A media creation task distribution method and system based on multi-dimensional feature matching

Abstract

The application discloses a media creation task distribution method and system based on multi-dimensional feature matching, belonging to the field of data processing. The method first receives a multi-modal task requirement containing a text report, a reference material set and a structured constraint, and generates a multi-dimensional structured task feature vector group through joint semantic understanding and feature decomposition. Relying on the extraction of features from the author's historical works and the calculation of the historical originality index, a static capability portrait is constructed. Through the release of style transfer innovation challenge tasks, the creation output is evaluated and the innovation index is calculated to generate a dynamic capability vector. Combined with the task feature vector, the author's static portrait and dynamic vector, the original adaptation score is obtained through the prediction network, and the risk is corrected according to the two types of original innovation indexes to generate an author ranking recommendation list. In the task execution, the similarity of the creation content, the reference material and the copyright library features is compared in real time, and a hierarchical copyright guidance prompt is triggered, which can improve the efficiency of media creation task distribution and control the copyright risk.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for distributing media creation tasks based on multidimensional feature matching. Background Technology

[0002] With the rapid development of the digital media industry, the content creation track is becoming increasingly segmented, and the creation tasks in various formats such as short videos, text and images, and audio are surging.

[0003] Currently, most commonly used media creation task distribution methods based on multi-dimensional feature matching in the industry are traditional manual distribution or simple keyword matching distribution. The main process of this conventional technical solution is as follows: First, the task publisher submits task requirements containing basic text descriptions through tables, documents, or manual communication. In some scenarios, a small amount of reference material is also attached, which does not form a standardized and structured presentation of requirements. Second, the distribution platform or administrators perform simple classification of task requirements and manually extract core keywords, such as creation type, industry field, basic style, etc. Then, based on the extracted keywords, the platform searches for the historical work tags of creators, filters out creators whose tags initially match the keywords, generates a simple recommendation list, and pushes tasks to the creators. Because it only relies on historical work tags, it is easy to have a situation where "matching is possible but not of high quality". At the same time, the high degree of human involvement results in low distribution efficiency and makes it difficult to adapt to multi-task, fast-paced content production scenarios. Summary of the Invention

[0004] To address the aforementioned technical issues, this application provides a method and system for distributing media creation tasks based on multidimensional feature matching.

[0005] Firstly, this application provides a media creation task distribution method based on multi-dimensional feature matching, employing the following technical solution: A media creation task distribution method based on multidimensional feature matching includes: Receive multimodal task requirements input from the task publisher, wherein the multimodal task requirements include text briefings, reference material sets, and structured constraints; The multimodal task requirements are subjected to joint semantic understanding and feature deconstruction to generate a structured task feature vector set, which includes content theme vector, style and aesthetic vector, platform audience vector and constraint target vector. Based on the creator's historical works, extract the corresponding historical content and stylistic characteristics of the creator, calculate the historical originality index, and construct a static capability profile. By issuing style transfer innovation challenge tasks to creators, evaluating the corresponding creative outputs submitted by creators, calculating the corresponding current style transfer innovation index, and generating dynamic capability vectors; Based on the structured task feature vector group, the creator's static ability profile, and the dynamic ability vector, the original adaptation potential score is calculated through the adaptation prediction network. Based on the current style migration innovation index and the historical originality index, the original adaptation potential score is risk-corrected to obtain the final recommendation potential score, and a sorted recommendation list is generated based on the final recommendation potential scores of all creators; During the process of creators performing recommended tasks, feature vectors of the corresponding work content are extracted in real time and compared with the features of the reference material set and the preset high-risk copyright library. Based on the comparison results, graded guidance prompts are triggered.

