An AI tag generation and completion system and method based on media resources

Abstract

The application discloses an AI tag generation and completion system and method based on media, and relates to the technical field of broadcast television and network audio-visual media management. The system comprises: a multi-modal intelligent identification engine, which selectively calls networking search, subtitle analysis, ASR transcription or image recognition components to generate multi-source structured description information according to the completeness of media information; a standardized tag generation module, which generates tag data containing basic metadata, three-level classification tags and key scene timestamps according to preset tag standard specifications; a dynamic completion and verification module, which is used for field integrity verification, format correction and multi-data source fault tolerance switching; and a business adaptation interface module, which synchronizes the tag data to a downstream business system. The application realizes automatic tag generation for media with different information completeness through a differentiated generation strategy, solves the problems of missing tags for unpopular media, non-uniform tag formats, lack of key scene timestamps and the like, and improves the tag generation efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of broadcast television and online audiovisual media asset management technology, specifically to an AI tag generation and completion system and method based on media assets. Background Technology

[0002] With the accelerated transformation of businesses towards full IP and intelligentization, the media asset library is experiencing explosive growth, and users' demands for searching, recommending, and personalized consumption of media content are becoming increasingly diversified. However, existing film and television and product media asset management systems have the following significant limitations: First, some niche, self-made, or replay-type media assets lack relevant reference information on the internet, and the tags are severely missing. Popular media asset tags do not fully follow a unified standard, and the format is chaotic and the dimensions are incomplete, making it difficult to support accurate application.

[0003] Secondly, almost all media assets lack key plot markers with timestamps, making it impossible to achieve intelligent functions such as plot jumps and extraction of highlight scenes, which restricts the upgrading of the user's viewing experience.

[0004] Third, existing tag completion relies on manual input, which is costly, time-consuming, and prone to human error. It lacks an integrated solution based on artificial intelligence and has not formed a closed-loop mechanism of "information collection - tag generation - dynamic completion - business adaptation". As a result, the existing tag system is unable to meet the needs of emerging business scenarios such as AI search, personalized recommendation, and intelligent marketing, and the value of media assets is not fully explored.

[0005] Therefore, there is an urgent need for an intelligent solution that can automatically generate and complete media asset tags, support full-scenario adaptation of different types of media assets, and ensure the standardization and compatibility of tag data, so as to at least solve some of the above problems. Summary of the Invention

[0006] To overcome the shortcomings mentioned above, this invention aims to provide an AI-based tag generation and completion system and method for media assets. By integrating core technologies such as network search, multimodal recognition, and large language models, it constructs a fully intelligent solution encompassing "intelligent recognition - tag generation - dynamic completion - business empowerment." This solution automatically generates and completes tags for film and television and product media assets, resolving issues such as missing tags, inconsistent formats, and incomplete data. It supports applications in multiple business scenarios, including AI search, personalized recommendations, and intelligent marketing.

[0007] To achieve the above objectives, in one aspect, the present invention provides the following technical solution: an AI tag generation and completion system based on media assets, comprising: The multimodal intelligent recognition engine is used to receive media assets to be processed and, based on the determination result of the completeness of the media asset information, selectively call one or more processing methods from the network search plugin, subtitle text analysis component, ASR speech transcription component or image recognition component to generate multi-source structured description information. The standardized tag generation module is connected to the multimodal intelligent recognition engine and is used to generate tag data containing basic metadata, three-level classification tags and key plot timestamps according to the multi-source structured description information and the preset media asset tag standard specifications. The three-level classification tags include first-level mandatory single tags, second-level mandatory multiple tags and third-level conditional completion tags. The dynamic completion and verification module is connected to the standardized label generation module. It is used to perform field integrity verification, format verification and automatic correction on the label data, and automatically switch to other recognition methods when a single data source fails, and periodically synchronize the latest Internet information to update existing labels. The business adaptation interface module communicates with the dynamic completion and verification module and is used to synchronize the verified tag data to the media asset library and downstream business systems through a standardized data interface.

[0008] A further technical solution of the present invention: The multimodal intelligent recognition engine is an intelligent engine based on an Agent architecture. The Agent architecture has built-in local media asset CMS database query tool, online search plugin, locally privately deployed ASR speech transcription tool, online image recognition interface, locally privately deployed large language model, preset rule engine and tag whitelist library, and media asset knowledge graph query interface.

