Artificial intelligence full-process auditing method and system for web series

Abstract

The application is suitable for the technical field of network drama auditing, and provides an artificial intelligence full-process auditing method and system for network drama, which comprises the following steps: analyzing a script based on a natural language model, and constructing a script knowledge graph comprising a macro-story layer, a meso-plot layer and a micro-element layer; the macro-story layer is used for representing the core theme and value orientation of the whole drama; the meso-plot layer automatically divides the script into multiple plot units; the micro-element layer is used for representing the entity elements and element development track within each plot unit; based on the script knowledge graph, dynamic risk identification is performed on the script to determine potential risk plot units; the rough cut sample and the script knowledge graph are aligned, and the risk index of the sample segment corresponding to the potential risk plot unit is calculated. The multi-level script knowledge graph constructed based on the natural language model can comprehensively and systematically understand the script content, and improves the accuracy and consistency of the auditing result.

Description

Technical Field

[0001] This invention relates to the field of online drama review technology, specifically to an AI-powered full-process review method and system for online dramas. Background Technology

[0002] Because online dramas have relatively low barriers to entry, spread quickly, and have a wide audience, their content quality varies greatly, and some works may have issues such as value deviations, inappropriate plots, or illegal elements.

[0003] Traditional online drama review methods primarily rely on manual review. Reviewers need to meticulously examine the script word by word and repeatedly watch sample episodes, using their professional knowledge and experience to determine if the content poses any risks. This method is not only inefficient, but also suffers from subjectivity and limitations, making it difficult to guarantee the accuracy and consistency of the review results.

[0004] In recent years, artificial intelligence (AI) technology has been widely applied in various fields, and it has also shown potential in content moderation. For example, some basic moderation technologies based on natural language processing and image recognition can initially identify simple violations in text and images. However, these technologies are mostly limited to single-level analysis, lacking a systematic and interconnected understanding of the script and sample content. They cannot accurately grasp the overall structure of the story, plot development, and the complex relationships between various elements, thus making it difficult to achieve comprehensive and accurate risk identification and effective compliance processing. Therefore, there is a need to provide an AI-powered end-to-end moderation method and system for online dramas, aiming to solve the above problems. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an AI-based full-process review method and system for online dramas, so as to solve the problems existing in the above-mentioned background technology.

[0006] This invention is implemented as follows: an AI-powered end-to-end review method for online dramas, the method comprising the following steps:

[0007] The script is analyzed based on a natural language model, and a script knowledge graph is constructed, which includes a macro-story layer, a meso-plot layer, and a micro-element layer. The macro-story layer is used to represent the core theme and value orientation of the whole play. The meso-plot layer automatically divides the script into multiple plot units and constructs the causal and emotional evolution relationships between plot units. The micro-element layer is used to represent the entity elements and the development trajectory of elements within each plot unit.

[0008] Based on the script knowledge graph and combined with contextual information, dynamic risk identification is performed on the script to determine potential risk plot units;

[0009] Receive rough cut sample, align the rough cut sample with the script knowledge graph, and calculate the risk index of the sample segment corresponding to the potential risk plot unit;

[0010] Sample clips with a risk index higher than the risk threshold are identified as risk clips. Based on generative adversarial networks, alternative visual solutions that are similar in artistic expression and compliant are generated for the risk clips.

[0011] As a further aspect of the present invention, the step of constructing a script knowledge graph comprising a macro-story layer, a meso-plot layer, and a micro-element layer specifically includes:

[0012] Semantic analysis of the script was conducted using text summarization and topic modeling to determine the core themes and value orientations of the entire play.

[0013] Based on script format and semantic turning point detection, the script is segmented into multiple continuous plot units; a semantic embedding vector is generated for each plot unit, and the causal chain and emotional evolution relationship between plot units are constructed by calculating the similarity and sequence relationship between unit vectors.

[0014] Extract the physical elements from the script, including characters, actions, objects, and dialogue. For core elements that run through multiple plot units, record the changes in the state and the evolution of relationships of the core elements in different plot units to form the development trajectory of the elements.

