feedback content. Based on this, as... Figure 9 As shown in the upper half of the figure, the negative feedback text from multiple platforms, the content text of the negative feedback content from multiple platforms, and the object labels of the objects that publish the negative feedback text from multiple platforms can be input into the large language model. This allows the large language model to analyze the reasons for the conflict between the object labels of the objects that publish the negative feedback text from multiple platforms and the content text of the negative feedback content from multiple platforms (i.e., the reasons for the anomalies), as well as the negative category system to which the controversy risk of the negative feedback content from multiple platforms belongs (i.e., the anomaly type).

[0119] Furthermore, the content text of multi-platform negative feedback content, the object tags of the objects that publish multi-platform negative feedback text, conflict judgments, anomaly types of multi-platform negative feedback content, and anomaly reasons can be used as training data. This allows the third-party agent to learn the mapping relationship between the content text of multi-platform negative feedback content, the object tags of the objects that publish multi-platform negative feedback text, and the conflict judgments, anomaly types, and anomaly reasons used as training data. Thus, when faced with new content text and object tags, it can accurately predict conflict judgments, anomaly types, and anomaly reasons. The base model of the third-party agent can be a large language model (e.g., the Qwen2.5-VL model). This setup enables the third-party agent to possess general language capabilities, facilitating its understanding of the textual logic within the content text of multi-platform negative feedback content, the object tags of the objects that publish multi-platform negative feedback text, conflict judgments, anomaly types, and anomaly reasons.

[0120] It should be noted that by using the content text of multi-platform negative feedback content, the object tags of the objects that publish multi-platform negative feedback text, conflict judgments, anomaly types of multi-platform negative feedback content, and anomaly causes as training data to train the third-party agent, the third-party agent can fully learn the mapping relationship between the content text of multi-platform negative feedback content, the object tags of the objects that publish multi-platform negative feedback text, and the conflict judgments, anomaly types of multi-platform negative feedback content, and anomaly causes used as training data. This makes it easier to accurately predict conflict judgments, anomaly types, and anomaly causes when faced with the content text of the content to be analyzed and the object tags of the target objects.

[0121] In one possible implementation, the third agent is trained by splitting the training data into fine-tuning training data and reinforcement training data; the third agent is then fine-tuned under supervision based on the fine-tuning training data; and the third agent is then fine-tuned under reinforcement based on the reinforcement training data.

[0122] Among them, supervised fine-tuning and reinforcement fine-tuning are two progressive training phases. The core goal is to upgrade the base model, such as the Qwen2.5-VL model, from being able to understand multimodal content to being able to accurately predict conflict judgments, anomaly types, and anomaly causes.

[0123] Supervised fine-tuning, as we understand it, refers to a supervised learning process that optimizes the model parameters of a third agent by using labeled training data (i.e., the content text of negative feedback content across multiple platforms and the object labels of the objects that published the negative feedback text) – output (i.e., conflict judgments, anomaly types of negative feedback content across multiple platforms, and anomaly causes) – on top of the base model of the third agent. This fine-tuning optimizes the model parameters by minimizing prediction errors. In this context, the fine-tuned training data can also be understood as training data used to train the third agent to learn the correct logic for predicting conflict judgments, anomaly types, and anomaly causes.

[0124] Furthermore, while the base model of the third agent can understand the content text of negative feedback content across multiple platforms, as well as the object tags of the objects that published the negative feedback text, it does not actively output conflict judgments, anomaly types of negative feedback content across multiple platforms, or the reasons for the anomalies. By using fine-tuning training data to supervise the fine-tuning of the third agent, the base model of the third agent can develop preliminary reasoning capabilities, learning to infer the reasons for conflicts between content text and object tags, as well as the types of conflicts, based on the content text and object tags.

[0125] It's important to clarify that reinforcement fine-tuning refers to binding the model's output to a reward function (such as the GRPO algorithm) based on the supervised fine-tuning of the third agent. Reinforcement learning algorithms are then used to further optimize the model parameters of the third agent, making its output increasingly closer to the desired outcome. Reinforcement training data refers to training data used to ensure the output of the supervised fine-tuned third agent is more reliable and accurate. By using reinforcement training data to fine-tune the supervised fine-tuned third agent, not only can the inference results be made more accurate and stable, but it also ensures that the third agent's output does not deviate from the correct knowledge learned during the supervised fine-tuning stage.

[0126] In one possible implementation, the supervised fine-tuning third agent is reinforced based on the reinforcement training data. Specifically, the content text and object labels from the reinforcement training data are input into the third agent to obtain its output. When the conflict determination indicates a conflict between the object label and content text of an object that publishes negative feedback text on multiple platforms, the model parameters of the third agent are adjusted based on the first reward weight. When the conflict determination indicates a conflict between the object label and content text of an object that publishes negative feedback text on multiple platforms, and the anomaly type of the negative feedback content is consistent with the anomaly type in the reinforcement training data, the model parameters of the third agent are adjusted based on the second reward weight. When the conflict determination indicates a conflict between the object label and content text of an object that publishes negative feedback text on multiple platforms, and the anomaly type of the negative feedback content is consistent with the anomaly type in the reinforcement training data, and the anomaly cause of the negative feedback content is consistent with the anomaly cause in the reinforcement training data, the model parameters of the third agent are adjusted based on the third reward weight. The process returns to the step of inputting the content text and object labels from the reinforcement training data into the third agent until the third agent converges.

[0127] In this method, the content text and object labels from the reinforcement training data are used as input data. After being input into a third-party intelligent agent, the output of the third-party intelligent agent includes the conflict judgment of object labels and content text of objects that publish negative feedback text on multiple platforms, the anomaly type of the negative feedback content on multiple platforms, and the reason for the anomaly. Moreover, as... Figure 9 As shown, after each training iteration, the third agent generates an output result. The output result of the second iteration is more accurate than the output result of the first iteration, and the output result of the third iteration is more accurate than the output result of the second iteration.