[0009] A further technical solution of the present invention: The multimodal intelligent recognition engine determines the processing path of media assets according to the following judgment process: The first step is to search the local media asset library and combine it with online search. If complete metadata can be obtained, it is determined to be a popular and searchable media asset, and online search is initiated for processing. The second step is to determine if the metadata is incomplete but the media asset contains a valid subtitle file. If so, the media asset is determined to have a subtitle file and the subtitle text analysis method is started for processing. The third step involves determining if the metadata is incomplete and there is no valid subtitle file. If so, it is considered a niche or self-made media asset without reference information, and ASR speech transcription and image recognition are initiated for fusion processing.

[0010] A further technical solution of the present invention: When the media asset to be processed simultaneously meets multiple conditions, the multimodal intelligent recognition engine adopts a priority acceptance and multi-source data fusion strategy, and acquires data according to the priority order of local media asset CMS library, online search information, subtitle text analysis results, ASR speech transcription results, and image recognition results. For cases where the same field is inconsistent in different data sources, the confidence weight is determined according to the weight of official information over encyclopedic information, encyclopedic information over text extraction information, and text extraction information over image recognition information. The complementary information from different sources is merged, deduplicated, and then uniformly stored in the database.

[0011] A further technical solution of the present invention: the extraction of the key plot timestamps adopts an unsupervised rule-driven approach, including: For narrative media assets, the entire audio is transcribed using ASR speech transcription. Key plot points are located based on at least one of the following methods: dialogue density mutation points, emotional curve peaks, and scene switching frequency. The start time of the scene switching is used as the start time of the timestamp, and the end time of the plot conflict or scene ending is used as the end time, generating second-level key plot timestamps. For documentary media assets, basic timestamps are automatically generated based on visual shot switching detection and scene recognition results, combined with narrative rhythm rules.

[0012] A further technical solution of the present invention: the types of key plot timestamps include at least one or more of the following: plot turning points, plot climaxes, character growth segments, famous scene segments, and emotional highlight segments.

[0013] A further technical solution of the present invention: the dynamic completion and verification module includes: The integrity verification unit is used to perform integrity checks on required fields and, when missing fields are found, to perform intelligent completion according to the priority of local media asset library, online search plugin, subtitle text extraction, and media asset knowledge graph matching. The format normalization unit is used to unify time-type fields into a predetermined standard format, regularize region-type fields into a predetermined structure, and automatically remove special characters and redundant prefixes from label fields, as well as perform deduplication and sorting. The multi-source conflict resolution unit is used to make a confidence-weighted resolution based on the weights of official information, encyclopedic information, text-extracted information, and image recognition information when information from multiple data sources is inconsistent. The dynamic update unit is used to periodically connect to the network for incremental synchronization of external information and automatically trigger a second review of all fields after each update.

[0014] On the other hand, the present invention provides the following technical solution: a method for generating and completing AI tags based on media assets, applied to the aforementioned AI tag generation and completion system based on media assets, comprising the following steps: S1: Receive media assets to be processed, and based on the determination result of the completeness of the media asset information, selectively perform one or more processing methods among network search, subtitle text analysis, ASR speech transcription or image recognition to generate multi-source structured description information; S2: Based on the multi-source structured description information and the preset media asset tag standard specifications, generate tag data containing basic metadata, three-level classification tags and key plot timestamps, wherein the three-level classification tags include first-level mandatory single tags, second-level mandatory multiple tags and third-level conditional completion tags; S3: Performs field integrity verification, format verification, and automatic correction on the generated tag data. When a data source fails, it automatically switches to other recognition methods and regularly updates existing tags with the latest information from the Internet. S4: Output the verified tag data to the media asset library or downstream business system through a standardized data interface.

[0015] A further technical solution of the present invention: the step S1 of selecting the processing method based on the determination result of the media asset information completeness specifically includes: If complete metadata can be obtained, it is determined to be a popular and searchable media asset, and an online search is initiated. If the metadata is incomplete but a valid subtitle file is available, it is determined to be media asset with a subtitle file, and the subtitle text analysis method is initiated. If the metadata is incomplete and there is no valid subtitle file, it is determined to be an obscure or unreferenced media resource, and ASR speech transcription and image recognition are initiated for fusion processing.

