Film and television content label processing method and terminal based on large model

Abstract

The invention discloses a film and television content label processing method and terminal based on a large model, and belongs to the technical field of artificial intelligence and film and television content processing, and the method comprises the steps: pre-integrating to form a multi-source film and television data set; performing fine adjustment on a plurality of specified artificial intelligence large models in advance; pre-deploying a task automatic triggering module and constructing a cue word engineering library; acquiring input film and television content, automatically selecting unrechecked content for the film and television content to trigger a task, calling corresponding cue words through the cue word engineering library, and respectively outputting error correction results of film and television content labels through each specified artificial intelligence large model after fine tuning; judging and correcting the error correction result of the output film and television content label by adopting a preset artificial intelligence model, and mapping the correction result to a label library; and after AI auditing is carried out, the identification recheck authentication state of the film and television content is updated. The accuracy, efficiency and diversity of film and television label processing are improved, and the requirements of scenes such as film and television content management, recommendation and analysis are met.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and film and television content processing technology, and in particular to a method, apparatus, server and storage medium for film and television content tagging based on a large model. Background Technology

[0002] With the rapid development of internet technology, the way film and television content is viewed has shifted from traditional television to streaming media platforms, and users have increasingly higher requirements for the accuracy of film and television content retrieval and recommendation. In the existing film and television content management and application technologies, tags, as a core element of content classification and retrieval, directly affect user experience and platform operational efficiency due to their accuracy and richness. However, traditional film and television tag processing methods have significant shortcomings: First, in terms of tag error correction, existing systems mainly rely on manual review or simple rule matching, making it difficult to identify tag errors in complex semantic scenarios, such as incorrectly tagging a suspense film as a romance film; second, in the tag generation stage, traditional methods often use keyword extraction or small-scale machine learning models, resulting in tags that lack semantic depth and diversity, failing to accurately reflect the rich characteristics of film and television content; third, in terms of tag mapping, the conversion efficiency between tag systems of different platforms or data sources is low, and manually formulated mapping rules are difficult to adapt to the dynamic changes and massive growth of film and television data. These problems lead to low efficiency in film and television content management, low user retrieval accuracy, and seriously restrict the intelligent development of film and television platforms.

[0003] Therefore, existing technologies still need improvement and development. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method, apparatus, server, and storage medium for film and television content tag processing based on a large model. It also provides a method for error correction, generation, and mapping of film and television content tags based on a large model. This invention improves the accuracy, efficiency, and diversity of film and television tag processing, meeting the needs of scenarios such as film and television content management, recommendation, and analysis. It has the advantages of improving the accuracy of film and television tag error correction, achieving automated tag correction and mapping, and enhancing the efficiency of film and television content management.

[0005] This application provides a method for processing film and television content tags based on a large model, and the technical solution is as follows: A method for processing film and television content tags based on a large model, comprising: A. Pre-retrieve relevant film information from film and television media asset databases and integrate them to form a multi-source film and television dataset; B. Use the film and television tag data in the multi-source film and television dataset to fine-tune multiple specified artificial intelligence models in advance, and optimize the performance of each specified artificial intelligence model on the film and television tagging task. C. A pre-deployed task automatic triggering module is used to automatically select unreviewed content from the film and television content pool to trigger tasks, and to build a prompt word engineering library for calling corresponding prompt words for different tag tasks; D. Obtain the input film and television content, automatically select unchecked content to trigger the task through the deployed task automatic trigger module, call the corresponding prompt words through the prompt word engineering library, and then output the error correction results of the film and television content tags through the fine-tuned specified artificial intelligence models respectively. D. Use a pre-defined artificial intelligence model to evaluate and correct the error correction results of the output film and television content tags, and map the correction results to the tag library; E. Perform AI review on the mapped tags, and update the identification and re-certification status of the film and television content after approval.

[0006] The aforementioned film and television content tagging method based on a large model further includes, before the step of pre-retrieving film-related information from a film and television media asset library and integrating it to form a multi-source film and television dataset: GPT, BERT, Qwen, and DEEPSEEK R1 models are pre-stacked to form a multi-model collaborative architecture.

[0007] The aforementioned film and television content tagging method based on a large model includes the step of pre-retrieving film-related information from a film and television media asset library and integrating it to form a multi-source film and television dataset, which includes: In advance, basic information about the film is retrieved from the film and television media asset database, and user comments, ratings, social media popularity data, and relevant data from authoritative film and television lists and news reports are obtained using web crawlers; The acquired basic information about the film, as well as corresponding user comments, ratings, social media popularity data, and relevant data from authoritative film and television lists and news reports, are processed through data washing, and unstructured plot summaries and comments are converted into structured data. The cleaned and transformed data is integrated into multi-source film and television data, forming a multi-source film and television dataset that includes film and television content features, user feedback, and external evaluations, and is divided into training set, validation set, and test set according to a predetermined ratio.

[0008] The aforementioned method for processing film and television content tags based on large models includes the following steps: pre-tuning multiple specified large AI models using film and television tag data from the multi-source film and television dataset to optimize the performance of each specified large AI model on the film and television tagging task. The film and television tag data in the training set of the multi-source film and television dataset is used in advance to fine-tune multiple specified large artificial intelligence models and optimize the performance of each specified large artificial intelligence model on the film and television tag task. The multiple specified large artificial intelligence models include: GPT, BERT and Qwen models. The film and television tag data includes: incorrect and correct label pairs and reference data for generating labels of unlabeled films and television shows. The DEEPSEEK R1 model is pre-tested using the validation set of the multi-source film and television dataset, and the parameters are adjusted based on the validation set to accurately evaluate the output results of the multi-model.

[0009] The aforementioned large-scale model-based film and television content tagging method, wherein the pre-deployed task automatic triggering module is used to automatically select unreviewed content from the film and television content pool to trigger tasks, and to construct a prompt word engineering library for calling corresponding prompt words for different tag tasks, includes the following steps: A pre-deployed task automatic triggering module is used to sort the content in the film and television content pool by popularity and automatically select unreviewed content to trigger tasks. A pre-built prompt word library is used to call the corresponding prompt words for different tag tasks.

[0010] The aforementioned method for processing film and television content tags based on large models includes the following steps: acquiring input film and television content, automatically selecting unchecked content to trigger tasks through a deployed task automatic triggering module, calling corresponding prompt words through a prompt word library, and then outputting the error correction results of film and television content tags through fine-tuned designated artificial intelligence large models. Get the input video content and extract the original tags of the input video content; The deployed task automatic triggering module automatically selects unreviewed content from film and television content to trigger tasks, and identifies plot elements in the unreviewed content. Based on the identified plot elements, determine whether they conflict with the original tags of the film and television content; if a conflict is found, call the corresponding conflict warning words from the warning word library. Based on the conflict warning words, the error correction results of the film and television content tags are output by each of the specified AI models after fine-tuning.

[0011] The aforementioned method for processing film and television content tags based on a large model includes the following steps: evaluating and correcting the error correction results of the output film and television content tags using a predetermined artificial intelligence model, and mapping the corrected results to a tag library. The DEEPSEEK R1 model, a pre-defined artificial intelligence model, is used to evaluate the error correction results of the output film and television content tags, and the tags corresponding to the identified plot elements are selected as correction tags. The function of generating tags and mapping tag libraries is invoked from the pre-built prompt word engineering library to map the corrected tags to the standard tags of the corresponding film and television content in the film and television content tag library.