[0006] By employing the aforementioned technical solutions and deeply deconstructing multimodal task requirements, vague creative needs are transformed into precise, structured feature vectors. This transforms task requirements from abstract descriptions into clear, quantifiable, and matchable indicators. This not only reduces the information gap between task publishers and creators but also allows creators to quickly and accurately grasp the core requirements of the task, reducing initial communication costs and misunderstandings. The method constructs a dual-dimensional creator evaluation system of "static ability profile + dynamic ability vector," which fully respects the creator's accumulated creative skills and stylistic traits while simultaneously capturing their innovative potential and style transfer ability in real time. This avoids stereotypical matching caused by relying solely on past works, providing more opportunities for innovative creators to stand out. Simultaneously, it allows task publishers to find suitable candidates who meet basic ability requirements and can deliver surprising creative results. Based on this, the final recommendation potential score, derived through adaptive prediction network calculations and a risk correction mechanism, can accurately match task requirements while effectively mitigating the risk of insufficient originality during the creative process. This improves the original quality of creative content from the source and reduces the potential for copyright disputes. The real-time copyright similarity comparison and tiered guidance during task execution further construct a dynamic copyright risk prevention and control system. This system can promptly identify potential problems during the creation process and provide guidance, preventing creators from having their work invalidated due to copyright issues after investing a lot of effort. This ensures the smoothness of the creation process and the compliance of the final result.

[0007] Optionally, the steps of performing joint semantic understanding and feature deconstruction on the multimodal task requirements specifically include: A pre-trained multimodal model is used to jointly analyze the text briefing and the reference material set to achieve semantic space alignment; From the aligned semantic information, we deconstruct the semantic features of four dimensions: content theme, style aesthetics, platform audience, and constraint objectives. The semantic features of the four dimensions are mapped to the corresponding vector spaces to generate the content theme vector, the style aesthetic vector, the platform audience vector, and the constraint target vector.

[0008] Optionally, the steps of issuing style transfer innovation challenges to creators, evaluating the corresponding creative outputs submitted by creators, and calculating the corresponding current style transfer innovation index specifically include: The style transfer innovation challenge task is issued to creators, requiring them to create a new theme based on the style characteristics of specified reference materials; Receive creative outputs submitted by creators in response to the aforementioned challenge task; Calculate the content similarity score between the creative output and the specified reference material in terms of specific content elements; The degree to which the creative output inherits the core stylistic characteristics of the specified reference material is evaluated to obtain a style conformity score; The current style migration innovation index is calculated according to the formula: Current Style Migration Innovation Index = Style Conformity Score × [1 - |Content Similarity Score - Expected Difference Baseline Value|].

[0009] By adopting the aforementioned technical solution, the challenge task design, centered on "specified style + entirely new theme," clarifies the direction of ability assessment for creators while reserving sufficient space for innovation. This avoids extreme cases of mere imitation or deviation from the style, allowing creators' style transfer and original innovation capabilities to be fully demonstrated within the same task scenario. In the evaluation of creative output, the dual consideration of content similarity and style conformity scores enables a comprehensive review of the creative results. This accurately identifies the creator's grasp and inheritance of the core stylistic characteristics of the reference materials, while clearly distinguishing the degree of originality of the content. This avoids plagiarism accusations arising solely from style similarity and eliminates deviations that prioritize originality while ignoring stylistic requirements.

[0010] Optionally, the step of risk-correcting the original fit potential score based on the current style migration innovation index and the historical originality index specifically includes: Based on the level of innovation requirements set by the task issuer and the content theme vector and style aesthetic vector of the task, an innovation orientation weight coefficient is generated. The final recommendation potential score is calculated using the formula: Final Recommendation Potential Score = Original Adaptation Potential Score × [Basic Retention Coefficient + Innovation Incentive Coefficient × (Innovation Orientation Weight Coefficient × Current Style Migration Innovation Index + (1 - Innovation Orientation Weight Coefficient) × Historical Originality Index)].

[0011] By adopting the above technical solution, we can fully consider the creator's past original ability foundation and reflect their current dynamic performance in style transfer and innovation in real time. This avoids the evaluation lag caused by relying solely on historical data and avoids the randomness brought about by looking only at the results of a single challenge task. As a result, the final recommendation potential score can more comprehensively and objectively reflect the creator's true ability and risk level in completing the target task.