[0016] A further technical solution of the present invention: Step S1 further includes: when the media asset to be processed simultaneously meets multiple conditions, data is acquired according to the priority order of local media asset library, online search information, subtitle text analysis results, ASR speech transcription results, and image recognition results; for cases where the same field is inconsistent in different data sources, a confidence weighted decision is made according to the weights of official information over encyclopedic information, encyclopedic information over text extraction information, and text extraction information over image recognition information, and complementary information from different sources is merged and deduplicated.

[0017] Optionally, the generation of the three-level classification tags in step S2 includes: generating a first-level mandatory single tag as the core category of the media asset for each media asset according to a predetermined standard specification, generating at least one second-level mandatory multi-tag as the media asset subject matter type dimension, and having AI intelligently complete the third-level conditional tags based on the content of the media asset.

[0018] Optionally, the generation of key plot timestamps in step S2 includes: For narrative media assets: complete the full audio transcription, extract keywords of plot conflict and dramatic sentences, perform performance evaluation and scoring of plot segments, and select high-value core segments; call a multimodal visual model to verify the visual relevance of high-scoring text segments and filter out invalid and noisy segments; output second-level timestamps with the start time of scene transition as the start time and the end time of plot conflict fallback or scene end. For documentary media assets: the frequency of scene switching is used to determine key plot points, and intervals where the number of scene switching reaches a predetermined number per minute are marked as key plot points, or the corresponding timestamps are generated by splitting the narrative structure.

[0019] Optionally, the field integrity verification and automatic correction in step S3 includes: performing integrity checks on required fields, and performing intelligent completion according to a fixed priority chain when missing fields are found, prioritizing local media asset libraries, followed by online search plugins, supplementing with subtitles and ASR transcribed text, and verifying with media asset knowledge graphs; for format errors, standardizing time fields to the standard year-month-day format, standardizing region fields to a unified structure, and automatically removing special characters and redundant prefixes from tag fields and completing deduplication and sorting; for self-made or niche media assets, using local standardized fallback values, and ensuring that in the case of missing third-level tags, AI automatically generates no less than a predetermined number of valid tags based on the plot, scene, and theme.

[0020] Optionally, it also includes step S5: continuously monitor the availability of each data source, and automatically downgrade to the next available data source according to a preset priority in the event of any data source failure, return of null value, exception or timeout, so as to ensure the continuity and accuracy of the label generation process.

[0021] Compared with existing technologies, the beneficial effects of this technical solution are as follows: By adopting a differentiated generation strategy based on the completeness of media asset information, it achieves automated tag generation for different types of media assets, solves the problem of missing tags for unpopular and self-made media assets, and improves tag coverage and generation efficiency; by generating key plot markers with timestamps, it supports intelligent functions such as plot jumps and highlight segment extraction, enriching the application scenarios of media assets and improving user experience; through multi-source data fusion and automatic verification and completion mechanisms, it reduces the cost of manual intervention and human error, and improves the accuracy and standardization of tag generation.

[0022] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the module architecture of the AI ​​tag generation and completion system based on media assets of the present invention; Figure 2 This is a flowchart illustrating the AI ​​tag generation and completion method based on media assets according to the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figure 1-2 In this embodiment, the process of generating and completing tags for a popular theatrical film using the system of the present invention is described.

[0027] Enter the description of the media asset: Title: Related Popular Movie Titles Type: Theatrical film Information completeness: The local media asset database contains only partial information, with some fields missing. System processing steps: 1. Media Asset Input: The system loads the media asset file "Related Popular Movie Names" and triggers initialization verification; 2. Agent query local media asset library: Search and match the existing media asset library, and the returned results are missing some basic metadata (missing director, actors, complete plot, standard classification). 3. The agent determines the content to be of a popular and searchable type: based on information such as the title and popularity characteristics, and completes the information via the internet; 4. The agent activates the online search plugin: automatically completes missing authoritative metadata such as director, actors, release region, release date, official genre, and plot synopsis; 5. Use the subtitle NLP analysis component to extract subtitle text, and combine it with a multimodal model to identify scenes, characters, and other content in keyframes of the video; 6. Generate first- and second-level mandatory tags according to the self-constructed three-level tag system, and use AI to intelligently complete the fine-grained third-level tags; 7. Generate key plot timestamps based on dialogue density abrupt changes, sentiment curve peaks, and scene switching frequency; 8. Enter the dynamic completion and validation module to complete field validation, format standardization, and conflict data adjudication; 9. Output standard structured tag data through the business adaptation interface.