[0012] A film and television content tagging processing device based on a large model, wherein the device includes: The multi-source integration module is used to retrieve relevant film information from the film and television media asset library in advance and integrate it to form a multi-source film and television dataset; The model fine-tuning module is used to fine-tune multiple specified large artificial intelligence models in advance using the film and television tag data in the multi-source film and television dataset, and optimize the performance of each specified large artificial intelligence model on the film and television tagging task. The task triggering and prompt word module is used to pre-deploy the automatic task triggering module, which is used to automatically select unchecked content from the film and television content pool to trigger tasks, and to build a prompt word engineering library for calling corresponding prompt words for different tag tasks; The tag correction module is used to acquire input film and television content. The deployed task automatic triggering module automatically selects unchecked content from the film and television content to trigger the task, and calls the corresponding prompt words through the prompt word engineering library. Then, the fine-tuned specified artificial intelligence models output the error correction results of the film and television content tags respectively. The tag mapping module is used to evaluate and correct the error correction results of the output film and television content tags using a pre-defined artificial intelligence model, and then map the correction results to the tag library. The AI ​​review module is used to review the mapped tags using AI, and update the identification and re-certification status of the film and television content after it passes the review.

[0013] A server includes a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors, the one or more programs comprising the method for performing any one of the methods.

[0014] A computer-readable storage medium, wherein, when instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform any of the methods described above.

[0015] As can be seen from the above, the film and television content tagging method, device, server and computer-readable storage medium provided in this application, by integrating multi-source data, fine-tuning the multi-model collaborative architecture, automatically triggering tasks and prompt word engineering, and combining AI judgment and correction mechanisms, solves the problems of low accuracy and poor efficiency in traditional tagging. It has the advantages of improving the accuracy of film and television tag error correction, realizing automated tag correction and mapping, and improving the efficiency of film and television content management.

[0016] This invention integrates multiple large language models (such as GPT, BERT, Qwen, DEEPSEEK R1, etc.) and combines prompt word engineering with an automated task triggering mechanism to achieve efficient and accurate processing of film and television content tags; furthermore, the embodiments of this invention also have the following advantages: 1) Improved label accuracy: Through multi-model evaluation and error correction, the label error rate is reduced. For example, the accuracy rate is improved by 80% in the label classification of complex plot films and television shows.

[0017] 2) Improved tag processing efficiency: Automated triggering and streamlined processing improve efficiency by 10 times compared to manual processing, meeting the real-time tagging needs of massive amounts of film and television content.

[0018] 3) Enhanced tag diversity: Generates rich personalized tags (such as "cyberpunk action film", "box office over 100 million", "blockbuster", "explosive opening", "must-see in a lifetime", "millions in revenue sharing"), improving the performance of film and television content in recommendation, search and other scenarios. Attached Figure Description

[0019] 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 recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the film and television content tagging method based on a large model provided in Embodiment 1 of the present invention.

[0021] Figure 2 This is a flowchart illustrating the film and television content tagging method based on a large model provided in Embodiment 2 of the present invention.

[0022] Figure 3 The principle block diagram of an embodiment of the film and television content tagging processing device based on a large model provided by the present invention.

[0023] Figure 4 This is a block diagram illustrating the internal structure of the server provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0025] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0026] In existing technologies, film and television content tagging primarily relies on manual review or rule-based systems, which suffer from insufficient semantic understanding and low processing efficiency. For example, an online film and television platform experienced search failures due to tagging errors; a science fiction film was incorrectly labeled as a romance film, preventing users from finding their target content through tag filtering. Traditional methods struggle to identify science fiction elements implicit in plot synopses and cannot update outdated tagging systems in a timely manner.

[0027] To address these issues, the R&D team discovered that existing single models could not cover complex semantic scenarios, and insufficient integration of multi-source data limited tag quality. Analysis of user complaint cases revealed that many erroneous tags stemmed from misidentification of plot elements. Attempts to use pre-trained models to handle this issue encountered domain adaptation problems, leading them to consider leveraging multi-model collaboration to improve generalization capabilities. To address the need for dynamic updates, they devised an automated task-triggered mechanism for real-time error correction.

[0028] Therefore, this application proposes a method for processing film and television content tags based on a large model, as described in the following embodiments: Example 1 like Figure 1 As shown in Embodiment 1 of the present invention, a method for processing film and television content tags based on a large model includes the following steps: Step S100: First, retrieve relevant film information from the film and television media asset database and integrate it to form a multi-source film and television dataset; In this embodiment of the invention, retrieving film-related information from the film and television media asset library includes retrieving basic film information from the film and television media asset library, including film name, director, actors, plot synopsis, original tag information, etc.; and also includes user comments, ratings, social media popularity data related to the film obtained by web crawling, as well as relevant data from authoritative film and television lists and news reports; then integrating the above multi-source data to form a dataset containing film and television content characteristics, user feedback, and external evaluations, which can be divided into training set, validation set, and test set in a [7:2:1] ratio.

[0029] Step S200: Use the film and television tag data in the multi-source film and television dataset in advance to fine-tune multiple specified artificial intelligence models and optimize the performance of each specified artificial intelligence model on the film and television tag task; The multiple specified large artificial intelligence models mentioned in the embodiments of the present invention may include GPT (General Text Generation Model), BERT (Text Understanding Model), Qwen (QWen) model, and DEEPSEEK R1 model; In this embodiment of the application, the model fine-tuning step includes: fine-tuning the GPT, BERT, and Qwen models using film and television label data (such as incorrect and correct label pairs, and generated label references for unlabeled films and televisions) in the training set to optimize their performance on film and television labeling tasks.

[0030] For the DEEPSEEK R1 model, the parameters are adjusted based on the validation set to enable it to accurately evaluate the output results of multiple models.

[0031] Step S300: Pre-deploy the automatic task triggering module, which is used to automatically select unreviewed content from the film and television content pool to trigger tasks, and build a prompt word engineering library for calling corresponding prompt words for different tag tasks; In this embodiment, an automatic task triggering module is pre-deployed and a popularity threshold is set (e.g., the number of views exceeds [X] times). The content in the film and television content pool is sorted by popularity, and unreviewed content is automatically selected to trigger the task. A prompt word engineering library is also built to call the corresponding prompt words for different tag tasks (e.g., calling "original tag error correction prompt words" for films and television shows with complex plots).

[0032] Step S400: Obtain the input film and television content, automatically select unchecked content to trigger the task through the deployed task automatic triggering module, call the corresponding prompt words through the prompt word engineering library, and then output the error correction results of the film and television content tags through the fine-tuned specified artificial intelligence models respectively. In this step of the embodiment, for example, the input film and television content is obtained (such as a science fiction war film, the original label is "romance film"). After the task is triggered, the prompt word "The label 'romance film' conflicts with the war and science fiction elements in the plot synopsis. Please correct it" is called. The error correction results are output by the finely adjusted GPT, BERT and Qwen models respectively (such as "science fiction war film", "war science fiction", "science fiction action war film").