[0128] Specifically, in each round of training, the content text of the multi-platform negative feedback content from the new reinforcement training data, along with the object labels of the objects that published the multi-platform negative feedback text, are input to the third agent. This causes the third agent to output a conflict judgment regarding the object labels and content text of the objects that published the multi-platform negative feedback text, the anomaly type of the multi-platform negative feedback content, and the reason for the anomaly. After this, as... Figure 9 As shown, this embodiment introduces a reward function. The basic logic of the reward function is to give rewards in order from easy to difficult (i.e., to adjust the model parameters of the third agent based on the reward weight).

[0129] The first step is a binary judgment reward. When the conflict judgment in the third agent's output indicates a conflict between the object label and content text of the object that published negative feedback text on multiple platforms, it means that the third agent has learned the core ability to distinguish between positive and negative feedback, and can proceed to the next stage. If the conflict judgment in the third agent's output indicates that there is no conflict between the object label and content text of the object that published negative feedback text on multiple platforms, the training round is terminated directly, and no reward is given.

[0130] The second step is type classification reward. When the anomaly type of the multi-platform negative feedback content in the third agent's output matches the anomaly type in the reinforcement training data, it indicates that the third agent has learned the ability to identify anomaly types, and it can proceed to the next stage. If the anomaly type of the multi-platform negative feedback content in the third agent's output does not match the anomaly type in the reinforcement training data, the current round of training is terminated, and the first reward weight is assigned (i.e., the model parameters of the third agent are adjusted based on the first reward weight).

[0131] The third step is the explanation reward. When the cause of the anomaly in the multi-platform negative feedback content in the third agent's output matches the cause of the anomaly in the reinforcement training data, it indicates that the third agent has learned to explain the conflict between the object label and content text of the object that published the multi-platform negative feedback text using natural language. At this point, a third reward weight can be assigned. If the cause of the anomaly in the multi-platform negative feedback content in the third agent's output does not match the cause of the anomaly in the reinforcement training data, a second reward weight is assigned. The second reward weight is greater than the first reward weight, and the third reward weight is greater than the second reward weight.

[0132] Understandably, after calculating the reward weights or terminating the current training round, the process returns to inputting the content text and object labels from the reinforcement training data into the third agent, initiating a new round of iterative training until the model converges. In other words, the third agent repeats the steps of input-output-reward weight calculation-model parameter update until it converges, or until the reward weights no longer increase, indicating that the third agent has the ability to accurately predict conflicting judgments, anomaly types, and causes. Through this process, the third agent can gradually optimize the accuracy of its inference by generating correct output results and obtaining reward weights in a loop.

[0133] It's important to clarify that adjusting the model parameters of the third agent based on reward weights specifically means that if an output receives a high reward (e.g., a third reward weight), the third agent will strengthen the reasoning path corresponding to that output; conversely, if an output receives a low reward (e.g., a first reward weight) or no reward, the third agent will weaken the reasoning path corresponding to that output. In essence, "strengthening the reasoning path corresponding to this output" can be understood as making the third agent more inclined to use the successful reasoning logic to generate outputs when encountering similar object tags and content text in the future. The core of this approach is to amplify the weight of correct reasoning paths by adjusting the model parameters of the third agent.

[0134] In one possible implementation, after inputting the content text and object labels from the reinforcement training data into a third agent to obtain the output of the third agent, the method further includes: inputting the content text and object labels from the reinforcement training data into a reference model to obtain the output of the reference model; determining deviation penalty weights based on the difference between the output of the reference model and the output of the third agent; and adjusting the model parameters of the third agent based on the deviation penalty weights.

[0135] The reference model is the third agent after supervised fine-tuning, and the model parameters of the reference model are fixed. During the iterative training of the third agent, the reference model can be used as a benchmark model to ensure that the third agent does not deviate from the correct logic during reinforcement fine-tuning.

[0136] Understandably, the output of a third-party agent is generally not a simple structured conclusion, but rather a probability distribution corresponding to each reasoning outcome. These probability distributions represent the probabilistic tendency of the output result. For example, for conflict judgments, the explanation model will output the probability distributions for conflict and non-conflict. For anomaly type judgments, the explanation model will output the probability distributions corresponding to various anomaly types. For anomaly cause generation, the explanation model will output the probability distribution of each word in its corresponding position. Based on this, the third-party agent might randomly select anomaly types or anomaly causes with high probability distributions from past outputs in order to obtain reward weights. This would lead to a serious deviation in the third-party agent's reasoning logic.

[0137] For the reasons mentioned above, the content text and object labels in the reinforcement training data can be used as input data to the reference model to obtain its output. Since the reference model serves as a benchmark, its parameters are fixed, and its output conforms to correct logic. Therefore, the probability distribution in the reference model's output can be compared with the probability distribution in the third agent's output to calculate the KL divergence, which measures the difference between the reference model's and the third agent's output. A larger KL divergence indicates a greater difference between the reference model's and the third agent's output, and a further deviation of the third agent from correct reasoning logic. Conversely, a smaller KL divergence indicates a smaller difference between the reference model's and the third agent's output, and a closer alignment of the third agent's reasoning logic with the reasoning logic learned through supervised fine-tuning.

[0138] It should be noted that the KL divergence can be used as a deviation penalty weight, and the model parameters of the third agent can be adjusted based on this weight. A larger KL divergence results in a larger deviation penalty weight, and a smaller KL divergence results in a smaller deviation penalty weight. This approach ensures that the third agent's reasoning logic does not deviate from the reasoning logic learned through supervised fine-tuning.

[0139] The complete process of content analysis methodology is explained in detail below.

[0140] Reference Figure 10 , Figure 10 This is a schematic diagram of an optional architecture for the content analysis method provided in the embodiments of this disclosure.