[0015] As a further aspect of the present invention, the step of dynamically identifying risks in the script and determining potential risk plot elements specifically includes:

[0016] Input the entity elements in the micro-element layer into the risk pattern library to determine whether a risk warning is triggered. When triggered, identify the risk elements and initiate a three-layer context check.

[0017] Conduct macro-level verification: check whether risk factors contradict the core themes and value orientations of the macro-level narrative; if they do, they are identified as potential risks.

[0018] Conduct mid-level verification: determine the causal and emotional evolution relationship associated with the corresponding risk elements, and determine whether the risk elements are punished in the subsequent plot units corresponding to the causal and emotional evolution relationship. If they are not punished, they are identified as potential risks.

[0019] Conduct micro-level verification: analyze the role development trajectory corresponding to the risk factors, determine whether the role development trajectory is consistent with the role motivation, and identify potential risks when it is inconsistent.

[0020] As a further aspect of the present invention, the step of aligning the rough cut sample with the script knowledge graph and calculating the risk index of the sample segment corresponding to the potential risk plot unit specifically includes:

[0021] Identify the dialogue text of the sample, perform semantic similarity matching between the dialogue text and the micro-element layer of the script knowledge graph, use this as an anchor point to establish the mapping relationship between the timeline of the sample and the plot unit, and complete the timeline alignment.

[0022] Identify sample clips corresponding to potential risk plot units and extract sample clip features, including character features, object features, action features, and sound effect features for each shot;

[0023] The sample features are input into a multivariate risk base database for risk element matching to obtain the risk base corresponding to the features of people, objects, actions and sound effects, and the risk index is calculated.

[0024] As a further aspect of the present invention, the step of generating alternative visual solutions for risky segments that are artistically similar and compliant specifically includes:

[0025] Based on the matching risk elements, compliance constraints are determined, and the compliance constraints and risk segments are input into a pre-trained video-to-video translation generative adversarial network.

[0026] Under the supervision of the discriminator, the differences between the generated segments and the risk segments in non-risk-related artistic features are minimized to generate candidate alternative video segments that meet compliance constraints.

[0027] As a further aspect of the present invention, when training a video-to-video translation generative adversarial network, pairs of video segments are first collected as a training set. Each training sample contains an original segment and a corresponding compliant reference segment. The generator adopts an encoder-transformer-decoder structure, with the original frame sequence as input, and the discriminator is used to distinguish between generated frames and real compliant frames.

[0028] Another objective of this invention is to provide an AI-powered end-to-end review system for online dramas, the system comprising:

[0029] The script graph construction module is used to analyze scripts based on natural language models and construct a script knowledge graph that includes a macro-story layer, a meso-plot layer, and a micro-element layer. The macro-story layer is used to represent the core theme and value orientation of the entire play. The meso-plot layer automatically divides the script into multiple plot units and constructs the causal and emotional evolution relationships between plot units. The micro-element layer is used to represent the entity elements and the development trajectory of elements within each plot unit.

[0030] The potential risk identification module is used to dynamically identify potential risks in the script based on the script knowledge graph and in combination with contextual information, and to determine potential risk plot units.

[0031] The risk index calculation module is used to receive rough cut samples, align the rough cut samples with the script knowledge graph, and calculate the risk index of the sample segments corresponding to the potential risk plot units.

[0032] The alternative solution production module is used to identify sample segments with a risk index higher than the risk threshold as risk segments. Based on adversarial generative network technology, it generates alternative visual solutions that are similar in artistic expression and compliant for risk segments.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] This invention utilizes a multi-layered script knowledge graph constructed based on a natural language model. This graph enables a comprehensive and systematic understanding of script content, dynamically identifying risks by incorporating contextual information. This reduces the subjectivity and limitations of manual review, improving the accuracy and consistency of review results. By deeply analyzing the script at the macro-story, meso-plot, and micro-element levels, it accurately grasps the core themes, plot development, and relationships between various elements, thus comprehensively and precisely identifying potential risky plot units and effectively preventing omissions. Furthermore, for identified risky segments, adversarial generative network technology is used to generate artistically similar and compliant alternative visual solutions, providing creators with specific modification guidance. Attached Figure Description

[0035] Figure 1 This is a flowchart of an AI-powered end-to-end review method for online dramas.