[0033] Step S500: Use a predetermined artificial intelligence model to evaluate and correct the error correction results of the output film and television content tags, and map the correction results to the tag library; In this embodiment, regarding result evaluation and mapping, the above results can be evaluated using a fine-tuned DEEPSEEK R1 model, selecting "science fiction war film" as the correction tag. Simultaneously, the "generate tag and tag library mapping prompt" is invoked to map "science fiction war film" to the "science fiction war category" standard tag in the tag library.

[0034] Step S600: Perform AI review on the mapped tags, and update the identification and re-certification status of the film and television content after approval.

[0035] In this embodiment of the application, regarding AI review and updating, the mapped tags are reviewed by AI (such as checking whether they conform to the tag library rules and whether the semantics are contradictory). After passing the review, the identification re-review and certification status of the film and television content is updated, thus completing the entire tag processing process.

[0036] The above-described embodiments of this application propose a technical solution that pre-integrates film and television media asset libraries and external data to form a multi-source dataset, fine-tunes multiple large language models to optimize tag processing performance, deploys an automatic triggering module to select unreviewed content, constructs a prompt word engineering library for targeted calling model, outputs error correction results through multi-model collaboration, and updates the tag status after AI review.

[0037] The multi-source film and television dataset refers to a structured data collection that integrates basic film information, user reviews, and authoritative rankings. Specifically, it can be achieved by using web crawlers to obtain social media data and then cleaning the unstructured plot summaries to convert them into structured fields, providing multi-dimensional features for model training. Fine-tuning the large AI model refers to domain-adaptive optimization of the pre-trained model. This can be achieved using a multi-task learning framework, adding comparison samples of incorrect and correct labels to the training set to improve the model's ability to identify label inconsistencies. The automatic task triggering module is an automated processing mechanism based on content popularity. Specifically, it can be achieved by setting a playback threshold ranking algorithm, automatically initiating a review process when the playback volume of film and television content exceeds a predetermined value. The prompt word library is a set of instructions categorized by label type. Specifically, it can be achieved using natural language template generation technology to design dedicated prompt statements for different tasks such as emotion tags and theme tags.

[0038] Specifically, the system first extracts basic data from a film and television database, and constructs a training set by combining crawled comments and ranking information. Models such as GPT and BERT are fine-tuned to recognize the plausibility of the co-occurrence of "science fiction" and "romance" elements. When a newly released film reaches 100,000 views, the module automatically triggers to select that content for detection. A prompt word library calls "identify genre conflicts" commands, driving multiple models to analyze the plot text. After each model outputs correction suggestions, a review model selects the optimal solution, ultimately updating the science fiction genre identifier in the tag library.

[0039] Compared with existing technologies, this application breaks through the bottleneck of single-model processing and improves the recognition accuracy of complex scenes through multi-model collaboration. Traditional methods rely on fixed rules to update tags, while this application implements a dynamic triggering mechanism based on content popularity. Existing technologies require manual construction of mapping rules, while this application achieves intelligent tag system adaptation through prompt word engineering.

[0040] Through the above technical solutions, this application effectively solves the problem of semantic bias in tags, such as accurately identifying the dominant elements in films with mixed themes. It also improves the timeliness of tag updates, completing a re-examination within 72 hours after film and television content reaches a popularity threshold. Furthermore, it enhances the compatibility of the tag system by automatically mapping to adapt to the tag specifications of different platforms.

[0041] This application further proposes to pre-stack GPT, BERT, Qwen, and DEEPSEEK R1 models to form a multi-model collaborative architecture.

[0042] Stacking refers to combining multiple models with different architectures, which can be achieved through model parallelism or result fusion, such as by weighted integration of the outputs of each model through an API interface. Multi-model collaborative architecture refers to leveraging the complementary strengths of different models; for example, GPT excels at generating tasks, BERT excels at semantic understanding, Qwen supports long Chinese text processing, and DEEPSEEK R1 has evaluation capabilities, thus covering the diverse needs of film and television tagging.

[0043] Specifically, before the film and television tagging process begins, this application stacks the GPT, BERT, Qwen, and DEEPSEEK R1 models to form a multi-model collaborative architecture. The GPT model generates descriptions of plot elements, the BERT model extracts semantic features, the Qwen model handles tag association analysis for long Chinese texts, and the DEEPSEEK R1 model evaluates the outputs of all models. The models interact through an interface layer; for example, the semantic features extracted by BERT are used as input for GPT tag generation, and the association analysis results from Qwen are cross-validated with the evaluation results from DEEPSEEK R1, ultimately forming a unified tag processing decision.

[0044] Compared to existing technologies, traditional film and television tagging methods typically employ a single model for tag generation or error correction, such as using only the BERT model for semantic matching. This leads to incomplete semantic coverage when handling complex scenarios. This application, however, utilizes a multi-model collaborative architecture that leverages the strengths of generative, interpretive, and evaluative models simultaneously. For instance, when processing film and television content containing metaphorical plots, GPT generates potential tag candidates, BERT verifies semantic consistency, and DEEPSEEK R1 arbitrates conflicting tags, thereby enhancing the ability to handle complex scenarios.

[0045] Through the above technical solution, this application solves the problems of incomplete label generation coverage and semantic understanding bias caused by the single model in traditional methods. The multi-model collaborative architecture can handle the generation, understanding, and evaluation needs of film and television content separately by combining different models. For example, when identifying labels for niche films, the Qwen model can parse the implicit features in long texts, while the DEEPSEEK R1 model corrects the results based on authoritative data, thereby improving the accuracy and robustness of label processing.

[0046] This application further proposes to first retrieve basic film information from a film and television media asset database, and to use web crawlers to obtain corresponding user comments, ratings, social media popularity data, and relevant data from authoritative film and television lists and news reports; to clean the obtained basic film information, corresponding user comments, ratings, social media popularity data, and relevant data from authoritative film and television lists and news reports, and to convert unstructured plot summaries and comments into structured data; to integrate the cleaned and transformed data into multi-source film and television data, forming a multi-source film and television dataset that includes film and television content features, user feedback, and external evaluations, and to divide it into training set, validation set, and test set according to a predetermined ratio.

[0047] The film and television media asset library refers to a database system that stores film metadata and related information. It can be implemented using a distributed database or a relational database, and is used to centrally manage basic attributes such as film titles, directors, actors, and release years. Web crawlers are programs that automatically scrape publicly available data from the internet based on preset rules. They can be implemented using the Scrapy framework or the Selenium tool, and are used to obtain dynamic data such as user comments and social media interactions. Data cleaning refers to the process of deduplicating, standardizing, and handling outliers in raw data. It can be implemented using regular expression matching or rule-based filtering algorithms to eliminate data redundancy and noise interference. Structured data transformation refers to converting natural language text into a data format with clearly defined fields. It can be implemented using named entity recognition technology or template-based parsing methods to extract quantifiable features such as plot keywords and emotional tendencies. Multi-source film and television datasets refer to standardized data sets that integrate multi-dimensional information. They can be stored in JSON format or relational table structures to support subsequent model training and validation needs.

[0048] Specifically, basic film information is exported in batches from a film and television media asset library via an API interface. Simultaneously, a web crawler periodically collects user comment data from third-party platforms and social media trending data. During the cleaning process, invalid characters are filtered using regular expressions, duplicate comments are merged using text similarity algorithms, and a sentiment analysis model is used to label the polarity of unstructured comments. In the structuring transformation stage, the plot synopsis is converted into a character relationship graph using entity extraction technology, and the rating data is normalized to a 0-1 range. The integrated dataset is randomly divided into training, validation, and test sets in a 7:2:1 ratio to ensure no content overlap between different subsets.