[0141] Step 1001: Call the first intelligent agent to obtain object information and historical operation information.

[0142] Step 1002: Filter out the first target content.

[0143] Step 1003: Semantic clustering of content titles. Specifically, this may include the following steps: for each first target content, perform semantic vector transformation on the content title of the first target content to obtain multiple title semantic vectors; perform vector clustering on the title semantic vectors using a clustering algorithm to obtain multiple negative title clusters, and the semantics of multiple title semantic vectors in each negative title cluster are similar.

[0144] Step 1004: Semantic clustering of content fragment features. Specifically, this may include the following steps: For each first target content, uniformly sample the first target content to obtain multiple content fragments corresponding to the first target content; extract content fragment features from the content fragments using a visual encoder; perform vector clustering on the content fragment features corresponding to each first target content using a clustering algorithm to obtain multiple negative visual feature clusters, and the semantics of the multiple content fragment features in each negative visual feature cluster are similar.

[0145] Step 1005: Invoke the first intelligent agent to determine object tags. Specifically, this may include the following steps: using the first intelligent agent to analyze negative title clusters, negative visual feature clusters, the target object's feedback to the first target content, and object information to obtain object tags when the target object browses the content. After obtaining the object tags when the target object browses the content, the steps further include: when historical operation information is updated, filtering out second target content with a viewing rate lower than a first viewing threshold from the newly added historical operation information; performing semantic clustering on the content titles of the second target content to obtain multiple new negative title clusters; performing semantic clustering on the content fragment features of the second target content to obtain multiple new negative visual feature clusters; and updating the object tags based on the new negative title clusters and new negative visual feature clusters.

[0146] Step 1006: Invoke the second agent to obtain multiple content fragments and the corresponding text of the content to be analyzed. Specifically, this may include the following steps: obtaining the sampling time interval; sampling the content to be analyzed based on the sampling time interval to obtain multiple content fragments corresponding to the content to be analyzed.

[0147] Step 1007: Extract content elements and anomalous elements. Specifically, this may include the following steps: for each content segment, extract multiple content segment features; perform anomalous element identification processing on the content segment features to obtain anomalous elements.

[0148] Step 1008: Invoke the second intelligent agent to use the scene knowledge graph to determine the first and second abnormal risk contents. Specifically, this may include the following steps: using the scene knowledge graph, determine the first abnormal risk contents associated with the content element, the risk level corresponding to the content element, the second abnormal risk contents associated with the abnormal element, and the risk level corresponding to the abnormal element, respectively. The scene knowledge graph is a knowledge graph used to describe the association between clusters of abnormal risk contents and clusters of objects.

[0149] Furthermore, using a scene knowledge graph, the following steps are taken to determine the first abnormal risk content associated with the content element, the risk level corresponding to the content element, the second abnormal risk content associated with the abnormal element, and the risk level corresponding to the abnormal element: First, determine the first content entity corresponding to the content element and the second content entity corresponding to the abnormal element in the scene knowledge graph; second, based on the scene knowledge graph, determine the first object entity associated with the first content entity, the association between the first content entity and the first object entity, and the risk level of the association between the first content entity and the first object entity, and determine the risk level of the association between the first content entity and the first object entity as the risk level corresponding to the content element, and determine the association between the first content entity and the first object entity as the first abnormal risk content; third, based on the scene knowledge graph, determine the second object entity associated with the second content entity, the association between the second content entity and the second object entity, and the risk level of the association between the second content entity and the second object entity, and determine the risk level of the association between the second content entity and the second object entity as the risk level corresponding to the abnormal element, and determine the association between the second content entity and the second object entity as the second abnormal risk content.

[0150] The scenario knowledge graph is constructed as follows: negative feedback text from multiple platforms is obtained, consisting of negative feedback content published by multiple objects on various platforms; content entities and object entities are extracted from the negative feedback text; the associations between object entities and content entities are analyzed based on the negative feedback text, and the risk level of these associations is determined; content entities and object entities are identified as nodes, the associations between them are identified as node association information, and the risk level of these associations is identified as node association attribute information; and a scenario knowledge graph is constructed based on each node, node association information, and node association attribute information.

[0151] Step 1009: Determine the anomaly risk level and content risk description. This may specifically include the following steps: determining the anomaly risk level of the content to be analyzed based on the risk levels corresponding to the content elements and the anomaly elements; and determining the content risk description based on the first and second anomaly risk contents.

[0152] Step 1010: Analyze the object tags, anomaly risk levels, and content risk descriptions through a third-party intelligent agent.

[0153] Step 1011: Generate analysis results for the content to be analyzed. Specifically, this may include the following steps: Analyzing object tags, anomaly risk levels, and content risk descriptions using a third-party intelligent agent to obtain conflict judgments regarding object tags and content risk descriptions, anomaly types of the content to be analyzed, and anomaly causes; generating analysis results based on the conflict judgments regarding object tags and content risk descriptions, the anomaly types of the content to be analyzed, and the anomaly causes.

[0154] It is worth noting that the third agent is trained in the following way: acquiring negative feedback text from multiple platforms, the content text of negative feedback content from multiple platforms, and the object tags of the objects that publish negative feedback text from multiple platforms, where the negative feedback content is the content targeted by the negative feedback text from multiple platforms; using a large language model to analyze the negative feedback text from multiple platforms and the content text of negative feedback content from multiple platforms, obtaining the object tags of the objects that publish negative feedback text from multiple platforms, the conflict judgment of the content text of negative feedback content from multiple platforms, the anomaly type of negative feedback content from multiple platforms, and the anomaly cause; using the content text of negative feedback content from multiple platforms, the object tags of the objects that publish negative feedback text from multiple platforms, the conflict judgment, the anomaly type of negative feedback content from multiple platforms, and the anomaly cause of anomaly as training data to train the third agent.