[0036] Figure 2 This is a flowchart for constructing a script knowledge graph in an AI-powered end-to-end review method for online dramas.

[0037] Figure 3 This is a flowchart for identifying potential risk plot elements in an AI-powered end-to-end review method for online dramas.

[0038] Figure 4 This is a flowchart for calculating the risk index in an AI-powered end-to-end review method for online dramas.

[0039] Figure 5 A flowchart for generating alternative visual solutions in an AI-powered end-to-end review method for online dramas.

[0040] Figure 6 This is a schematic diagram of the structure of an AI-powered end-to-end review system for online dramas. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0042] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0043] like Figure 1 As shown in the figure, this invention provides an AI-powered end-to-end review method for online dramas, the method comprising the following steps:

[0044] S100 analyzes scripts based on natural language models and constructs a script knowledge graph that includes a macro-story layer, a meso-plot layer, and a micro-element layer.

[0045] S200, Based on the script knowledge graph and combined with contextual information, perform dynamic risk identification on the script and determine potential risk plot units;

[0046] S300, Receive rough cut sample, align the rough cut sample with the script knowledge graph, and calculate the risk index of the sample segment corresponding to the potential risk plot unit;

[0047] S400 identifies sample segments with a risk index higher than the risk threshold as risk segments. Based on adversarial generative network technology, it generates alternative visual solutions for risk segments that are similar in artistic expression and comply with regulations.

[0048] It should be noted that existing manual review and basic AI review technologies can often only identify obvious violations. They struggle to analyze and accurately identify potentially risky content and complex structures, leading to omissions. Furthermore, when risky content is discovered, traditional methods typically only require creators to make modifications, lacking specific, artistically similar, and compliant alternatives. Creators may lack guidance and struggle to effectively improve their work, impacting quality and motivation. This invention aims to address these issues.

[0049] In this embodiment of the invention, a natural language processing (NLP) model is first used to analyze the script. Recommended NLP models include ChatGLM3, Qwen1.5-7B / 14B, and Baichuan2-13B, which are pre-trained on high-quality Chinese corpora and support long texts. Then, a multi-layered script knowledge graph is constructed. The macro-story layer represents the core theme and value orientation of the entire script. The meso-plot layer automatically divides the script into multiple plot units and constructs causal and emotional evolution relationships between these units. The micro-element layer represents the entity elements (actions, dialogues, objects, etc.) and their development trajectories within each plot unit, primarily referring to character development trajectories. This multi-layered script knowledge graph enables a comprehensive and systematic understanding of the script content, facilitating dynamic risk identification based on contextual information, reducing the subjectivity and limitations of manual review, and improving the accuracy and consistency of the review results. Next, the script is analyzed in depth from the macro-story level, meso-plot level, and micro-element level, based on the contextual relationships of the script knowledge graph. This allows for an accurate grasp of the core theme, plot development, and relationships between various elements, thus comprehensively and accurately identifying potential risk plot units and effectively preventing omissions during review. Next, users need to upload rough cut samples. This embodiment of the invention automatically aligns the rough cut samples with the script knowledge graph and calculates the risk index of the sample segments corresponding to potential risk plot units. This significantly shortens the review cycle, improves review efficiency, and allows online dramas to be released to audiences faster. Then, sample segments with risk indices higher than a pre-set risk threshold are identified as risky segments. Based on adversarial generative network technology, alternative visual solutions that are artistically similar and compliant are generated for the risky segments. This provides creators with specific modification guidance, helping them quickly complete compliant modifications while ensuring the artistic quality of their work, thus promoting the healthy development of the online drama industry.