[0049] Compared to existing technologies, traditional methods primarily rely on static metadata from a single source, lacking dynamic collection of user feedback and external evaluations, thus limiting the dimensions of tag generation. This application integrates multi-source heterogeneous data to construct a comprehensive dataset containing content features, user behavior, and industry evaluations, providing a more comprehensive semantic understanding foundation for model training. While existing technologies often employ manual rule configuration for data cleaning, this application introduces an automated text processing workflow, significantly improving data processing efficiency.

[0050] Through the above technical solution, this application effectively solves the problem of insufficient label accuracy caused by the single data dimension in traditional film and television label processing methods. The fusion of multi-source data enhances the model's ability to understand film and television content from multiple perspectives, the structured transformation process improves the usability of non-textual data, and the dataset partitioning mechanism ensures the scientific nature of model training, providing reliable data support for subsequent label correction and generation tasks.

[0051] This application further proposes to fine-tune several designated large-scale artificial intelligence models in advance using film and television tag data from the training set of a multi-source film and television dataset to optimize their performance on the film and television tagging task. The designated large-scale artificial intelligence models include GPT, BERT, and Qwen models. The film and television tag data includes incorrect and correct label pairs and reference data for generating labels for unlabeled films and television shows. At the same time, the DEEPSEEK R1 model is used with a validation set of the multi-source film and television dataset to adjust the parameters and accurately evaluate the output results of the multiple models.

[0052] The training set of the multi-source film and television dataset refers to a cleaned, structured dataset collected from film and television media asset libraries and external data sources. Specifically, it can be implemented by dividing the data into training, validation, and test sets in a 7:2:1 ratio using a data bucketing method, providing the sample features and label correspondences required for model training. Fine-tuning of the large AI model refers to parameter optimization based on a pre-trained language model, which can be achieved using a cross-entropy loss function combined with a contrastive learning framework to improve the model's ability to identify incorrect patterns in film and television labels. The incorrect and correct label pairs refer to manually labeled error correction datasets, generated through a combination of manual review and semi-automatic annotation tools, used to construct training samples for supervised learning. DEEPSEEK R1 model parameter tuning refers to optimizing the evaluation model based on validation set results, which can be achieved using a grid search combined with a Bayesian optimization algorithm to establish a confidence evaluation mechanism for the output results of multiple models.

[0053] Specifically, by iteratively training the GPT, BERT, and Qwen models using the training set, the models learn the mapping relationship between incorrect and correct labels. For example, samples mislabeled as "fantasy" instead of "science fiction" are used to correct the model's output. A dynamic weight adjustment strategy is employed during training to balance the distribution of samples with different error types. The validation set data is input into the DEEPSEEK R1 model for parameter optimization, such as adjusting the temperature coefficient in the attention mechanism to accurately identify contradictory results from multiple models. After training, the GPT, BERT, and Qwen models output candidate correction labels, and the DEEPSEEK R1 model scores the confidence of these candidates based on the evaluation criteria of the validation set optimization.

[0054] Compared to existing technologies, current film and television tagging methods typically employ a single model for training and lack a validation set optimization mechanism, leading to the model's tendency to overfit noisy labels in the training data. This application reduces the risk of single-model bias through multi-model collaborative training, while utilizing independent validation sets to optimize the parameter settings of the evaluation model, enabling the system to identify potential contradictions in the outputs of different models. Furthermore, the data construction method for incorrect and correct label pairs overcomes the limitations of traditional methods that rely on manual rule matching, allowing the model to learn label error correction patterns under complex semantics.

[0055] Through the above technical solutions, this application solves the problem of insufficient model generalization ability in traditional film and television tag processing methods and improves the ability to distinguish semantically similar tags; the established validation set optimization mechanism effectively avoids the situation of overfitting the evaluation model with the training data and ensures the reliability of the evaluation of the output results of multiple models; the training data construction method of erroneous tag pairs enhances the model's ability to process unlabeled film and television content, enabling the system to automatically generate reference tags that conform to industry standards.

[0056] This application further proposes a pre-deployed task automatic triggering module, and pre-sets a popularity threshold for playback volume to sort the content in the film and television content pool by popularity and automatically select unreviewed content to trigger tasks; and pre-builds a prompt word engineering library to call corresponding prompt words for different tag tasks.

[0057] The automatic task triggering module is a program component that automatically filters content to be processed based on preset rules. Specifically, it can be implemented using a distributed task scheduling framework combined with a popularity calculation algorithm. By monitoring metrics such as play counts and click-through rates of film and television content in real time, it dynamically filters high-popularity content that has not yet been re-checked. This module filters low-value data by setting thresholds and prioritizes high-profile content, thereby improving resource allocation efficiency.

[0058] A prompt word library refers to a dataset that stores standardized instruction templates. Specifically, it can be constructed using graph databases combined with natural language processing techniques. By establishing a mapping relationship between labeled task types and prompt word templates, accurate invocation in different scenarios can be achieved. This library ensures the standardization and task adaptability of large model inputs by structurally storing prompt word templates for scenarios such as conflict detection and semantic correction.

[0059] Specifically, the automatic task triggering module periodically scans the video content pool and calculates popularity scores based on playback statistics. When the popularity score of a video content exceeds a preset threshold, the system automatically marks it as an unreviewed task and adds it to the processing queue. Simultaneously, the prompt word library dynamically matches instruction templates based on the task type. For example, when a conflict is detected between plot elements and original tags, a dedicated prompt word containing steps such as conflict analysis and tag candidate set generation is invoked to drive the large model to perform targeted error correction.

[0060] Compared to existing technologies, traditional methods rely on manual screening of content to be processed and lack prioritization, resulting in delays in processing highly popular content. This application achieves dynamic optimization and allocation of data processing resources through an automated threshold triggering mechanism. Existing technologies often use statically configured prompt words, which cannot adapt to multi-task scenarios. This application, however, constructs a structured prompt word library that can dynamically match conflict resolution templates based on the identification results of plot elements, significantly improving the scenario adaptability of tag correction.

[0061] Through the above technical solution, this application solves the problems of low task triggering efficiency and insufficient adaptability of prompt words in traditional film and television tagging. It automatically filters high-value content using popularity thresholds, avoiding processing delays caused by manual intervention; and through precise invocation of a structured prompt word library, it ensures the standardization of large model input instructions in different conflict scenarios, thereby improving the accuracy and processing efficiency of tag correction results.

[0062] This application further proposes to acquire input film and television content, extract the original tags of the input film and television content; automatically select unchecked content from the deployed task triggering module to trigger tasks, and identify plot elements in the unchecked content; determine whether the identified plot elements conflict with the original tags of the film and television content; if they conflict, call the corresponding conflict prompt words through the prompt word engineering library; and output the error correction results of the film and television content tags through each fine-tuned specified artificial intelligence model based on the conflict prompt words.