[0012] Optionally, the step of comparing the similarity with the features of the reference material set and the preset high-risk copyright library, and triggering a graded guidance prompt based on the comparison result, specifically includes: Calculate the first real-time similarity between the feature vector of the work content and the feature vector of the reference material set; Calculate the second real-time similarity between the feature vector of the work content and the feature vector of works in a preset high-risk copyright database; When the first real-time similarity exceeds the first preset threshold, a prompt message is triggered to guide the creator to adjust the creation to avoid excessive imitation of the reference material; When the second real-time similarity exceeds a second preset threshold that is higher than the first preset threshold, an early warning message is triggered to alert the creator that the current content is at risk of infringement. When both the first real-time similarity and the second real-time similarity are lower than the first preset threshold, it is determined that the current work content is in a safe and original state, the current feature vector is recorded in the safe creation log, and background monitoring continues at the normal frequency.

[0013] By employing the aforementioned technical solution, and calculating the real-time similarity between the work content and the reference material set and the high-risk copyright database, a dual-dimensional monitoring of the creative content is achieved. This focuses on both the degree to which the creator borrows from the reference materials and closely monitors potential copyright infringement risks, allowing potential problems in the creative process to be captured in a timely manner. Setting different levels of preset thresholds and corresponding prompts and warning mechanisms allows for targeted intervention measures based on the degree of risk. When a creator shows a tendency to excessively imitate the reference materials, timely prompts can guide them to adjust their creative direction, injecting more original elements while retaining the stylistic characteristics of the references, thus avoiding the predicament of homogenized content. When the similarity between the creative content and works in the high-risk copyright database reaches a high level, strong warning messages can promptly remind the creator to pay attention to copyright issues, avoiding subsequent copyright disputes due to unintentional borrowing or similarity, and reducing unnecessary losses for the creator.

[0014] Optionally, the task distribution method further includes: After the task is completed, collect the actual effect data of the task, which includes the publisher's satisfaction score and the originality judgment result of the finished product; The final recommendation potential score is combined with the actual performance data to form a training sample; The fitness prediction network is iteratively optimized using the training samples to reduce the prediction error of the original fitness potential score.

[0015] Optionally, the task distribution method further includes: When the final recommendation potential scores of all creators are lower than the preset recommendation threshold, the gap between the structured task feature vector group and the available creator ability profile is analyzed to identify the core ability gap. Based on the core capability gaps, corresponding learning materials and micro-tasks are matched in the resource library to generate and push adaptive capability improvement paths to relevant creators.

[0016] By employing the aforementioned technical solution, when all creators' final recommendation potential scores fall below the recommendation threshold, it signifies a significant gap between the current task requirements and the creators' abilities that cannot be bridged through conventional matching. In this case, analyzing the gap between structured task feature vector sets and available creator ability profiles to identify core capability gaps allows for precise identification of the root cause of the problem, avoiding low-quality results and resource waste caused by blindly searching for creators or forcibly pushing tasks forward. Furthermore, matching corresponding learning materials and micro-tasks based on core capability gaps and generating adaptive capability improvement paths pushed to relevant creators not only provides creators with clear directions for capability improvement but also transforms abstract capability requirements into executable learning and practice tasks. This allows creators to quickly address their shortcomings through targeted training and gradually acquire the ability to undertake high-requirement tasks. For task publishers, this mechanism is no longer simply waiting for suitable creators to appear but actively participating in the cultivation of creator capabilities, shortening the task implementation cycle and reserving higher-quality creator resources for subsequent tasks.