[0050] like Figure 2 As shown, in a preferred embodiment of the present invention, the step of constructing a script knowledge graph including a macro-story layer, a meso-plot layer, and a micro-element layer specifically includes:

[0051] S101, using text summarization and topic modeling to perform semantic analysis on the script, to determine the core theme and value orientation of the entire play;

[0052] S102, based on script format and semantic turning point detection, divides the script into multiple continuous plot units; generates a semantic embedding vector for each plot unit, and constructs the causal chain and emotional evolution relationship between plot units by calculating the similarity and sequence relationship between unit vectors;

[0053] S103: Extract the entity elements from the script. Entity elements include characters, actions, objects, and dialogue. For core elements that run through multiple plot units, record the state changes and relationship evolution of the core elements in different plot units to form the development trajectory of the elements.

[0054] In this embodiment of the invention, text summarization and topic modeling, combined with semantic analysis of the entire script, are first used to automatically summarize the core theme of the entire drama and determine its value orientation, thus obtaining the macro-story layer. Then, based on the script format (such as scene titles) and semantic turning point detection, semantic turning points can be detected by using a language model to calculate semantic abrupt changes between paragraphs, dividing the script into multiple continuous plot units, each describing a relatively complete plot segment. Next, a semantic embedding vector is generated for each plot unit. Specifically, for each segmented plot unit, all its textual content (scene descriptions, dialogues, action cues) is input into a pre-trained narrative-aware language model. This model, trained on film and television script data, can understand the unique narrative structure of the script; the model outputs a fixed-dimensional semantic embedding vector, which comprehensively represents the core event, character interaction, and emotional tone of the unit. Then, the causal correlation between the embedding vectors of continuous plot units is calculated. First, by analyzing the tense, modality, and event state changes within the unit, a sequence model is used to predict causal probability. Second, the projection and correlation of the unit vector in a specific "narrative causality" semantic space are calculated. Plot unit pairs with a correlation higher than a threshold are marked as having a potential causal relationship, and causal chain hypotheses are generated. Additionally, the embedding vector of each plot unit is mapped to a continuous emotional dimension value (e.g., "pleasure-excitement") through an emotion classifier. An emotional curve for the entire play is plotted according to the time sequence of plot units. The inflection points, trends, and periodicity of the emotional curve are analyzed to construct emotional contrasts, progressions, or turning points between plot units. Causal relationships and emotional evolution relationships are added as connection attributes to the meso-plot layer of the script knowledge graph, connecting the nodes of each plot unit to form a dynamic network structure with narrative logic. Finally, characters, actions, objects, and dialogue elements in the script are extracted. For core elements that run through multiple plot units, the state changes and relationship evolution of these core elements within different plot units are recorded, forming an element development trajectory. This yields the micro-element layer.

[0055] like Figure 3 As shown, in a preferred embodiment of the present invention, the step of dynamically identifying risks in the script and determining potential risk plot elements specifically includes:

[0056] S201: Input the entity elements in the micro-element layer into the risk pattern library, determine whether a risk warning is triggered, and when triggered, identify the risk elements and start the three-layer context check.

[0057] S202, conduct macro-level verification: check whether the risk factors are contrary to the core theme and value orientation of the macro story layer. If they are contrary, they are identified as potential risks.

[0058] S203, conduct mid-level verification: determine the causal and emotional evolution relationship associated with the corresponding risk element, and determine whether the risk element is punished in the subsequent plot unit corresponding to the causal and emotional evolution relationship. If it is not punished, it is identified as a potential risk.

[0059] S204, conduct micro-level verification: analyze the role development trajectory corresponding to the risk factors, determine whether the role development trajectory is consistent with the role motivation, and identify potential risks if it is inconsistent.