[0063] Among them, the "Unreviewed Content Triggering Task" refers to the automatic triggering process for film and television content that has not undergone tag review by the system. This can be implemented by setting playback thresholds or timestamp filtering mechanisms to prioritize high-traffic or newly added content. "Plot Element Recognition" refers to the structured analysis of core plot points, character relationships, and scene features in film and television content. Specifically, it can use natural language processing models to extract key entities and their semantic relationships to detect the consistency between tags and content. "Conflict Hints" refers to instruction templates designed for scenarios where tags and plot elements contradict each other. Specifically, it can generate hint statements containing descriptions of the contradictions and correction requirements based on preset conflict types, guiding the large model to focus on correcting erroneous tags.

[0064] Specifically, when film and television content is input into the system, the original tags are extracted and associated with an automatically triggered queue of unchecked content. A natural language processing model is used to extract entities and analyze relationships within the plot text, generating a semantic graph containing core elements. This graph is then semantically matched with the original tags. If a logical contradiction is detected between a tag and a plot element, the corresponding conflict warning word is invoked based on the type of contradiction. For example, if the tag is "comedy" but the plot element contains numerous violent scenes, the system invokes the "genre tag conflict" warning word, triggering multiple finely tuned large models to generate correction suggestions in parallel. Each model, based on the contradiction description in the conflict warning word and the tag mapping rules in the training data, outputs correction results containing candidate tags such as "action film" and "dark humor."

[0065] Compared to existing technologies, traditional methods rely on manual review or single models for label verification, resulting in low efficiency and susceptibility to subjective judgment. This application prioritizes unreviewed content through an automatic triggering mechanism, establishes objective conflict detection standards by combining plot element recognition technology, and generates diverse correction suggestions using a multi-model collaborative mechanism, effectively solving the label correction bias problem in complex semantic scenarios.

[0066] Through the above technical solutions, this application achieves efficient and automated detection of film and television tag errors, avoiding processing delays caused by manual review; it accurately locates the type of erroneous tag by analyzing the semantic conflict between plot elements and tags; and it utilizes a multi-model parallel processing mechanism to ensure the diversity and accuracy of error correction results, significantly improving the consistency between the tag system and film and television content.

[0067] This application further proposes to use a pre-defined artificial intelligence model, DEEPSEEK R1, to evaluate the error correction results of the output film and television content tags, select the tags corresponding to the identified plot elements as correction tags, and call the tag generation and tag library mapping function of the pre-built prompt word engineering library to map the correction tags to the standard tags of the corresponding film and television content in the film and television content tag library.

[0068] The DEEPSEEK R1 model refers to an evaluation model whose parameters have been adjusted using a validation set of multi-source film and television datasets. Specifically, it can be implemented using a neural network architecture based on an attention mechanism. A multi-task learning framework is used to quantitatively evaluate the semantic matching degree of the tags, selecting candidate tags with the strongest relevance to plot elements from multiple error correction results. The tag generation and tag library mapping prompt function refers to a pre-defined set of standardized mapping rules, specifically implemented using a tag association matrix based on a knowledge graph. Semantic similarity calculations establish a correspondence between temporary corrected tags and the standard tag system, eliminating semantic ambiguity between different tag systems.

[0069] Specifically, when multiple fine-tuned large models output different error correction results, the DEEPSEEK R1 model first scores the semantic matching degree of each result, such as assigning weights to the tags "science fiction" and "suspense" in plots containing alien creature elements. During the scoring process, the model focuses on analyzing the logical relationship between the tags and the identified plot elements; for example, it prioritizes retaining the "science fiction" tag when the plot contains time loop scenes. Subsequently, it calls the mapping function through the prompt word engineering library, for example, mapping the user-generated "mind-bending" to "suspenseful reasoning" in the standard tag library. This process automatically completes the tag standardization process through preset semantic expansion rules.

[0070] Compared with existing technologies, traditional tag mapping methods rely on manually building static mapping tables, which makes it difficult to handle the semantic expansion of newly generated tags. This application combines a dynamic prompt word engineering library with knowledge graph technology to automatically identify potential associations between temporary tags and the standard system. For example, it can automatically associate the non-standard "dark style" with the dual tag combination of "crime theme" and "violent elements" in the standard tag library, realizing intelligent mapping of cross-dimensional tag systems.

[0071] Through the above technical solution, this application effectively solves the technical problem of low tag mapping efficiency in dynamic film and television data environments. By automatically selecting the optimal corrected tags and completing standardized mapping through an AI model, the processing delays and mapping errors caused by manual intervention are avoided. For example, when batch processing user-generated tags, the system can complete the automatic conversion of thousands of non-standard tags in milliseconds, significantly improving the update efficiency and semantic consistency of the film and television content tag library.

[0072] The present invention will be further described in detail below through another specific application embodiment.

[0073] Example 2 like Figure 2 The diagram shown is a technical architecture diagram of the film and television content tagging method based on a large model provided in the second specific application embodiment of this invention. Figure 2 The document demonstrates the entire process from triggering a task from a film and television content pool, through prompt word engineering, large model stacking processing, tag engineering processing, data processing and preparation, to finally completing AI review and tag re-verification. The interaction relationships between the various modules (automatic task triggering, prompt word engineering, large model stacking, tag engineering processing, data preparation, and data processing) are clearly presented, reflecting the technical logic and innovative architecture of this invention.

[0074] like Figure 2 As shown, the automatic task trigger is used to build a "film and television content pool," prioritizing content that has not undergone re-verification based on popularity, and automatically triggering tag correction, generation, and mapping tasks. For example, newly released film and television content with high viewership is processed first to ensure the timeliness and accuracy of popular content tags.

[0075] The prompt word engineering module is used to design various types of prompt words, including: original label error correction prompt words, classification target label generation prompt words, multi-result evaluation prompt words, and label generation and label library mapping prompt words.

[0076] The original tag correction prompt is used to guide the model to analyze semantic contradictions or classification errors in the original tag (such as "The movie tag 'romance' conflicts with the war theme in the plot synopsis, please correct it").

[0077] The classification target label generation prompt words are used to generate standard classification labels based on film and television information (such as "generate type labels based on 'director, actors, plot synopsis'").

[0078] The multi-result evaluation prompts are used to assist the model in comparing and evaluating the output results of multiple large models (such as "compare the labels generated by GPT and BERT, and select the result with the most accurate semantics and in line with the classification standards of the film and television industry").

[0079] The generated tag and tag library mapping prompt is used to map the generated tag to the standard tag library (e.g., "map 'mind-bending drama' to 'suspense drama' in the tag library").

[0080] like Figure 2 As shown, the large model stacking process includes: The error correction and generation model integrates models such as GPT, BERT, and Qwen, leveraging their respective advantages (such as GPT's generation flexibility and BERT's bidirectional semantic understanding) to correct and generate film and television tags.

[0081] The result evaluation model is used to comprehensively evaluate the output results of multiple models using DEEPSEEK R1 to ensure the accuracy and consistency of the labels.

[0082] Model fine-tuning is used to fine-tune the above model using labeled data in the film and television field (including film titles, directors and actors, plot summaries, user reviews, authoritative rankings, etc.) to make it more suitable for film and television labeling tasks.

[0083] like Figure 2 As shown, the label engineering process includes: summarizing the processing results of each model (GPT processing results, BERT processing results, Qwen processing results), performing label correction (deleting erroneous labels on the content) and mapping (mapping the generated labels to the standard label pool labels) after evaluation by DEEPSEEK R1, and finally affixing the mapped labels to the content.