[0017] Secondly, this application provides a media creation task distribution system based on multi-dimensional feature matching, which adopts the following technical solution: A media creation task distribution system based on multidimensional feature matching includes: The task requirement receiving module is used to receive multimodal task requirements input by the task publisher. The multimodal task requirements include text briefings, reference material sets, and structured constraints. The task feature deconstruction module is used to perform joint semantic understanding and feature deconstruction on the multimodal task requirements to generate a structured task feature vector group, which includes content theme vector, style aesthetic vector, platform audience vector and constraint target vector. The creator profile building module is used to: extract the creator's historical content and style characteristics based on the creator's historical works, and calculate the historical originality index to build a static capability profile; and evaluate the corresponding creative output submitted by the creator by issuing style transfer and innovation challenge tasks to the creator, and calculate the corresponding current style transfer and innovation index to generate a dynamic capability vector. The task matching and recommendation module is used to: calculate the original adaptation potential score through an adaptation prediction network based on the structured task feature vector group, the creator's static ability profile, and the dynamic ability vector; perform risk correction on the original adaptation potential score based on the current style transfer innovation index and the historical originality index to obtain the final recommendation potential score; and generate a ranked recommendation list based on the final recommendation potential scores of all creators. The execution process monitoring module is used to extract the feature vector of the corresponding work content in real time during the process of the creator performing the recommended task, and compare it with the features of the reference material set and the preset high-risk copyright library. Based on the comparison results, it triggers graded guidance prompts.

[0018] Thirdly, this application provides a computer device that adopts the following technical solution: A computer device includes a memory, a processor, and a computer program stored in the memory, the processor executing the computer program to implement the media creation task distribution method based on multidimensional feature matching as described in the first aspect.

[0019] Fourthly, this application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described in the first aspect, a media creation task distribution method based on multidimensional feature matching.

[0020] In summary, this application includes at least one of the following beneficial technical effects: By decomposing the demand for multimodal creative tasks, a two-dimensional creator evaluation system of static ability profile and dynamic innovation assessment is built. Combined with the fit prediction network and innovation risk correction, intelligent and accurate distribution is achieved, breaking down the information barriers between supply and demand, avoiding the problems of rigid matching and inflexible evaluation standards, reducing communication and misunderstanding biases, and improving the quality of task and creator fit.

[0021] It enables dynamic prevention and control of copyright risks throughout the entire creative process, and avoids excessive imitation and infringement risks in a timely manner through real-time similarity classification and comparison, ensuring the compliance of creative results; at the same time, it forms a closed-loop mechanism of task effect feedback iteration, accurate identification of capability gaps and adaptive improvement, continuously optimizing matching accuracy and cultivating high-quality creator resources. Attached Figure Description

[0022] Figure 1 This is a first flowchart of an embodiment of the method of this application; Figure 2 This is a second flowchart of an embodiment of the method of this application; Figure 3 This is a third flowchart of an embodiment of the method of this application; Figure 4 This is the fourth flowchart of an embodiment of the method of this application. Detailed Implementation

[0023] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-4 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0024] The first embodiment of this application discloses a media creation task distribution method based on multi-dimensional feature matching. (Refer to...) Figure 1 The media creation task distribution method based on multidimensional feature matching includes S110-S170: S110, receive multimodal task requirements input by the task issuer. Multimodal task requirements include text briefing, reference material set and structured constraints; S120 performs joint semantic understanding and feature deconstruction on multimodal task requirements to generate a structured task feature vector set, which includes content theme vector, style and aesthetic vector, platform audience vector, and constraint target vector. S130, based on the creator's historical works, extracts the corresponding historical content and style characteristics of the creator, calculates the historical originality index, and constructs a static capability profile; S140 evaluates the corresponding creative output submitted by creators by issuing style transfer innovation challenge tasks, calculates the corresponding current style transfer innovation index, and generates a dynamic capability vector. S150: Based on the structured task feature vector group and the creator's static ability profile and dynamic ability vector, the original adaptation potential score is calculated through the adaptation prediction network. S160, based on the current style migration innovation index and the historical originality index, performs risk correction on the original adaptation potential score to obtain the final recommendation potential score, and generates a sorted recommendation list based on the final recommendation potential scores of all creators; S170: During the process of creators performing recommended tasks, the feature vector of the corresponding work content is extracted in real time and compared with the features of the reference material set and the preset high-risk copyright library. Based on the comparison results, a graded guidance prompt is triggered.

[0025] Specifically, for step S110, a dedicated multimodal task publishing platform is built, which supports task publishers to submit requirements through web pages and API interfaces. The platform has preset standardized input templates to ensure that the requirement information is complete and in a consistent format.