[0060] In this embodiment of the invention, a risk pattern library is pre-set, where each pattern defines triggering conditions, which are directly related to entity elements at the micro-level. Entity elements at the micro-level are input into the risk pattern library for matching. The risk pattern library automatically determines whether a risk warning is triggered. When triggered, a macro-level check is first performed: it checks whether the risk element contradicts the core theme and value orientation of the macro-story layer. If it does, it is identified as a potential risk; otherwise, it is considered normal. Then, a meso-level check is performed: it determines the causal and emotional evolution relationship associated with the corresponding risk element, and determines whether the risk element is punished in the subsequent plot unit corresponding to the causal and emotional evolution relationship. If it is not punished, it means the risky behavior has not been criticized, and it is identified as a potential risk; if it is punished, the plot unit is a normal plot. Finally, a micro-level check is performed: it analyzes the character development trajectory corresponding to the risk element, and determines whether the character development trajectory conforms to the character's motivation. If it does not conform, for example, if the risky behavior occurs on a positive character, it is identified as a potential risk; if it does conform, the corresponding plot unit is a normal plot. During the three-layer context check, as long as one layer is identified as a potential risk, the corresponding plot unit is the potential risk plot unit.

[0061] like Figure 4 As shown, in a preferred embodiment of the present invention, the step of aligning the rough cut sample with the script knowledge graph and calculating the risk index of the sample segment corresponding to the potential risk plot unit specifically includes:

[0062] S301, Identify the dialogue text of the sample, perform semantic similarity matching between the dialogue text and the micro-element layer of the script knowledge graph, use this as an anchor point to establish the mapping relationship between the timeline of the sample and the plot unit, and complete the timeline alignment.

[0063] S302, identify the sample clips corresponding to the potential risk plot units, and extract the sample clip features. The sample clip features include the character features, object features, action features and sound effect features of each shot.

[0064] S303: Input the sample features into the multivariate risk base database for risk element matching to obtain the risk base corresponding to the features of people, objects, actions and sound effects, and calculate the risk index.

[0065] In this embodiment of the invention, further risk analysis is conducted in conjunction with the rough-cut sample. First, the dialogue text of the sample needs to be identified. Semantic similarity matching is then performed between the dialogue text and the micro-element layer of the script knowledge graph. Using the matching relationship between the dialogue text and the micro-element layer as anchor points, a mapping relationship is established between the timeline of the sample and the plot units, aligning the timeline of the rough-cut sample with that of the plot units in the script knowledge graph. Then, the sample segments corresponding to potential risk plot units are determined, and the character features, object features, action features, and sound effect features of each shot in the sample segments are extracted. It should be noted that this embodiment of the invention pre-constructs a risk base library containing a large number of risk element features. Each risk element corresponds to a risk base number. Inputting the sample features into the multivariate risk base library for risk element matching outputs several matching risk base numbers. All risk base numbers are then summed to obtain the risk index.

[0066] like Figure 5 As shown, in a preferred embodiment of the present invention, the step of generating alternative visual solutions for risky segments that are artistically similar and compliant specifically includes:

[0067] S401, determine compliance constraints based on matching risk elements, and input the compliance constraints and risk segments into a pre-trained video-to-video translation generative adversarial network;

[0068] S402, under the supervision of the discriminator, minimize the difference between the generated segment and the risk segment in non-risk-related artistic features, and generate candidate alternative video segments that meet compliance constraints.

[0069] In this embodiment of the invention, compliance constraints are determined based on matching risk elements. These compliance constraints aim to reduce the corresponding risk elements and serve as control signals. The compliance constraints and risk segments are then input into a pre-trained video-to-video translation generative adversarial network (GAN). Under the supervision of the GAN's discriminator, the differences between the generated segments and the risk segments in non-risk-related artistic features are minimized, generating candidate alternative video segments that satisfy the compliance constraints.