[0155] Based on this, by invoking the first intelligent agent to obtain the target object's object information and historical operation information, and determining the object tags when the target object browses content based on the object information and historical operation information, it is possible to clearly identify the target object's operation information when viewing content it likes and when viewing content it dislikes. Then, by invoking the second intelligent agent to obtain multiple content fragments and corresponding text of the content to be analyzed, and determining the anomaly risk level and content risk description of the content to be analyzed based on these multiple content fragments and text, a more comprehensive understanding of the anomaly content and its risk level can be achieved by integrating both visual and textual content levels. This process effectively improves the accuracy of content analysis. Furthermore, because different objects have different tolerance levels for anomalies in the content to be analyzed—some objects may accept anomalies while others may not—a third intelligent agent can be invoked to combine the previously obtained object tags from the target object's browsing behavior to analyze the anomaly risk level and content risk description, thereby determining the target object's acceptance level of the content to be analyzed and obtaining the final analysis result. If the target audience cannot accept the content to be analyzed, the analysis results will detail the anomaly type and cause of the anomaly, providing a comprehensive analysis of why the target audience cannot accept the content. Compared to existing schemes that rely on manual identification of the specific content triggering negative feedback from the target audience, this embodiment, through joint training of a first, second, and third agent, can not only analyze the content to be analyzed from multiple modal levels without human intervention, improving the accuracy of content analysis, but also determine the anomaly type and cause of the content to be analyzed by combining the object tags used by the target audience when browsing the content, making the analysis results more aligned with the target audience's browsing preferences. Furthermore, by introducing a multi-agent processing approach, with different agents responsible for different analysis processes, more refined analysis can be achieved through mutual cooperation, thereby improving the accuracy of content analysis.

[0156] It is understood that although the steps in the above flowcharts are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated in this embodiment, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the above flowcharts may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0157] Reference Figure 11 , Figure 11 This is a schematic diagram of an optional structure of the content analysis device 1100 provided in an embodiment of the present disclosure. The content analysis device 1100 includes: The first acquisition module 1101 is used to call the first intelligent agent to acquire the object information and historical operation information of the target object, and determine the object tag when the target object browses the content based on the object information and historical operation information. The historical operation information is the operation information when the target object browses the content. The second acquisition module 1102 is used to call the second intelligent agent to acquire multiple content fragments and content text corresponding to the content to be analyzed, and to determine the abnormal risk level and content risk description of the content to be analyzed based on the multiple content fragments and content text. The content text is the text used to describe the content to be analyzed. The determination module 1103 is used to invoke a third agent to determine the analysis results of the content to be analyzed based on object labels, anomaly risk levels, and content risk descriptions. The analysis results indicate the anomaly type and cause of the content to be analyzed. The first, second, and third agents are obtained through joint training.

[0158] Furthermore, historical operational information includes the target audience's view rate for each piece of content, as well as the target audience's feedback on the content; The first acquisition module 1101 mentioned above is specifically used for: Filter out the primary target content whose viewership is below the first viewership threshold; Semantic clustering of the titles of the primary target content yields multiple negative title clusters; Semantic clustering is performed on the content fragment features of the first target content to obtain multiple negative visual feature clusters; The first intelligent agent is used to analyze negative title clusters, negative visual feature clusters, feedback content of the target object to the first target content, and object information to obtain object tags when the target object browses the content.

[0159] Furthermore, the aforementioned first acquisition module 1101 is specifically used for: For each primary target content, the title of the primary target content is transformed into a semantic vector, resulting in multiple title semantic vectors; By using a clustering algorithm to perform vector clustering on the title semantic vectors, multiple negative title clusters are obtained, and the semantics of multiple title semantic vectors in each negative title cluster are similar.

[0160] Furthermore, the aforementioned first acquisition module 1101 is specifically used for: For each first target content, the first target content is sampled uniformly to obtain multiple content fragments corresponding to the first target content; Extract content fragment features from content fragments using a visual encoder; By using a clustering algorithm to perform vector clustering on the content fragment features corresponding to each first target content, multiple negative visual feature clusters are obtained, and the semantics of multiple content fragment features in each negative visual feature cluster are similar.

[0161] Furthermore, the aforementioned first acquisition module 1101 is specifically used for: When historical operation information is updated, select the second target content with a viewing rate lower than the first viewing threshold from the newly added historical operation information; Semantic clustering of the titles of the second target content yields several new negative title clusters; Semantic clustering of content fragment features of the second target content yields multiple new negative visual feature clusters; The object labels are updated based on the newly added negative title cluster and the newly added negative visual feature cluster.

[0162] Furthermore, the second acquisition module 1102 mentioned above is specifically used for: Extract content elements from the text and identify anomalous elements in the content fragments. Content elements indicate content with actual semantic meaning, while anomalous elements indicate visual content that poses a risk. Using a scenario knowledge graph, the first abnormal risk content associated with the content element, the risk level corresponding to the content element, the second abnormal risk content associated with the abnormal element, and the risk level corresponding to the abnormal element are determined respectively. The scenario knowledge graph is a knowledge graph used to describe the association between the cluster of abnormal risk content and the cluster of objects. The abnormal risk level of the content to be analyzed is determined based on the risk level corresponding to the content element and the risk level corresponding to the abnormal element, and the content risk description is determined based on the first abnormal risk content and the second abnormal risk content.

[0163] Furthermore, the second acquisition module 1102 mentioned above is specifically used for: Retrieve negative feedback text from multiple platforms. The negative feedback text from multiple platforms consists of negative feedback content published by multiple individuals on the platforms. Extract content entities and object entities from negative feedback text across multiple platforms; Based on multi-platform negative feedback text analysis, the association between object entities and content entities is determined, and the risk level of the association between object entities and content entities is determined. Content entities and object entities are identified as nodes, the association between object entities and content entities is identified as node association information, the risk level of the association between object entities and content entities is identified as node association attribute information, and a scenario knowledge graph is constructed based on each node, node association information, and node association attribute information.