[0026] The text brief should be filled in according to the fixed structure of "core objective of the task - delivery requirements - time nodes - supplementary explanation". For example, the text brief of a brand's short video promotion task can be filled in as "core objective: promote new skin care products, highlighting moisturizing effects; delivery requirements: 3 short videos of 15-30 seconds each, 1 set of supporting text and image copy; time nodes: deliver the first draft within 7 working days; supplementary explanation: adapt to Douyin and Xiaohongshu platforms, with a style that fits the young female audience".

[0027] The reference material set supports uploading various file types, including images (such as JPG and PNG formats with a resolution of at least 1080P), videos (such as MP4 formats with a frame rate of 25fps and a bit rate of at least 5Mbps), text (such as Word and TXT formats), and audio (such as MP3 formats). It also supports uploading or associating historical dissemination data links for materials. The number of reference materials per task cannot exceed 20, and the size of a single file cannot exceed 100MB. The platform automatically performs format verification and virus scanning on uploaded materials. After verification and virus scanning pass, the materials are stored in the distributed material library, generating a unique material ID and associating it with the corresponding task.

[0028] Structured constraints are set through a combination of platform dropdown selection and custom input. Preset constraint types include delivery format constraints (such as short video resolution and text word count), copyright constraints (such as prohibiting the use of unauthorized materials), style constraints (such as prohibiting exaggerated marketing statements), and originality constraints (such as an originality rate of no less than 80%). Custom constraints must clearly define quantitative standards and avoid vague expressions. For example, "an originality rate of no less than 80%" must be supplemented with the explanation that "determined by the platform's copyright detection tool, a duplication rate of less than 20% with existing works is considered compliant."

[0029] Furthermore, when publishing a task, the task publisher can explicitly set the "innovation requirement level" (e.g., conservative, moderate, radical) using a slider control or selection box. This level is independent of structured constraints and is used to generate innovation-oriented weight coefficients later, avoiding circular dependencies with the constraint target vector. After all input information is submitted, the platform automatically generates a task ID, completing the reception and archiving of multimodal task requirements.

[0030] S120, the specific steps for joint semantic understanding and feature deconstruction for multimodal task requirements include: A pre-trained multimodal model is used to jointly analyze the text briefing and reference material set to achieve semantic space alignment. From the aligned semantic information, semantic features in four dimensions—content theme, style aesthetics, platform audience, and constraint target—are deconstructed. The semantic features in the four dimensions are mapped to the corresponding vector spaces to generate content theme vector, style aesthetics vector, platform audience vector, and constraint target vector.

[0031] Specifically, for step S120, a pre-trained multimodal large model (such as the BLIP-2 model) is used as the core analysis tool. This model supports joint encoding of text and multimodal data such as images and videos, and can effectively achieve semantic space alignment.

[0032] First, the text briefing is segmented and stop words are removed to extract core keywords (such as "skincare products, moisturizing, young women, TikTok"). At the same time, frames are extracted from images and videos in the reference material set (e.g., one frame is extracted from every 2 seconds of video). The audio is processed for speech-to-text conversion. All multimodal data are uniformly input into the encoding layer of the multimodal big data model. The semantic association between the text and each modality material is realized through the cross-attention mechanism, and the semantic space is aligned to ensure that information from different modalities can be compared and analyzed under the same semantic dimension.

[0033] Subsequently, based on the aligned semantic information, the model's feature deconstruction module extracts semantic features in four dimensions: The content theme dimension focuses on extracting the core creative direction and core expressive elements of the task. For example, the content theme features of the aforementioned skincare product promotion task are "demonstration of skincare product moisturizing effects and presentation of usage scenarios for young women"; the style aesthetic dimension extracts the visual style (such as color tone and composition), language style (such as colloquial and artistic), and auditory style (such as background music type) of the reference materials. For example, if the reference materials are in a fresh Japanese style, the style aesthetic features would be "light color tone, natural lighting, colloquial narration, and soft background music"; the platform audience dimension retrieves historical dissemination data of similar content from the platform's historical database based on the content features of the reference materials (or reads historical dissemination data associated with the materials) to extract audience characteristics such as age, gender, and interest tags, for example, "women aged 18-28 who are interested in beauty and skincare and like short video content"; and the constraint target dimension extracts the core quantitative indicators in the structured constraints, such as "originality rate ≥ 80%, delivery time ≤ 7 working days, and compatibility with both Douyin and Xiaohongshu platforms."