[0070] When training the video-to-video translation generative adversarial network, a large number of paired video clips are collected as the training set. Each sample contains an original clip and a corresponding "compliant reference clip" that has been manually reviewed and modified. This data needs to cover various risk types and artistic styles. The generator adopts an encoder-transformer-decoder structure, taking the original frame sequence as input, and a discriminator to distinguish generated frames from real compliant frames. The composite loss function in the network includes adversarial loss, content reconstruction loss, and style and temporal consistency loss. The adversarial loss drives the generated content distribution to converge towards the compliant video distribution; the content reconstruction loss constrains the consistency between the generated clip and the original clip in non-risk areas (such as background and character identity); and the style and temporal consistency loss ensures that the lighting, tone, and motion rhythm of the generated clip are consistent with the original clip. During iterative training, data pairs are input into the network, and the generator and discriminator are jointly optimized through backpropagation. During training, conditional inputs (such as risk type labels) guide the generation direction until the network can stably output visually natural and content-compliant video conversion results.

[0071] like Figure 6 As shown in the figure, this invention also provides an AI-powered end-to-end review system for online dramas, the system comprising:

[0072] The script graph construction module 100 is used to analyze scripts based on natural language models and construct a script knowledge graph including a macro story layer, a meso plot layer, and a micro element layer. The macro story layer is used to represent the core theme and value orientation of the whole play. The meso plot layer automatically divides the script into multiple plot units and constructs the causal and emotional evolution relationships between plot units. The micro element layer is used to represent the entity elements and the development trajectory of elements within each plot unit.

[0073] The potential risk identification module 200 is used to dynamically identify potential risks in the script based on the script knowledge graph and in combination with context information, and to determine potential risk plot units.

[0074] The risk index calculation module 300 is used to receive rough cut samples, align the rough cut samples with the script knowledge graph, and calculate the risk index of the sample segments corresponding to the potential risk plot units.

[0075] The alternative solution production module 400 is used to identify sample segments with a risk index higher than the risk threshold as risk segments. Based on adversarial generative network technology, it generates alternative visual solutions that are similar in artistic expression and compliant for risk segments.

[0076] In a preferred embodiment of the present invention, the script graph construction module 100 includes:

[0077] Thematic and value-based units are used to perform semantic analysis of the script using text summarization and theme models to determine the core themes and value orientations of the entire play.

[0078] The plot unit determination unit is used to segment the script into multiple consecutive plot units based on the script format and semantic turning point detection; a semantic embedding vector is generated for each plot unit, and the causal chain and emotional evolution relationship between plot units are constructed by calculating the similarity and sequence relationship between unit vectors.

[0079] The entity element identification unit is used to extract entity elements from the script. Entity elements include characters, actions, objects, and dialogues. For core elements that run through multiple plot units, the unit records the state changes and relationship evolution of the core elements in different plot units, forming an element development trajectory.

[0080] In a preferred embodiment of the present invention, the potential risk identification module 200 includes:

[0081] The risk factor early warning unit is used to input entity elements in the micro-element layer into the risk pattern library, determine whether a risk warning is triggered, and when triggered, identify the risk element and initiate a three-layer context check.

[0082] The macro-level verification unit is used to conduct macro-level verification: checking whether risk factors are contrary to the core themes and value orientations of the macro-level story layer; when they are contrary, they are identified as potential risks.

[0083] The meso-level verification unit is used to conduct meso-level verification: determine the causal and emotional evolution relationship associated with the corresponding risk element, determine whether the risk element is punished in the subsequent plot unit corresponding to the causal and emotional evolution relationship, and if it is not punished, it is identified as a potential risk;

[0084] The micro-level verification unit is used to conduct micro-level verification: analyze the role development trajectory corresponding to the risk factors, determine whether the role development trajectory is consistent with the role motivation, and identify potential risks when it is inconsistent.

[0085] In a preferred embodiment of the present invention, the risk index calculation module 300 includes:

[0086] The timeline alignment unit is used to identify the dialogue text of the sample, perform semantic similarity matching between the dialogue text and the micro-element layer of the script knowledge graph, and use this as an anchor point to establish the mapping relationship between the timeline of the sample and the plot unit, thus completing the timeline alignment.

[0087] The sample feature extraction unit is used to determine the sample segments corresponding to the potential risk plot units and extract sample features. The sample features include the character features, object features, action features and sound effect features of each shot.