[0028] In this embodiment, the process of generating and completing tags for a niche, self-made media asset using the system of the present invention is described.

[0029] Enter media asset description: Title: Related Niche Self-Made Media Asset Names Type: Self-made cultural documentary Information completeness: extremely low; the local media asset library lacks complete metadata, publicly available internet information, standard subtitles, and a tagging system. System processing steps: 1. Media Asset Input: The system imports video files from the "Related Less Popular Self-Made Media Asset Names" list; 2. Agent search of local media asset database: No valid information found, determined to be self-produced / niche media asset. 3. Call the video frame extraction component: Extract keyframes from the video for image recognition; 4. Call the ASR speech transcription component: convert audio to text and generate the original transcribed text; 5. Call the local LLM large model component: correct, complete, and semantically normalize low-quality ASR results with accents or grammatical errors; 6. Call the multimodal image recognition component: perform urban scene, human scene, and daily life scene recognition on keyframes; 7. Call the knowledge graph construction component: extract entities such as region, population characteristics, festival background, and life status to construct a content association graph; 8. Tag Generation Component: Automatically generates required first- and second-level tags, and AI intelligently completes third-level tags; 9. Timestamp Extraction Component: Generates key plot timestamps based on scene transitions, character interviews, and narrative passages; 10. Verification and Input: The dynamic completion verification module uses a unified format to output standardized structured data.

[0030] In some embodiments, for example, the preset media asset labeling standard specification is as follows: The first-level mandatory single tag is the core category of the media asset, and only one tag can be selected for each media asset; The secondary category requires multiple tags, which are media asset themes / types. Each media asset can have one or more tags selected. The third-level conditional completion tag is a fine-grained feature dimension of media assets. It is not mandatory and is intelligently completed by AI based on the content of the media assets. It is used to accurately characterize the features of media assets and support scenarios such as refined recommendation and search.

[0031] In some embodiments, exemplarily, the specific model modules used in the multimodal intelligent recognition engine include: The ASR speech transcription module has a pre-set speech recognition model and is deployed locally and privately. The image recognition module uses a pre-set visual recognition model and is deployed via an online interface. The person and scene recognition module has a pre-set scene and person recognition training model, which can be deployed through an online interface; The text semantic analysis module is pre-configured with a large language model and is deployed locally in a private environment. The tag matching and error correction module uses a preset rule engine and a tag whitelist library, and is deployed locally in a private environment.

[0032] In some embodiments, exemplarily, the specific rules for extracting key plot timestamps are as follows: Key plot timestamp extraction rules are customized for different scenarios, along with scene classification, classic scene extraction, and an exception handling mechanism, as detailed below: 1. Rules for extracting timestamps from narrative media assets (1) Text preprocessing and value scoring: The entire audio is transcribed using the ASR speech transcription module, and invalid empty subtitles, advertising subtitles, and irrelevant content at the beginning and end of the film are automatically filtered out. Keywords of plot conflict and dramatic sentence structure are extracted, and semantic analysis and plot understanding are completed. The preset plot quantification scoring algorithm is called to give the plot segments a quantification score of 0-10 points, and high-value core segments with scores of 8 points or above are selected.

[0033] (2) Multi-dimensional image verification: Call the image recognition module to identify scene, props, costumes and environmental features. Simultaneously combine the character and scene recognition module to complete the matching of character identity, character features, city scene and performance scene. Verify the image correlation of high-scoring text paragraphs and filter out invalid noise paragraphs.

[0034] (3) Precise timestamp positioning: The starting time of the scene change of the verified paragraph is the timestamp start time, and the ending time is the plot conflict fall / scene end point; the timestamp is uniformly in seconds, and the duration of a single interval is 30 seconds to 5 minutes.

[0035] 2. Rules for Supplementing Timestamps for Documentary Media Assets For documentaries, operas, sports events, and other content lacking obvious dramatic conflict, a dual approach of visuals and narrative is used to supplement timestamps: Visual dimension: Based on the image recognition module and the character and scene recognition module, the frequency of scene switching is used to determine the key plot points. The intervals that switch 3 or more times per minute are marked as key plot points.