[0034] Then, a feature mapping layer (such as a multilayer perceptron MLP) is used to map the semantic features of the four dimensions to a vector space of preset dimensions, generating content theme vectors, style aesthetics vectors, platform audience vectors, and constraint target vectors. The vectors of the four dimensions are then concatenated in sequence to form a structured task feature vector group. The vector group is stored in the task feature database and associated with the corresponding task ID.

[0035] For step S130, a creator's historical portfolio management system is built. This system is linked to the creator's account and automatically synchronizes all tasks completed by the creator on the platform and the creative works uploaded by the creator. It covers multiple types of content such as text, images, videos, and audio. Each work is associated with metadata information such as creation time, task type, delivery effect, and originality score.

[0036] Based on this system, a multimodal feature extraction tool is used to extract the creator's historical content and style features: for text-based works, the BERT model is used to extract text semantic features, sentence structure features, and keyword preferences; for image and video works, the CNN model is used to extract visual features (hue, composition, and camera language); for audio works, features such as audio rhythm, timbre, and narration style are extracted, and all features are integrated into a 1024-dimensional historical feature vector.

[0037] Simultaneously, the creator's historical originality index is calculated. The calculation logic is as follows: Historical Originality Index = (Sum of originality rates of all works ÷ Total number of works) × w1 + (Number of works without copyright complaints ÷ Total number of works) × w2 + (Average positive evaluation scores obtained by original works ÷ 10) × w3, where w1, w2, and w3 are preset weight coefficients, and w1 + w2 + w3 = 1. The originality rate is determined by the platform's copyright detection tool (comparing the platform's copyright library and third-party copyright databases), with a value range of 0-1. The number of works without copyright complaints refers to the number of works that have not received any copyright complaints and have passed the review in the past 12 months. The positive evaluation score is calculated as the average of the originality scores (0-10 points) given by the task publisher. The final historical originality index ranges from 0 to 1, rounded to two decimal places.

[0038] Finally, the historical feature vector and historical originality index are fused and mapped through a feature mapping layer (such as a multilayer perceptron MLP) to construct a static capability profile vector of the creator (e.g., 1024 dimensions). The semantic labels mapped by this profile include core dimensions such as creative field (e.g., beauty, food, technology), preferred style (e.g., Japanese fresh style, hardcore science popularization), originality level (divided into four levels: S, A, B, and C based on the historical originality index), preferred platform (e.g., Douyin, Xiaohongshu), and delivery efficiency. The static capability profile is updated in real time, and the historical feature vector and historical originality index are optimized synchronously after each task is completed.

[0039] Reference Figure 2 In S140, the steps of issuing style transfer innovation challenge tasks to creators, evaluating the corresponding creative outputs submitted by creators, and calculating the corresponding current style transfer innovation index specifically include S210-S250: S210 presents creators with a style transfer innovation challenge, requiring them to create a completely new theme based on the stylistic characteristics of specified reference materials. S220 receives creative outputs submitted by creators in response to the challenge task; S230, Calculate the content similarity score between the creative output and the specified reference material on specific content elements; S240 assesses the degree to which the creative output inherits the core stylistic characteristics of the specified reference material, and obtains a style conformity score; S250, according to the formula: Current Style Transfer Innovation Index = Style Conformity Score × [1 - |Content Similarity Score - Expected Difference Baseline Value|], the current style transfer innovation index is calculated.