[0164] Furthermore, the second acquisition module 1102 mentioned above is specifically used for: Identify the first content entity corresponding to the content element and the second content entity corresponding to the abnormal element in the scene knowledge graph; Based on the scene knowledge graph, the risk levels of the first object entity associated with the first content entity, the association between the first content entity and the first object entity, and the association between the first content entity and the first object entity are determined. The risk level of the association between the first content entity and the first object entity is determined as the risk level corresponding to the content element. The association between the first content entity and the first object entity is determined as the first abnormal risk content. Based on the scene knowledge graph, the risk levels of the second object entity associated with the second content entity, the association between the second content entity and the second object entity, and the association between the second content entity and the second object entity are determined. The risk level of the association between the second content entity and the second object entity is determined as the risk level corresponding to the abnormal element, and the association between the second content entity and the second object entity is determined as the second abnormal risk content.

[0165] Furthermore, the second acquisition module 1102 mentioned above is specifically used for: For each content segment, extract multiple content segment features from the content segment; Abnormal elements are identified by performing abnormal element identification processing on the features of the content fragments.

[0166] Furthermore, the aforementioned determining module 1103 is specifically used for: By analyzing object tags, anomaly risk levels, and content risk descriptions using a third-party intelligent agent, we can obtain conflict judgments on object tags and content risk descriptions, anomaly types of the content to be analyzed, and anomaly causes. Analysis results are generated based on the conflict judgment of object tags and content risk descriptions, the anomaly type of the content to be analyzed, and the cause of the anomaly.

[0167] Furthermore, the aforementioned determining module 1103 is specifically used for: Get the negative feedback text from multiple platforms, the content text of the negative feedback from multiple platforms, and the object tags of the object that published the negative feedback text from multiple platforms. The negative feedback content from multiple platforms is the content that the negative feedback text from multiple platforms targets. Using a large language model, we analyze negative feedback text and content text from multiple platforms to obtain conflict judgments of object tags for objects that publish negative feedback text and content text of negative feedback content from multiple platforms, as well as the abnormal types and causes of abnormalities in negative feedback content from multiple platforms. The training data includes the text of negative feedback content from multiple platforms, the object tags of the objects that publish negative feedback text from multiple platforms, conflict judgments, the anomaly types of negative feedback content from multiple platforms, and the reasons for the anomalies, which are used to train the third agent.

[0168] Furthermore, the aforementioned determining module 1103 is specifically used for: The training data is split into fine-tuning training data and reinforcement training data; Supervised fine-tuning of the third agent is performed based on the fine-tuning training data. Based on the reinforcement training data, the third agent after supervision fine-tuning is further reinforced and fine-tuned.

[0169] Furthermore, the aforementioned determining module 1103 is specifically used for: The content text and object labels in the reinforcement training data are used as input data and fed into the third agent to obtain the output results of the third agent. The output results include the conflict judgment of object labels and content text of objects that publish negative feedback text on multiple platforms, the anomaly type of negative feedback content on multiple platforms, and the anomaly reason. When the conflict determination indicates that there is a conflict between the object tag and the content text of the object that publishes negative feedback text on multiple platforms, the model parameters of the third agent are adjusted based on the first reward weight. When the conflict judgment indicates that the object label and content text of the object that publishes negative feedback text on multiple platforms conflict, and the anomaly type of the negative feedback content on multiple platforms is consistent with the anomaly type in the reinforcement training data, the model parameters of the third agent are adjusted based on the second reward weight, and the second reward weight is greater than the first reward weight. When the conflict judgment indicates that the object tag and content text of the object that publishes negative feedback text on multiple platforms conflict, and the anomaly type of the negative feedback content on multiple platforms is consistent with the anomaly type in the reinforcement training data, and the anomaly cause of the negative feedback content on multiple platforms is consistent with the anomaly cause in the reinforcement training data, the model parameters of the third agent are adjusted based on the third reward weight, and the third reward weight is greater than the second reward weight. Return to the steps where the content text and object labels from the reinforcement training data are used as input data and fed into the third agent until the third agent converges.

[0170] Furthermore, the aforementioned determining module 1103 is specifically used for: The content text and object labels in the reinforcement training data are used as input data and fed into the reference model to obtain the output of the reference model. The reference model is a third agent after supervised fine-tuning. Based on the difference between the output of the reference model and the output of the third agent, the deviation penalty weight is determined. The model parameters of the third agent are adjusted based on the deviation penalty weights.

[0171] Furthermore, the second acquisition module 1102 mentioned above is specifically used for: Obtain the sampling time interval; The content to be analyzed is sampled based on the sampling time interval to obtain multiple content fragments corresponding to the content to be analyzed.

[0172] The aforementioned content analysis device 1100 and content analysis method are based on the same inventive concept. By invoking a first intelligent agent to obtain object information and historical operation information of the target object, and determining the object tags when the target object browses content based on the object information and historical operation information, it can clearly identify the operation information of the target object when viewing content it likes and when viewing content it dislikes. Then, by invoking a second intelligent agent to obtain multiple content fragments and corresponding content text of the content to be analyzed, and determining the abnormal risk level and content risk description of the content to be analyzed based on the multiple content fragments and content text, it can comprehensively understand the abnormal content and the degree of risk of the abnormal content in the content to be analyzed by integrating visual and textual content levels. This process effectively improves the accuracy of content analysis. Furthermore, because different objects have different tolerance levels for abnormal content in the content to be analyzed, some objects may be able to accept abnormal content in the content to be analyzed, while others may not. Based on this, a third intelligent agent can be invoked to combine the previously obtained object tags when the target object browses content to analyze the abnormal risk level and content risk description, in order to determine the target object's acceptance level of the content to be analyzed, and obtain the final analysis result. If the target audience cannot accept the content to be analyzed, the analysis results will detail the anomaly type and cause of the anomaly, providing a comprehensive analysis of why the target audience cannot accept the content. Compared to existing schemes that rely on manual identification of the specific content triggering negative feedback from the target audience, this embodiment, through joint training of a first, second, and third agent, can not only analyze the content to be analyzed from multiple modal levels without human intervention, improving the accuracy of content analysis, but also determine the anomaly type and cause of the content to be analyzed by combining the object tags used by the target audience when browsing the content, making the analysis results more aligned with the target audience's browsing preferences. Furthermore, by introducing a multi-agent processing approach, with different agents responsible for different analysis processes, more refined analysis can be achieved through mutual cooperation, thereby improving the accuracy of content analysis.