[0036] Narrative Dimensions: Opera content is broken down into opening, singing segments, and curtain call; humanistic documentaries are broken down into background setup, core narrative, and theme elevation; sports events are broken down into pre-match warm-up, official competition, decisive moment, and award ceremony to generate corresponding timestamps.

[0037] 3. Scene Classification and Classic Scene Extraction Rules (1) Combining the transcribed text from the ASR speech transcription module and the semantic analysis results from the text semantic analysis module, along with the image features identified by the image recognition module and the character and scene recognition module, the plot segments are divided into fixed types such as dialogue, narrative transition, plot conflict, emotional expression, action scene, and flashback.

[0038] (2) Extraction of classic scenes: From the marked timestamps, select plot turning points, famous scenes, exciting scenes, emotional scenes, core lines and fragments, and simultaneously mark the time interval, fragment type, key lines, characters and emotional attributes.

[0039] 4. Exception Handling Rules Subtitle errors, missing timestamps, ASR transcription anomalies: Automatically skip invalid paragraphs and automatically supplement reasonable timestamps based on the narrative rhythm and scene switching patterns of the preceding and following normal paragraphs.

[0040] Blurred character recognition in the footage: We supplemented the film with encyclopedia entries and plot explanations by relying on online searches, and calibrated the plot divisions and timestamp intervals.

[0041] The entire film has no strong dramatic plot: by default, the film is divided into segments evenly according to the basic narrative rhythm to generate timestamps, ensuring that the plot coverage of the entire film is no less than 90%.

[0042] In some embodiments, the dynamic completion and verification module incorporates an automatic verification and correction mechanism that includes field integrity verification, intelligent completion of missing information, data format normalization, multi-source information conflict resolution, anomaly tolerance, and dynamic update review. 1. Integrity Validation and Intelligent Completion: The system first performs integrity checks on required fields such as program name, primary tags, and secondary tags. If any missing fields are found, intelligent completion is immediately performed according to a fixed priority chain: local media asset CMS library first, online search plugin second, subtitle and ASR transcribed text extraction and supplementation, and media asset knowledge graph matching and verification. Among them, director, actor and other character information must be matched with the built-in character knowledge base with high confidence before it can be filled back. 2. Format normalization: To address formatting errors, the system automatically standardizes time fields to the YYYY year MM month DD day standard format, unifies region fields to a predetermined structure, automatically removes special characters and redundant prefixes from tag fields and performs deduplication and sorting, and uses commas to separate multiple names and standardizes their sorting. 3. Multi-source conflict resolution: When inconsistencies occur in information from multiple data sources such as local media asset library, online search, subtitle extraction, and image recognition, the system makes a confidence-weighted resolution based on the authority of official information > encyclopedic information > text extraction information > image recognition information. Conflicting data that cannot be automatically determined will be automatically marked as pending manual review and all original evidence will be retained. 4. Coverage for niche media assets: For self-made, niche, or media assets without internet reference information, the system automatically adopts standardized cover values ​​such as local self-made, unknown, and unpublished to avoid empty values; in case of subtitle errors, abnormal speech recognition, etc., the system retains the valid data from the previous step, does not forcibly overwrite or arbitrarily set empty, and ensures that if the third-level tags are missing, the AI ​​will automatically generate no less than 3 valid tags based on the plot, scene, and theme; 5. Dynamic Update and Review: The system regularly updates incrementally online, synchronizing the latest external information such as new director's works, film awards, and rating changes. After each update, a second review of all fields is automatically triggered to ensure that the tag data remains complete, standardized, accurate, and consistent in the long term, meeting the requirements for standardized media asset storage and business scenario calls.

[0043] In some embodiments, the tools / APIs invoked by the Agent used by this system are: The Agent has built-in and schedules the following tool components: local media asset CMS database query tool, online search plugin (for obtaining authoritative external information), local private ASR speech-to-text tool, online image recognition interface, local privately deployed large language model, preset rule engine and tag whitelist library, and media asset knowledge graph query interface.

[0044] In some embodiments, the Agent automatically determines the media asset type and processing path according to the following fixed decision-making process: 1. First, search the local media asset library + online search. If complete metadata can be obtained, it is determined to be a popular and searchable media asset. 2. If the metadata is incomplete, but the video contains a valid subtitle file, it is determined that the video has subtitle file media assets; 3. If no information can be found or there are no valid subtitles, it is judged as an obscure / self-made media resource without reference information.