[0040] Specifically, for step S140, based on the creator's static ability profile, creators who are initially matched with the current multimodal task requirements in terms of creative field and platform of expertise are selected, and style transfer innovation challenge tasks are released to them in a targeted manner. The release methods include platform message push, SMS reminder and email notification to ensure that creators receive them in a timely manner.

[0041] The challenge tasks are designed to closely align with the style of the reference materials and the new theme, clearly defining the task requirements, delivery format, submission deadline, and scoring criteria. The core requirement is to "create a new video based on the stylistic characteristics of the specified reference materials, ensuring stylistic continuity while avoiding content duplication." For example, if the current task's reference materials are "Japanese-style fresh and natural skincare product short videos," and the new theme is "Japanese-style fresh and natural beauty tool short videos," then the challenge task could be designed as follows: "Based on the three provided Japanese-style fresh and natural skincare product short videos (material IDs: XXX-XXX), inherit their light color tones, natural lighting, and conversational narration, create a 15-second short video promoting beauty tools. The theme should be the usage techniques of a beauty sponge. The use of shots, text, and background music from the reference materials is prohibited. The delivery format should be MP4 (1080P, 25fps), and the submission deadline is within 24 hours. Scoring will be based on style conformity and content originality." Meanwhile, the platform provides a dedicated submission portal for challenge tasks, clearly indicating the task deadline, making it convenient for creators to view task details, upload their creative output, and allowing creators to preview and modify their work before submission.

[0042] A challenge task creation output receiving system has been established and linked to the challenge task publishing portal. After completing their creations, creators can upload their outputs through a dedicated portal. The system automatically verifies whether the format and size of the works meet the requirements (consistent with the challenge task requirements). For example, short videos must be in MP4 format and 1080P resolution, and text must be in Word format and have a minimum of 300 words. If the format does not meet the requirements, the system will immediately display a prompt message informing the creator of the specific modification requirements until the verification is passed.

[0043] After successful verification, the system automatically assigns a unique output ID to the creation, linking it to the creator ID and challenge task ID. It also records the submission time. If the creator fails to submit before the deadline, the system will send three reminders (6 hours, 2 hours, and 30 minutes before the deadline). After the deadline, the submission portal will be closed, and the challenge task will be considered abandoned. Furthermore, the system allows creators to modify their work after submission but before the deadline. Each modification automatically overwrites the original output file, updates the submission time, and records the number of modifications. The number of modifications is included as a reference factor in subsequent dynamic ability assessments. All received creative outputs are stored in a temporary output database under encrypted storage to prevent leakage. After subsequent scoring is completed, they will be synchronized to the creator's portfolio system.

[0044] A cosine similarity-based content similarity calculation method, combined with multimodal content feature extraction technology, is used to calculate the similarity score between the creative output and the specified reference material on specific content elements. First, specific content elements are extracted from both the creative output and the reference material: For text-based outputs (copywriting, narration), core keywords, sentence structure, paragraph logic, and core viewpoints are extracted and transformed into 512-dimensional feature vectors using the BERT model; for image and video outputs, camera angles, scene settings, character actions, and visual elements (such as props and backgrounds) are extracted and transformed into 512-dimensional feature vectors using the CNN model; for audio outputs, background music melody, narration speed, and tone are extracted and transformed into 512-dimensional feature vectors.

[0045] Subsequently, the cosine similarity between the feature vectors of the created content and the reference material is calculated. The cosine similarity value ranges from 0 to 1. The higher the similarity, the higher the degree of repetition of content elements. This cosine similarity value is used as the content similarity score, rounded to two decimal places. For example, if the core content elements of the reference material short video are "a woman holding skincare products, a close-up of her face, and a scene with natural light by a window," and the content elements of the created short video are "a woman holding a beauty blender, a close-up of her face, and a scene with natural light by a window," after extracting the visual feature vectors of the two, the cosine similarity is calculated to be 0.35, which means the content similarity score is 0.35. This indicates that the content elements of the two are related to some extent, but the degree of repetition is low. If the created content directly uses clips from the reference material, the calculated content similarity score may reach 0.8 or higher, which is considered excessive content repetition.