[0173] The electronic device provided in this disclosure for performing the above-described content analysis method can be a terminal, as shown in the following embodiments. Figure 12 , Figure 12This is a partial structural block diagram of a terminal provided in an embodiment of the present disclosure. The terminal includes: a camera assembly 1210, a first memory 1220, an input unit 1230, a display unit 1240, a sensor 1250, an audio circuit 1260, a wireless fidelity (WiFi) module 1270, a first processor 1280, and a first power supply 1290, among other components. Those skilled in the art will understand that... Figure 12 The terminal structure shown does not constitute a limitation on the terminal and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0174] The camera assembly 1210 can be used to capture images or videos. Optionally, the camera assembly 1210 includes a front-facing camera and a rear-facing camera. Typically, the front-facing camera is located on the front panel of the terminal, and the rear-facing camera is located on the back of the terminal. In some embodiments, there are at least two rear-facing cameras, which are any one of a main camera, a depth-sensing camera, a wide-angle camera, and a telephoto camera, to achieve background blurring by fusion of the main camera and the depth-sensing camera, panoramic shooting by fusion of the main camera and the wide-angle camera, VR (Virtual Reality) shooting, or other fusion shooting functions.

[0175] The first memory 1220 can be used to store software programs and modules. The first processor 1280 executes various functional applications and data processing of the terminal by running the software programs and modules stored in the first memory 1220.

[0176] The input unit 1230 can be used to receive input numeric or character information, and to generate key signal inputs related to the terminal's settings and function control. Specifically, the input unit 1230 may include a touch panel 1231 and other input devices 1232.

[0177] The display unit 1240 can be used to display input or provided information, as well as various menus of the terminal. The display unit 1240 may include a display panel 1241.

[0178] Audio circuitry 1260, speaker 1261, and microphone 1262 provide an audio interface.

[0179] The first power source 1290 can be AC ​​power, DC power, a disposable battery, or a rechargeable battery.

[0180] The number of sensors 1250 can be one or more, and these sensors 1250 include, but are not limited to: accelerometers, gyroscopes, pressure sensors, optical sensors, etc. An accelerometer can detect the magnitude of acceleration along the three coordinate axes of a coordinate system established by the terminal. For example, an accelerometer can be used to detect the components of gravitational acceleration along the three coordinate axes. The first processor 1280 can control the display unit 1240 to display the user interface in either a horizontal or vertical view based on the gravitational acceleration signal acquired by the accelerometer. The accelerometer can also be used for collecting motion data from games or other applications.

[0181] The gyroscope sensor can detect the terminal's orientation and rotation angle. It can work in conjunction with an accelerometer to collect 3D user movements on the terminal. Based on the data collected by the gyroscope sensor, the first processor 1280 can perform the following functions: motion sensing (e.g., changing the UI based on the user's tilt), image stabilization during shooting, game control, and inertial navigation.

[0182] The pressure sensor can be installed on the side bezel of the terminal and / or on the lower layer of the display unit 1240. When the pressure sensor is installed on the side bezel of the terminal, it can detect the user's grip signal on the terminal, and the first processor 1280 can perform left / right hand recognition or quick operation based on the grip signal collected by the pressure sensor. When the pressure sensor is installed on the lower layer of the display unit 1240, the first processor 1280 can control the operable controls on the UI interface based on the user's pressure operation on the display unit 1240. The operable controls include at least one of button controls, scroll bar controls, icon controls, and menu controls.

[0183] An optical sensor is used to collect ambient light intensity. In one embodiment, the first processor 1280 can control the display brightness of the display unit 1240 based on the ambient light intensity collected by the optical sensor. Specifically, when the ambient light intensity is high, the display brightness of the display unit 1240 is increased; when the ambient light intensity is low, the display brightness of the display unit 1240 is decreased. In another embodiment, the first processor 1280 can also dynamically adjust the shooting parameters of the camera assembly 1210 based on the ambient light intensity collected by the optical sensor.

[0184] In this embodiment, the first processor 1280 included in the terminal can execute the content analysis method of the previous embodiment.

[0185] The electronic device provided in this disclosure for performing the above-described content analysis method can also be a server, see reference. Figure 13 , Figure 13This is a partial structural block diagram of a server provided in an embodiment of the present disclosure. The server can vary significantly due to different configurations or performance. It may include one or more second processors 1310 and second memories 1330, and one or more storage media 1340 (e.g., one or more mass storage devices) for storing application programs 1343 or data 1342. The second memories 1330 and storage media 1340 may be temporary or persistent storage. The program stored in the storage media 1340 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the server. Furthermore, the second processor 1310 may be configured to communicate with the storage media 1340 and execute the series of instruction operations stored in the storage media 1340 on the server.

[0186] The server may also include one or more secondary power supplies 1320, one or more wired or wireless network interfaces 1350, one or more input / output interfaces 1360, and / or one or more operating systems 1341, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0187] The second processor 1310 in the server can be used to execute content analysis methods.

[0188] This disclosure also provides a computer-readable storage medium for storing a computer program that executes the content analysis methods of the foregoing embodiments.