[0045] In some embodiments, the system adopts a high-priority data acceptance + multi-source data fusion strategy, with the specific logic as follows: 1. Data acquisition priority: Local media asset library > Online search > Subtitle NLP analysis > ASR speech recognition > Image recognition; 2. Handling conflicts within the same field: When multiple data sources have inconsistent results for the same field (such as director or type), the result with the highest confidence level is selected based on the weighted confidence level, and deduplication is performed automatically; 3. Complementary information processing: Complementary information from different sources (such as plot supplements, actor expansions, and tag expansions) is directly merged, deduplicated, normalized, and then uniformly stored in the database; 4. Final output: A set of standardized label data that is the most complete, accurate, and conflict-free.

[0046] In some embodiments, the key plot timestamps do not rely on manually labeled training data and adopt an unsupervised mode overall. The specific implementation logic is as follows: 1. First, the entire audio is transcribed using the ASR speech transcription module. Then, the text semantic analysis module analyzes the dialogue and narration text, extracts conflict keywords and dramatic sentences, and scores the highlights of the plot segments. 2. Call the image recognition module to identify the features of the scene, set, and props, and at the same time combine the character and scene recognition modules to verify the character features and scene attributes, and filter out invalid noise segments; 3. Based on the frequency of camera cuts, the sudden changes in dialogue density in ASR transcribed text, and the peak values ​​of the emotion curve obtained from large-scale semantic analysis, combined with a preset plot structure template matching algorithm, key plot nodes are accurately located. 4. Using the scene transition start point determined by visual recognition as the start time of the timestamp, and the end point of the plot conflict reversal and the scene ending as the end time, a standardized second-level key plot timestamp is generated. 5. For documentary media assets without obvious dramatic conflicts, such as documentaries, urban humanities documentaries, operas, and sports events, basic timestamps are automatically generated based on visual shot switching detection and scene recognition results, combined with narrative rhythm rules, to achieve full coverage of plots for all types of media assets.

[0047] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An AI-based tag generation and completion system based on media assets, characterized in that, include: The multimodal intelligent recognition engine is used to receive media assets to be processed and, based on the determination result of the completeness of the media asset information, selectively call one or more processing methods from the network search plugin, subtitle text analysis component, ASR speech transcription component or image recognition component to generate multi-source structured description information. The standardized tag generation module is connected to the multimodal intelligent recognition engine and is used to generate tag data containing basic metadata, three-level classification tags and key plot timestamps according to the multi-source structured description information and the preset media asset tag standard specifications. The three-level classification tags include first-level mandatory single tags, second-level mandatory multiple tags and third-level conditional completion tags. The dynamic completion and verification module is connected to the standardized label generation module. It is used to perform field integrity verification, format verification and automatic correction on the label data, and automatically switch to other recognition methods when a single data source fails, and periodically synchronize the latest Internet information to update existing labels. The business adaptation interface module communicates with the dynamic completion and verification module and is used to synchronize the verified tag data to the media asset library and downstream business systems through a standardized data interface.

2. The AI ​​tag generation and completion system according to claim 1, characterized in that, The multimodal intelligent recognition engine is an intelligent engine based on an Agent architecture. The Agent architecture includes a local media asset CMS database query tool, an online search plugin, a locally privately deployed ASR speech transcription tool, an online image recognition interface, a locally privately deployed large language model, a preset rule engine and tag whitelist library, and a media asset knowledge graph query interface.

3. The AI ​​tag generation and completion system according to claim 1, characterized in that, The multimodal intelligent recognition engine determines the processing path of media assets according to the following decision-making process: The first step is to search the local media asset library and combine it with online search. If complete metadata can be obtained, it is determined to be a popular and searchable media asset, and online search is initiated for processing. The second step is to determine if the metadata is incomplete but the media asset contains a valid subtitle file. If so, the media asset is determined to have a subtitle file and the subtitle text analysis method is started for processing. The third step involves determining if the metadata is incomplete and there is no valid subtitle file. If so, it is considered a niche or self-made media asset without reference information, and ASR speech transcription and image recognition are initiated for fusion processing.