[0088] The risk index calculation unit is used to input sample features into a multivariate risk base library for risk element matching, obtain the risk base corresponding to the features of people, objects, actions and sound effects, and calculate the risk index.

[0089] The above description only details the preferred embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0090] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, 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 various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0091] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0092] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the disclosure in the specification and embodiments. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

Claims

1. An artificial intelligence full-process auditing method for network drama films, characterized in that, The method comprises the following steps: analyzing the script based on a natural language model to construct a script knowledge graph comprising a macro-story layer, a meso-plot layer and a micro-element layer; the macro-story layer is used to represent the core theme and value orientation of the whole script; the meso-plot layer automatically divides the script into a plurality of plot units and constructs the cause-and-effect and emotional evolution relationship between the plot units; the micro-element layer is used to represent the entity elements and element development trajectories within each plot unit; based on the script knowledge graph, combining context information, dynamically identifying the risk of the script to determine the potential risk plot unit; receiving a rough-cut sample, aligning the rough-cut sample with the script knowledge graph, and calculating the risk index of the sample segment corresponding to the potential risk plot unit; determining the sample segment with a risk index higher than a risk threshold as a risk segment, and generating an alternative visual scheme similar in artistic expression and compliant for the risk segment based on a generative adversarial network technology; wherein the step of dynamically identifying the risk of the script to determine the potential risk plot unit comprises: inputting the entity elements in the micro-element layer into a risk pattern library to determine whether a risk warning is triggered, and when the risk warning is triggered, determining a risk element and starting a three-layer context check; performing macro-layer checking: checking whether the risk element is contrary to the core theme and value orientation of the macro-story layer, and when it is contrary, determining it as a potential risk; performing meso-layer checking: determining the cause-and-effect and emotional evolution relationship associated with the corresponding risk element, and determining whether the risk element is punished in the subsequent plot unit corresponding to the cause-and-effect and emotional evolution relationship, and when it is not punished, determining it as a potential risk; performing micro-layer checking: analyzing the role development trajectory corresponding to the risk element, and determining whether the role development trajectory is consistent with the role motivation, and when it is not consistent, determining it as a potential risk. 2.The network drama-oriented artificial intelligence full-process auditing method according to claim 1, characterized in that, The step of constructing the script knowledge graph comprising a macro-story layer, a meso-plot layer and a micro-element layer comprises: performing semantic analysis on the script by using a text summary and a theme model to determine the core theme and value orientation of the whole script; dividing the script into a plurality of continuous plot units based on script format and semantic turning point detection; generating semantic embedding vectors for each plot unit, and constructing the cause-and-effect chain and emotional evolution relationship between the plot units by calculating the similarity and sequence relationship between the unit vectors; extracting entity elements in the script, including roles, actions, objects and dialogues, and for core elements running through a plurality of plot units, recording the state changes and relationship evolution of the core elements in different plot units to form element development trajectories. 3.The network drama-oriented artificial intelligence full-process auditing method according to claim 1, characterized in that, The step of aligning the rough-cut sample with the script knowledge graph to calculate the risk index of the sample segment corresponding to the potential risk plot unit comprises: identifying the dialogue text of the sample, and performing semantic similarity matching between the dialogue text and the micro-element layer of the script knowledge graph to establish the mapping relationship between the sample timeline and the plot units, and complete the timeline alignment; determining the sample segment corresponding to the potential risk plot unit, and extracting sample features, including the character features, object features, action features and sound effect features of each shot; The sample feature is input into a multivariate risk base library for risk element matching to obtain risk bases corresponding to character features, object features, action features and sound effect features, and a risk index is calculated.

4. The AI full-process auditing method for network drama films according to claim 3, characterized in that, The step of generating an alternative visual scheme that is similar in artistic expression and compliant for the risk segment specifically includes: According to the matched risk elements, a compliance constraint condition is determined, and the compliance constraint condition and the risk segment are input into a pre-trained video-to-video translation generative adversarial network; Under the supervision of the discriminator, the difference between the generated segment and the risk segment in non-risk-related artistic features is minimized, and a candidate alternative video segment that meets the compliance constraint is generated.