[0189] This disclosure also provides a computer program product comprising a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the content analysis method described above.

[0190] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in this disclosure and the foregoing drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate to describe embodiments of this disclosure, for example, those that can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatuses.

[0191] It should be understood that in this disclosure, "at least one item" means one or more, and "more than one" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0192] It should be understood that in the description of the embodiments of this disclosure, "multiple" means two or more, "greater than", "less than", "exceeding" etc. are understood to exclude the number itself, and "above", "below", "within" etc. are understood to include the number itself.

[0193] In the several embodiments provided in this disclosure, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0194] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0195] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0196] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, 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. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0197] It should also be understood that the various implementation methods provided in this disclosure can be combined arbitrarily to achieve different technical effects. The above is a detailed description of the preferred embodiments of this disclosure. However, this disclosure is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this disclosure. All such equivalent modifications or substitutions are included within the scope defined by the claims of this disclosure.

Claims

1. A content analysis method, characterized in that, include: The first intelligent agent is invoked to obtain the object information and historical operation information of the target object. Based on the object information and the historical operation information, the object tag of the target object when browsing content is determined. The historical operation information is the operation information of the target object when browsing content. The second intelligent agent is invoked to obtain multiple content fragments corresponding to the content to be analyzed and the content text corresponding to the content to be analyzed. Based on the multiple content fragments and the content text, the abnormal risk level and content risk description of the content to be analyzed are determined. The content text is the text used to describe the content of the content to be analyzed. A third agent is invoked to determine the analysis result of the content to be analyzed based on the object label, the anomaly risk level, and the content risk description. The analysis result is used to indicate the anomaly type and cause of the content to be analyzed. The first agent, the second agent, and the third agent are obtained through joint training.

2. The content analysis method according to claim 1, characterized in that, The historical operation information includes the target object's view rate for each piece of content, and the target object's feedback content for the content; The step of determining the object tag when browsing content of the target object based on the object information and the historical operation information includes: Filter out the first target content whose view rate is lower than the first view threshold; Semantic clustering is performed on the content titles of the first target content to obtain multiple negative title clusters; Semantic clustering is performed on the content fragment features of the first target content to obtain multiple negative visual feature clusters; The first intelligent agent is used to analyze the negative title cluster, the negative visual feature cluster, the feedback content of the target object to the first target content, and the object information to obtain the object tags when the target object browses the content.

3. The content analysis method according to claim 2, characterized in that, The semantic clustering of the content titles of the first target content yields multiple negative title clusters, including: For each of the first target contents, the semantic vector transformation is performed on the content title of the first target content to obtain multiple title semantic vectors; The title semantic vectors are clustered using a clustering algorithm to obtain multiple negative title clusters, and the semantics of multiple title semantic vectors in each negative title cluster are similar.

4. The content analysis method according to claim 2, characterized in that, The semantic clustering of the content fragment features of the first target content yields multiple negative visual feature clusters, including: For each of the first target contents, the first target contents are uniformly sampled to obtain multiple content fragments corresponding to the first target contents; The visual encoder is used to extract content fragment features from the content fragment; By performing vector clustering on the content fragment features corresponding to each of the first target contents using a clustering algorithm, multiple negative visual feature clusters are obtained, and the semantics of the multiple content fragment features in each negative visual feature cluster are similar.

5. The content analysis method according to claim 2, characterized in that, After obtaining the object tags when browsing the content of the target object, the method further includes: When the historical operation information is updated, second target content with a viewing rate lower than the first viewing threshold is filtered out from the newly added historical operation information; Semantic clustering of the content titles of the second target content yields multiple new negative title clusters; Semantic clustering is performed on the content fragment features of the second target content to obtain multiple new negative visual feature clusters; The object labels are updated based on the newly added negative title cluster and the newly added negative visual feature cluster.

6. The content analysis method according to claim 1, characterized in that, The step of determining the abnormal risk level and content risk description of the content to be analyzed based on multiple content fragments and the content text includes: Extract content elements from the content text and identify abnormal elements in the content fragments. The content elements indicate content with actual semantic meaning, and the abnormal elements indicate visual content with potential risks. Using a scene knowledge graph, the first abnormal risk content associated with the content element, the risk level corresponding to the content element, the second abnormal risk content associated with the abnormal element, and the risk level corresponding to the abnormal element are determined respectively. The scene knowledge graph is a knowledge graph used to describe the association between abnormal risk content clusters and object clusters. The abnormal risk level of the content to be analyzed is determined based on the risk level corresponding to the content element and the risk level corresponding to the abnormal element, and the content risk description is determined based on the first abnormal risk content and the second abnormal risk content.

7. The content analysis method according to claim 6, characterized in that, The scene knowledge graph is constructed in the following way: Obtain negative feedback text from multiple platforms, where the negative feedback text refers to negative feedback content published by multiple objects on the platforms; Extract the content entities and object entities from the negative feedback text from the multi-platform platforms; Based on the multi-platform negative feedback text analysis, the association between the object entity and the content entity is determined, and the risk level of the association between the object entity and the content entity is determined. The content entity and the object entity are identified as nodes, the association between the object entity and the content entity is identified as node association information, the risk level of the association between the object entity and the content entity is identified as node association attribute information, and a scene knowledge graph is constructed based on each node, the node association information, and the node association attribute information.