[0084] like Figure 2 As shown, data processing and preparation include: Data preparation: Obtain raw film and television data (film title, director and actors, genre / release date, plot synopsis, tags, etc.) through media asset database retrieval, and obtain user-generated data (film reviews, ratings, interactions) and external data (authoritative rankings, news reports, popularity data, box office data) using web crawlers.

[0085] Data processing: The acquired data is cleaned (noise and duplicate data are removed), transformed (unstructured data is structured), and integrated to form a data training set for model fine-tuning and training.

[0086] AI review and re-inspection: The processed tags are reviewed by AI, and the content identifier re-inspection and certification status is updated after passing the review, ensuring the reliability of the entire tag processing process.

[0087] Specifically, refer to Figure 2 As shown in the figure, the second embodiment of this paper provides a method for processing film and television content tags based on a large model, which includes the following steps: S10, Data Acquisition Steps: Retrieve basic film information from the film and television media asset database; then proceed to S11; The basic information of the film includes: film title, director, actors, plot synopsis, original tags, etc.

[0088] S11. Use web crawlers to obtain user reviews, ratings, social media popularity data, and relevant data from authoritative film and television rankings and news reports; then proceed to S12. S12. Clean and transform the data, removing duplicate and erroneous comment data, and converting unstructured plot summaries and comments into structured data (such as extracting keywords and sentiment); then proceed to S13. S13. Integrate multi-source data to form a dataset containing film and television content features, user feedback, and external evaluations; that is, integrate original film and television data, user-generated data, and external data to form a dataset containing film and television content features, user feedback, and external evaluations; divide the dataset into training set, validation set, and test set in a ratio of [7:2:1]; then proceed to S14. S14. Model fine-tuning: Fine-tune the GPT, BERT, and Qwen models using film and television label data from the training set (such as incorrect and correct label pairs, and generated label references for unlabeled films and televisions) to optimize their performance on film and television labeling tasks; then proceed to S15. S15. For the DEEPSEEK R1 model, adjust the parameters based on the validation set to enable it to accurately evaluate the output results of multiple models; then proceed to S16. S16. Deploy the automatic task triggering module, set the popularity threshold (e.g., the number of views exceeds [X] times), sort the content in the film and television content pool by popularity, automatically select unreviewed content to trigger the task; and build a prompt word engineering library to call the corresponding prompt words for different tag tasks (e.g., call "original tag error correction prompt words" for films and television shows with complex plots); then proceed to S17. S17. Input film / TV content (e.g., a science fiction war film, originally labeled "romance"). After triggering the task, the prompt message "The label 'romance' conflicts with the war and science fiction elements in the plot synopsis. Please correct it" is displayed. The GPT, BERT, and Qwen models output the correction results respectively (e.g., "science fiction war film", "war science fiction", "science fiction action war film"). Then proceed to S18. S18, Result Evaluation and Mapping Steps: The results are evaluated using the DEEPSEEK R1 model, and "Science Fiction War Film" is selected as the correction tag. At the same time, "Generate Tag and Tag Library Mapping Prompt" is called to map "Science Fiction War Film" to the "Science Fiction War Category" standard tag in the tag library. Then proceed to S19. S19, AI review and update steps: AI review is performed on the mapped tags (such as checking whether they conform to the tag library rules and whether the semantics are contradictory). After passing the review, the identification re-inspection and certification status of the film and television content is updated to complete the entire tag processing process.

[0089] As can be seen from the above, the specific application embodiment of this application provides a film and television content tagging method based on a large model. This invention integrates original film and television data, user-generated data, and external data to enrich model input, enhance the generalization ability of tag processing, and realize the utilization of multi-source data. Furthermore, this invention creatively stacks models such as GPT, BERT, Qwen, and DEEPSEEK R1, combining the advantages of each model to achieve accuracy and diversity in tag processing, forming a multi-model collaborative architecture. Moreover, it adopts automatic task triggering, seamless integration of prompt word engineering and model processing, realizing automated and intelligent processes, improving tag processing efficiency, and reducing manual intervention. This invention also has the following advantages: 1) Improve label accuracy: Reduce label error rate through multi-model evaluation and error correction. For example, the accuracy rate is improved by 80% in the label classification of complex plot films and television shows.

[0090] 2) Improved processing efficiency: Automated triggering and streamlined processing improve efficiency by 10 times compared to manual processing, meeting the real-time tagging needs of massive amounts of film and television content.

[0091] 3) Enhance tag diversity: Generate rich personalized tags (such as "cyberpunk action movie", "box office over 100 million", "blockbuster", "explosive opening", "must-see in a lifetime", "millions in revenue sharing") to improve the performance of film and television content in recommendation, search and other scenarios.

[0092] Exemplary device like Figure 3 As shown, this embodiment of the invention provides a film and television content tagging processing device based on a large model. The device includes: The multi-source integration module 310 is used to retrieve relevant film information from the film and television media asset library in advance and integrate it to form a multi-source film and television dataset. The model fine-tuning module 320 is used to fine-tune multiple specified large artificial intelligence models in advance using the film and television tag data in the multi-source film and television dataset, and optimize the performance of each specified large artificial intelligence model on the film and television tagging task. The task triggering and prompt word module 330 is used to pre-deploy the automatic task triggering module, which is used to automatically select unchecked content from the film and television content pool to trigger tasks, and to build a prompt word engineering library for calling corresponding prompt words for different tag tasks; The tag correction module 340 is used to acquire the input film and television content. The deployed task automatic triggering module automatically selects the unchecked content of the film and television content to trigger the task, and calls the corresponding prompt words through the prompt word engineering library. Then, the fine-tuned specified artificial intelligence models output the error correction results of the film and television content tags respectively. The tag mapping module 350 is used to evaluate and correct the error correction results of the output film and television content tags using a pre-defined artificial intelligence model, and map the correction results to the tag library. The AI ​​review module 360 ​​is used to review the mapped tags using AI, and update the identification and re-certification status of the film and television content after it passes the review.

[0093] The multi-source integration module is a processing unit that integrates basic film and television information, user comments, and external data into multi-source data through structured processing. Specifically, it can be implemented using a distributed web crawler framework and data cleaning algorithms. For example, the Scrapy framework can be used to collect social media data and extract structured fields using regular expressions. Its purpose is to solve the problem of single data sources in traditional methods. The model fine-tuning module is a functional unit that optimizes the parameters of multiple large models based on the training set. Specifically, it can be implemented using LoRA lightweight fine-tuning technology. For example, inserting a low-rank adaptation layer into the BERT model can improve label classification accuracy. Its purpose is to enhance the model's ability to recognize complex semantic labels. The task triggering and prompt word module is an execution unit that automatically filters unchecked content based on a popularity threshold and matches prompt words. Specifically, it can use a sliding window algorithm to calculate the content popularity value in real time and match preset prompt word templates through vector retrieval. Its purpose is to improve the triggering efficiency of label processing tasks. The label error correction module is a computational unit that calls multiple models in parallel to generate error correction results. Specifically, it can use a GPU cluster to achieve load balancing of multi-model inference tasks. Its purpose is to reduce the risk of misjudgment by a single model through model collaboration. The tag mapping module is a functional unit that uses a validation set-adjusted evaluation model to correct error correction results. Specifically, it can employ an attention mechanism to enhance the correlation analysis between tags and plot elements, thereby improving the semantic accuracy of tag mapping. The AI ​​review module is a verification unit that performs consistency checks on the final tags. Specifically, it can employ a dual verification mechanism of rule engine and confidence threshold, thereby ensuring the reliability of tag updates.