5. The network drama-oriented artificial intelligence full-process auditing method according to claim 4, characterized in that, When training the video-to-video translation generative adversarial network, first, pairs of video segments are collected as a training set, each training sample including an original segment and a corresponding compliant reference segment, the generator adopts an encoder-transformer-decoder structure, and the original frame sequence is input, and the discriminator is used to distinguish between generated frames and true compliant frames.

6. An artificial intelligence full-process auditing system for network drama films, characterized in that, The system comprises: A script graph construction module is configured to analyze a script based on a natural language model and construct a script knowledge graph including a macro-story layer, a meso-plot layer and a micro-element layer; the macro-story layer is configured to represent core themes and value orientation of the whole script; the meso-plot layer is configured to automatically segment the script into a plurality of plot units and construct causal and emotional evolution relationships between the plot units; and the micro-element layer is configured to represent entity elements and element development trajectories within each plot unit. A potential risk identification module is configured to identify potential risks in the script based on the script knowledge graph and in combination with context information, and determine a potential risk plot unit. A risk index calculation module is configured to receive a rough-cut sample, align the rough-cut sample with the script knowledge graph, and calculate a risk index of a sample segment corresponding to the potential risk plot unit. An alternative scheme production module is configured to determine a sample segment with a risk index higher than a risk threshold as a risk segment, and generate an alternative visual scheme that is similar in artistic expression and compliant for the risk segment based on an adversarial generative network technology. The potential risk identification module includes: a risk element early warning unit configured to input entity elements in the micro-element layer into a risk pattern library to determine whether a risk early warning is triggered, and when the risk early warning is triggered, determine a risk element and start three-layer context verification; a macro-layer verification unit configured to perform macro-layer verification: check whether the risk element is contrary to the core themes and value orientation of the macro-story layer, and when the risk element is contrary to the core themes and value orientation of the macro-story layer, determine that the risk element is a potential risk; a meso-layer verification unit configured to perform meso-layer verification: determine a causal and emotional evolution relationship associated with the corresponding risk element, and determine whether the risk element is punished in a subsequent plot unit corresponding to the causal and emotional evolution relationship, and when the risk element is not punished, determine that the risk element is a potential risk; and a micro-layer verification unit configured to perform micro-layer verification: analyze a role development trajectory corresponding to the risk element, and determine whether the role development trajectory is consistent with a role motivation, and when the role development trajectory is not consistent with the role motivation, determine that the risk element is a potential risk.

7. The network drama-oriented artificial intelligence full-process auditing system according to claim 6, characterized in that, The script graph construction module includes: A theme value unit is configured to perform semantic analysis on the script by using a text summary and a theme model to determine core themes and value orientations of the whole script; A plot unit determination unit is configured to divide the script into a plurality of continuous plot units based on script format and semantic turning point detection, generate semantic embedding vectors for each plot unit, and construct a cause-and-effect chain and an emotional evolution relationship between the plot units by calculating similarity and sequence relationship between the vectors; An entity element determination unit is configured to extract entity elements in the script, including characters, actions, objects, and dialogues, record state changes and relationship evolutions of core elements in different plot units for core elements running through a plurality of plot units, and form an element development track. 8.The network drama-oriented artificial intelligence full-process auditing system according to claim 6, characterized in that, The risk index calculation module includes: A time axis alignment unit is configured to identify dialogue text of the sample, perform semantic similarity matching between the dialogue text and a micro-element layer of the script knowledge graph, use the semantic similarity matching as an anchor point, establish a mapping relationship between a sample time axis and a plot unit, and complete time axis alignment; A sample feature extraction unit is configured to determine a sample segment corresponding to a potential risk plot unit, and perform sample feature extraction, including character features, object features, action features, and sound effect features of each shot; A risk index calculation unit is configured to input the sample features into a multi-element risk base to perform risk element matching, obtain risk bases corresponding to the character features, the object features, the action features, and the sound effect features, and calculate a risk index.

Images