8. The content analysis method according to claim 6, characterized in that, The process of utilizing a scenario knowledge graph to determine the first abnormal risk content associated with the content element, the risk level corresponding to the content element, the second abnormal risk content associated with the abnormal element, and the risk level corresponding to the abnormal element includes: Determine the first content entity corresponding to the content element and the second content entity corresponding to the abnormal element in the scene knowledge graph; Based on the scenario knowledge graph, the risk level of the first object entity associated with the first content entity, the association between the first content entity and the first object entity, and the association between the first content entity and the first object entity is determined, and the risk level of the association between the first content entity and the first object entity is determined as the risk level corresponding to the content element, and the association between the first content entity and the first object entity is determined as the first abnormal risk content. Based on the scenario knowledge graph, the risk levels of the second object entity associated with the second content entity, the association between the second content entity and the second object entity, and the association between the second content entity and the second object entity are determined. The risk level of the association between the second content entity and the second object entity is determined as the risk level corresponding to the abnormal element, and the association between the second content entity and the second object entity is determined as the second abnormal risk content.

9. The content analysis method according to claim 6, characterized in that, The identification of abnormal elements in the content fragment includes: For each content segment, extract multiple content segment features from the content segment; The abnormal element is obtained by performing abnormal element identification processing on the features of the content fragment.

10. The content analysis method according to claim 1, characterized in that, The process of invoking a third-party intelligent agent to determine the analysis results of the content to be analyzed based on the object tag, the anomaly risk level, and the content risk description includes: The third intelligent agent analyzes the object tag, the abnormal risk level, and the content risk description to obtain the conflict judgment of the object tag and the content risk description, the abnormal type of the content to be analyzed, and the abnormal cause. The analysis results are generated based on the conflict judgment of the object tag and the content risk description, the anomaly type of the content to be analyzed, and the cause of the anomaly.

11. The content analysis method according to claim 10, characterized in that, The third agent is trained in the following way: Obtain the negative feedback text from multiple platforms, the content text of the negative feedback from multiple platforms, and the object tag of the object that published the negative feedback text from multiple platforms, wherein the negative feedback content from multiple platforms is the content targeted by the negative feedback text from multiple platforms; The large language model is used to analyze the negative feedback text from multiple platforms and the content text of the negative feedback content from multiple platforms to obtain the conflict judgment between the object tags of the object that published the negative feedback text from multiple platforms and the content text of the negative feedback content from multiple platforms, the abnormal type of the negative feedback content from multiple platforms and the abnormal reason. The third agent is trained using the content text of the multi-platform negative feedback content, the object tags of the objects that publish the multi-platform negative feedback text, the conflict judgment, the anomaly type of the multi-platform negative feedback content, and the anomaly cause as training data.

12. The content analysis method according to claim 11, characterized in that, The training of the third agent includes: The training data is split into fine-tuning training data and reinforcement training data; Based on the fine-tuning training data, the third agent is subjected to supervised fine-tuning. Based on the enhanced training data, the third agent after supervised fine-tuning is further enhanced and fine-tuned.

13. The content analysis method according to claim 12, characterized in that, The step of performing enhanced fine-tuning on the supervised fine-tuning third agent based on the enhanced training data includes: The content text and object labels in the reinforcement training data are used as input data and input into the third intelligent agent to obtain the output results of the third intelligent agent. The output results include the conflict judgment between the object labels and the content text of the object that published the negative feedback text on the multi-platform, the abnormal type of the negative feedback content on the multi-platform, and the abnormal reason. When the conflict determination indicates that there is a conflict between the object tag of the object that publishes the negative feedback text on the multi-platform platform and the content text, the model parameters of the third agent are adjusted based on the first reward weight; When the conflict determination indicates that the object tag of the object that publishes the negative feedback text on the multi-platform conflicts with the content text, and the anomaly type of the negative feedback content on the multi-platform is consistent with the anomaly type in the reinforcement training data, the model parameters of the third agent are adjusted based on the second reward weight, and the second reward weight is greater than the first reward weight. When the conflict determination indicates that the object tag of the object that publishes the negative feedback text on the multi-platform conflicts with the content text, and the anomaly type of the negative feedback content on the multi-platform is consistent with the anomaly type in the reinforcement training data, and the anomaly cause of the negative feedback content on the multi-platform is consistent with the anomaly cause in the reinforcement training data, the model parameters of the third agent are adjusted based on the third reward weight, and the third reward weight is greater than the second reward weight. Return to the step of inputting the content text and object labels from the reinforcement training data into the third agent until the third agent converges.

14. The content analysis method according to claim 13, characterized in that, After inputting the content text and object labels from the enhanced training data into the third agent to obtain the output of the third agent, the method further includes: The content text and object labels in the enhanced training data are used as input data and input into the reference model to obtain the output result of the reference model, which is a third agent after supervised fine-tuning; Based on the difference between the output of the reference model and the output of the third agent, the deviation penalty weight is determined. The model parameters of the third agent are adjusted based on the deviation penalty weights.

15. The content analysis method according to claim 1, characterized in that, The process of obtaining multiple content fragments corresponding to the content to be analyzed includes: Obtain the sampling time interval; Based on the sampling time interval, the content to be analyzed is sampled to obtain multiple content fragments corresponding to the content to be analyzed.

16. A content analysis device, characterized in that, include: The first acquisition module is used to call the first intelligent agent to acquire the object information and historical operation information of the target object, and determine the object tag of the target object when browsing content based on the object information and the historical operation information, wherein the historical operation information is the operation information of the target object when browsing content. The second acquisition module is used to call the second intelligent agent to acquire multiple content fragments corresponding to the content to be analyzed and the content text corresponding to the content to be analyzed, and to determine the abnormal risk level and content risk description of the content to be analyzed based on the multiple content fragments and the content text, wherein the content text is text used to describe the content of the content to be analyzed; The determination module is used to call a third agent to determine the analysis result of the content to be analyzed based on the object label, the anomaly risk level, and the content risk description. The analysis result is used to indicate the anomaly type and cause of the content to be analyzed. The first agent, the second agent, and the third agent are obtained through joint training.

17. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the content analysis method according to any one of claims 1 to 15.

18. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the content analysis method according to any one of claims 1 to 15.

19. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the content analysis method according to any one of claims 1 to 15.

Images