[0094] Specifically, when film and television content is input into the device, the multi-source integration module first collects relevant film information from the media asset library and external data sources. After data cleaning, it generates a training set containing structured tag data. The model fine-tuning module adjusts the parameters of large models such as GPT and BERT based on this dataset, enabling them to identify tag errors under complex semantics. The task triggering and prompt word module continuously monitors the popularity data of the film and television content pool. When it detects unreviewed content with a playback volume exceeding a set threshold, it automatically triggers a tag processing task and calls a prompt word template that matches the plot elements. The tag error correction module inputs the prompt words into multiple fine-tuned large models and generates differentiated error correction suggestions through parallel inference. The tag mapping module uses a deep semantic analysis model to perform weighted evaluation of the output results of multiple models, selects the corrected tags with the highest matching degree with the plot elements, and maps them to a standard tag library. The AI ​​review module performs logical consistency verification on the mapped tags and updates the review status of the content after confirming that there are no errors.

[0095] Compared to existing technologies, traditional tag processing methods rely on a single model, leading to significant deviations in error correction results. This application, however, achieves cross-validation of error correction results through a multi-model collaborative architecture. Existing technologies using static rules to trigger tasks are prone to resource waste; this application optimizes task triggering efficiency through a dynamic popularity threshold mechanism. Furthermore, traditional tag mapping, relying on manual rule bases, struggles to handle emerging film and television content. This application achieves dynamic adaptation of the tag system by combining a prompt word engineering library with a deep evaluation model.

[0096] Through the above technical solutions, this application overcomes the limitations of traditional methods in correcting complex semantic tags, improving the diversity and accuracy of tag generation results. The automated task triggering mechanism and multi-model collaborative processing significantly reduce the cost of manual intervention. The tag mapping method based on deep semantic analysis effectively improves the conversion efficiency between different tag systems and solves the problem of rigid mapping rules in traditional methods. The final closed-loop processing flow ensures the reliability and traceability of the tag update process.

[0097] Based on the above embodiments, the present invention also provides a server, the principle block diagram of which can be as follows: Figure 4 As shown. The server includes a processor, memory, network interface, display screen, and database connected via a system bus.

[0098] A server according to this application further includes one or more programs, wherein one or more programs are stored in a memory and configured to be executed by one or more processors, the one or more programs including a method for performing a large-model-based video content tagging process.

[0099] The memory refers to a data storage unit used to store multi-source film and television datasets, fine-tuned AI model parameters, and tag mapping rules. It can be implemented using solid-state drives or distributed storage systems, and its function is to provide basic data support for the data processing flow. The processor refers to a computing unit that executes program instructions, and can be implemented using a multi-core CPU or GPU cluster. Its function is to drive the automatic triggering of task modules and coordinate the parallel computation of multiple large models. The program includes method instructions that refer to the operational logic for automatically correcting and updating film and television tags. This can be implemented through modular code encapsulation, and its function is to transform the method flow of claims 1-7 into an executable sequence of computer instructions.

[0100] Specifically, after the memory loads the pre-built multi-source film and television dataset and the fine-tuned GPT and BERT model parameters, the processor starts the automatic task triggering module according to the execution cycle set by the program. For example, it automatically scans the film and television content pool every morning, selecting films with more than a predetermined threshold of views and no re-examination status as processing targets. By calling the prompt templates related to plot conflict detection in the prompt word engineering library, it drives multiple fine-tuned large models to perform semantic analysis on the original labels. The processor inputs the error correction results output by each model into the DEEPSEEK R1 model for confidence evaluation, and selects corrected labels with confidence scores higher than the set value to write into the label library. The review rule library set in the memory will perform compliance verification on the mapped labels, and update the re-examination certification mark of the film after passing the verification.

[0101] In some specific implementations, the memory can adopt a hierarchical storage architecture, storing frequently accessed prompt word engineering libraries in an in-memory database, while storing historical video datasets on a disk array. When performing tag conflict detection, the processor can dynamically allocate computing resources, for example, deploying the GPT model to a dedicated inference card and allocating data cleaning tasks to general-purpose computing cores. The program's execution logic can be designed as a pipelined structure, including three parallel threads: data preprocessing, multi-model inference, and result fusion.

[0102] Compared to existing technologies, current film and television tagging systems typically rely on independently deployed servers and manual operation interfaces, resulting in low efficiency in hardware and software collaboration during data processing. This application, however, achieves a high degree of integration of storage resources, computing resources, and business logic by embedding the method flow into server program instructions. For example, traditional systems require manual export of a list of tags to be processed and submission of each tag to the model interface for tag re-inspection, while this application achieves fully automated processing through the coordination of an automatic task triggering module and a processor scheduling mechanism.

[0103] Through the above technical solution, this application can effectively improve the real-time performance and reliability of film and television tag processing. The collaborative operation of the memory and processor enables efficient access and parallel computation of massive amounts of film and television data, while the encapsulated method logic ensures the standardized execution of the multi-model collaborative reasoning process. Compared to traditional manual review methods, this server can process more than ten times the number of films to be inspected per unit time, and the accuracy of erroneous tag detection is significantly improved. Simultaneously, the hardware-level resource management mechanism ensures the stability of the system during peak periods, avoiding processing delays caused by competition for computing resources.

[0104] This application further proposes a computer-readable storage medium that, when executed by the processor of an electronic device, enables the electronic device to perform a film and television content tagging method based on a large model. The method includes: pre-retrieving film-related information from a film and television media asset library and integrating it to form a multi-source film and television dataset; using this dataset to fine-tune multiple specified large models; deploying an automatic task triggering module and constructing a prompt word engineering library; selecting unreviewed content through the triggering module and calling prompt words, then using the fine-tuned large model to output tag correction results; using a predetermined model to evaluate and correct the correction results and mapping them to the tag library; and updating the review and certification status after AI review of the mapped tags.

[0105] Computer-readable storage media refers to a physical carrier capable of persistently storing program instructions, which can be implemented using solid-state drives, USB flash drives, or optical discs. The stored instruction set includes modular program code for data integration, model training, and task scheduling. A processor is a hardware unit that performs logical operations, which can be implemented using a multi-core CPU or GPU cluster. It automates the video tagging process by loading instructions from the storage medium. Instructions are program code that controls the operation of electronic devices, specifically including data cleaning algorithms, model fine-tuning parameter sets, and task trigger condition judgment logic. The entire tagging process is made intelligent through the serialized execution of this code.

[0106] Specifically, after the instructions in the storage medium are loaded by the processor, the electronic device will perform multi-source data acquisition and cleaning according to a predetermined program to construct a structured dataset containing user comments, ratings, and external evaluations. Subsequently, distributed computing resources are used to fine-tune large models such as GPT and BERT in parallel, optimizing their semantic understanding capabilities in the film and television tagging task. The task triggering module monitors the popularity data of the film and television content pool in real time. When it detects content with a playback volume exceeding a set threshold and not yet reviewed, it automatically calls the prompt word library to generate task instructions. The fine-tuned large model performs deep analysis of plot elements based on the prompt words, outputs tag correction suggestions, and optimizes the results through the DEEPSEEK R1 model. Finally, the corrected tags are mapped to the standard tag library to complete the state update.