4. The AI ​​tag generation and completion system according to claim 3, characterized in that, When the media assets to be processed meet multiple conditions simultaneously, the multimodal intelligent recognition engine adopts a priority-based acceptance and multi-source data fusion strategy. It acquires data according to the priority order of local media asset CMS library, online search information, subtitle text analysis results, ASR speech transcription results, and image recognition results. For cases where the same field is inconsistent in different data sources, it performs a confidence-weighted decision based on the weights of official information over encyclopedic information, encyclopedic information over text extraction information, and text extraction information over image recognition information. It also merges and deduplicates complementary information from different sources before uniformly storing it in the database.

5. The AI ​​tag generation and completion system according to claim 1, characterized in that, The extraction of key plot timestamps adopts an unsupervised rule-driven approach, including: For narrative media assets, the entire audio is transcribed using ASR speech transcription. Key plot points are located based on at least one of the following methods: dialogue density mutation points, emotional curve peaks, and scene switching frequency. The start time of the scene switching is used as the start time of the timestamp, and the end time of the plot conflict or scene ending is used as the end time, generating second-level key plot timestamps. For documentary media assets, basic timestamps are automatically generated based on visual shot switching detection and scene recognition results, combined with narrative rhythm rules.

6. The AI ​​tag generation and completion system according to claim 5, characterized in that, The types of key plot timestamps include at least one or more of the following: plot turning points, plot climaxes, character growth segments, iconic scenes, and emotional highlights.

7. The AI ​​tag generation and completion system according to claim 1, characterized in that, The dynamic completion and verification module includes: The integrity verification unit is used to perform integrity checks on required fields and, when missing fields are found, to perform intelligent completion according to the priority of local media asset library, online search plugin, subtitle text extraction, and media asset knowledge graph matching. The format normalization unit is used to unify time-type fields into a predetermined standard format, regularize region-type fields into a predetermined structure, and automatically remove special characters and redundant prefixes from label fields, as well as perform deduplication and sorting. The multi-source conflict resolution unit is used to make a confidence-weighted resolution based on the weights of official information, encyclopedic information, text-extracted information, and image recognition information when information from multiple data sources is inconsistent. The dynamic update unit is used to periodically connect to the network for incremental synchronization of external information and automatically trigger a second review of all fields after each update.

8. A method for generating and completing AI tags based on media assets, applied to an AI tag generation and completion system based on media assets as described in any one of claims 1 to 7, characterized in that, Includes the following steps: S1: Receive media assets to be processed, and based on the determination result of the completeness of the media asset information, selectively perform one or more processing methods among network search, subtitle text analysis, ASR speech transcription or image recognition to generate multi-source structured description information; S2: Based on the multi-source structured description information and the preset media asset tag standard specifications, generate tag data containing basic metadata, three-level classification tags and key plot timestamps, wherein the three-level classification tags include first-level mandatory single tags, second-level mandatory multiple tags and third-level conditional completion tags; S3: Performs field integrity verification, format verification, and automatic correction on the generated tag data. When a data source fails, it automatically switches to other recognition methods and regularly updates existing tags with the latest information from the Internet. S4: Output the verified tag data to the media asset library or downstream business system through a standardized data interface.

9. The AI ​​tag generation and completion method according to claim 8, characterized in that, The specific steps in step S1, which involve selecting a processing method based on the determination of the completeness of media asset information, include: If complete metadata can be obtained, it is determined to be a popular and searchable media asset, and an online search is initiated. If the metadata is incomplete but a valid subtitle file is available, it is determined to be media asset with a subtitle file, and the subtitle text analysis method is initiated. If the metadata is incomplete and there is no valid subtitle file, it is determined to be an obscure or unreferenced media resource, and ASR speech transcription and image recognition are initiated for fusion processing.

10. The AI ​​tag generation and completion method according to claim 8, characterized in that, Step S1 further includes: when the media asset to be processed meets multiple conditions at the same time, data is acquired according to the priority order of local media asset library, online search information, subtitle text analysis results, ASR speech transcription results, and image recognition results; for the case where the same field is inconsistent in different data sources, a confidence weighted decision is made according to the weight of official information over encyclopedic information, encyclopedic information over text extraction information, and text extraction information over image recognition information, and complementary information from different sources is merged and deduplicated.

Images