[0107] Compared to existing technologies, traditional storage media only store static processing rules or single model parameters, which cannot support complex tag processing workflows involving multiple models working together. This application enables electronic devices to dynamically schedule multiple fine-tuned large models by storing an instruction set containing complete processing logic, achieving closed-loop processing of tag correction, generation, and mapping. Existing technologies rely on manually configured tag mapping rules, while this application achieves intelligent mapping across tag systems by storing a prompt word engineering library and model optimization algorithms.

[0108] Through the above technical solution, this application effectively solves the problems of insufficient error correction capability, single generated tags, and low mapping efficiency in film and television tag processing. The multi-model collaborative instructions integrated in the storage medium enable electronic devices to recognize tag errors with complex semantics, such as accurately detecting logical conflicts between plot elements and original tags. By storing preset prompt word generation rules, it supports the generation of multi-dimensional tags covering content features and user feedback. The tag mapping algorithm embedded in the instructions can automatically adapt to different classification systems, improving the processing efficiency of cross-platform tag management.

[0109] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for processing film and television content tags based on a large model, characterized in that, include: Film-related information is retrieved from film and television media asset databases in advance and integrated to form a multi-source film and television dataset; The film and television tag data in the multi-source film and television dataset is used in advance to fine-tune multiple specified large artificial intelligence models and optimize the performance of each specified large artificial intelligence model on the film and television tag task. A pre-deployed task automatic triggering module is used to automatically select unchecked content from the film and television content pool to trigger tasks, and to build a prompt word engineering library for calling corresponding prompt words for different tag tasks; The system acquires input video content, automatically selects unchecked content from the deployed task triggering module to trigger tasks, calls corresponding prompt words from the prompt word library, and then outputs the error correction results of video content tags through each of the fine-tuned specified artificial intelligence models. A pre-defined artificial intelligence model is used to evaluate and correct the error correction results of the output film and television content tags, and the correction results are mapped to the tag library. The mapped tags are reviewed by AI, and the identification and certification status of the film and television content is updated after the review is passed.

2. The film and television content tagging method based on a large model according to claim 1, characterized in that, Before the step of pre-retrieving film-related information from a film and television media asset database and integrating it to form a multi-source film and television dataset, the following steps are also included: GPT, BERT, Qwen, and DEEPSEEK R1 models are pre-stacked to form a multi-model collaborative architecture.

3. The film and television content tagging method based on a large model according to claim 1, characterized in that, The step of pre-retrieving film-related information from a film and television media asset database and integrating it to form a multi-source film and television dataset includes: In advance, basic information about the film is retrieved from the film and television media asset database, and user comments, ratings, social media popularity data, and relevant data from authoritative film and television lists and news reports are obtained using web crawlers; The acquired basic information about the film, as well as the corresponding user comments, ratings, social media popularity data, and relevant data from authoritative film and television lists and news reports, are processed through data washing, and unstructured plot summaries and comments are converted into structured data. The cleaned and transformed data is integrated into multi-source film and television data, forming a multi-source film and television dataset that includes film and television content features, user feedback, and external evaluations, and is divided into training set, validation set, and test set according to a predetermined ratio.

4. The film and television content tagging method based on a large model according to claim 1, characterized in that, The steps of pre-tuning multiple specified large-scale artificial intelligence models using film and television tag data from the multi-source film and television dataset to optimize the performance of each specified large-scale artificial intelligence model on the film and television tagging task include: The film and television tag data in the training set of the multi-source film and television dataset is used in advance to fine-tune multiple specified large artificial intelligence models and optimize the performance of each specified large artificial intelligence model on the film and television tag task. The multiple specified large artificial intelligence models include: GPT, BERT and Qwen models. The film and television tag data includes: incorrect and correct label pairs and reference data for generating labels of unlabeled films and television shows. The DEEPSEEK R1 model is pre-tested using the validation set of the multi-source film and television dataset, and the parameters are adjusted based on the validation set to accurately evaluate the output results of the multi-model.

5. The film and television content tagging method based on a large model according to claim 1, characterized in that, The pre-deployed task automatic triggering module is used to automatically select unreviewed content from the film and television content pool to trigger tasks, and to build a prompt word engineering library for calling corresponding prompt words for different tag tasks. The steps include: A pre-deployed task automatic triggering module is used to sort the content in the film and television content pool by popularity and automatically select unreviewed content to trigger tasks. A pre-built prompt word library is used to call the corresponding prompt words for different tag tasks.

6. The film and television content tagging method based on a large model according to claim 1, characterized in that, The steps of acquiring input film and television content, automatically selecting unchecked content to trigger tasks through a deployed task triggering module, calling corresponding prompt words from a prompt word library, and then outputting the error correction results of film and television content tags through fine-tuned designated artificial intelligence models include: Get the input video content and extract the original tags of the input video content; The deployed task automatic triggering module automatically selects unreviewed content from film and television content to trigger tasks, and identifies plot elements in the unreviewed content. Based on the identified plot elements, determine whether they conflict with the original tags of the film and television content; if a conflict is found, call the corresponding conflict warning words from the warning word library. Based on the conflict warning words, the error correction results of the film and television content tags are output by each of the specified AI models after fine-tuning.

7. The film and television content tagging method based on a large model according to claim 6, characterized in that, The step of using a predetermined artificial intelligence model to evaluate and correct the error correction results of the output film and television content tags, and mapping the correction results to the tag library, includes: The DEEPSEEK R1 model, a pre-defined artificial intelligence model, is used to evaluate the error correction results of the output film and television content tags, and the tags corresponding to the identified plot elements are selected as correction tags. The function of generating tags and mapping tag libraries is invoked from the pre-built prompt word engineering library to map the corrected tags to the standard tags of the corresponding film and television content in the film and television content tag library.

8. A film and television content tagging processing device based on a large model, characterized in that, The device includes: The multi-source integration module is used to retrieve relevant film information from the film and television media asset library in advance and integrate it to form a multi-source film and television dataset; The model fine-tuning module is used to fine-tune multiple specified large artificial intelligence models in advance using the film and television tag data in the multi-source film and television dataset, and optimize the performance of each specified large artificial intelligence model on the film and television tagging task. The task triggering and prompt word module is used to pre-deploy the automatic task triggering module, which is used to automatically select unchecked content from the film and television content pool to trigger tasks, and to build a prompt word engineering library for calling corresponding prompt words for different tag tasks; The tag correction module is used to acquire input film and television content. The deployed task automatic triggering module automatically selects unchecked content from the film and television content to trigger the task, and calls the corresponding prompt words through the prompt word engineering library. Then, the fine-tuned specified artificial intelligence models output the error correction results of the film and television content tags respectively. The tag mapping module is used to evaluate and correct the error correction results of the output film and television content tags using a pre-defined artificial intelligence model, and then map the correction results to the tag library. The AI ​​review module is used to review the mapped tags using AI, and update the identification and re-certification status of the film and television content after it passes the review.

9. A server, characterized in that, It includes a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors, wherein the one or more programs include methods for performing any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device is able to perform the method as described in any one of claims 1-7.

Images