[0046] A combination of machine and human evaluation is used to assess the degree to which the creative output inherits the core stylistic features of specified reference materials, ultimately yielding a style conformity score (ranging from 0 to 1, rounded to two decimal places). Machine evaluation employs a pre-trained style recognition model (optimized and trained based on ResNet-50, specifically adapted for media creation styles) to extract core stylistic features from the reference materials (such as tone parameters, compositional proportions, language style keywords, and audio rhythm features), forming a style feature template. The stylistic features of the creative output are then extracted and matched against this template, calculating the matching degree to obtain the machine evaluation score (ranging from 0 to 1). The manual evaluation process involves 3-5 professional style reviewers (reviewers must have experience in style judgment within their respective creative fields; for example, senior creators in the beauty industry are invited for beauty-related tasks, and copywriting experts are invited for copywriting tasks). Reviewers score the creative output based on the core style characteristics of the reference material, evaluating it across three dimensions: "style consistency," "style completeness," and "style accuracy" (each dimension ranging from 0 to 10). The average score from all reviewers is then normalized to obtain the manual evaluation score (range 0-1). The final style conformity score = machine evaluation score × z + manual evaluation score × (1-z), where z is the preset weighting coefficient for the machine evaluation, ranging from 0 to 1. For example, if the reference material is a Japanese fresh style, the machine evaluation score is 0.8, and the three reviewers score it 8, 9, and 8 respectively, with an average score of 8.33, which is 0.833 after normalization. If z is set to 0.7, then the style conformity score = 0.8 × 0.7 + 0.833 × 0.3 ≈ 0.81.

[0047] The innovation index is calculated using the preset formula: "Current Style Transfer Innovation Index = Style Conformity Score × [1 - |Content Similarity Score - Expected Difference Baseline Value|]". The expected difference baseline value is a preset hyperparameter (typically 0.3-0.5) used to characterize a reasonable degree of innovation deviation. The calculation process retains four decimal places, and the final result is rounded to two decimal places, ranging from 0 to 1. A higher score indicates a stronger style transfer innovation ability. This formula uses an absolute value penalty term to penalize both excessively high content similarity (blankly copying reference materials) and excessively low content similarity (completely off-topic, deviating from the framework of reference materials), ensuring that true innovation is "giving new content while retaining the core style". For example, if the style conformity score is 0.81 and the expected difference baseline is set to 0.4; if the content similarity score is 0.35, then the innovation index = 0.81 × [1 - |0.35 - 0.4|] = 0.81 × 0.95 = 0.77; if the content similarity score is 0.9 (copying), then the innovation index = 0.81 × [1 - |0.9 - 0.4|] = 0.81 × 0.5 = 0.405; if the content similarity score is 0.1 (off-topic), then the innovation index = 0.81 × [1 - |0.1 - 0.4|] = 0.81 × 0.7 = 0.567.

[0048] Subsequently, a dynamic capability vector is generated. Scalar and discrete features such as the current style transfer innovation index, the completion status of the challenge task (e.g., on-time submission, number of revisions, etc.), and the quality score of the creative output (supplemented by reviewers, 0-1) are first transformed into a low-dimensional dense vector through an embedding network containing two fully connected layers. The first layer maps the input to a 64-dimensional dense vector, and the second layer maps it to a 512-dimensional dynamic capability vector. The activation function is ReLU. This embedding network is jointly trained end-to-end with the subsequent fitness prediction network. The dynamic capability vector is updated in real time, and is recalculated and updated after each challenge task or formal task is completed.

[0049] For step S150, an fitness prediction network based on multilayer perceptron (MLP) is used. This network can effectively fuse multi-dimensional vector features and perform regression prediction, thereby improving matching accuracy.

[0050] First, the generated structured task feature vector group, the generated creator static ability profile vector, and the generated dynamic ability vector are concatenated. An attention mechanism is used to assign different weights to the vector features from different sources (for example, the weight of the structured task feature vector group is set to 0.5, the weight of the static ability profile vector is set to 0.3, and the weight of the dynamic ability vector is set to 0.2). The weighted concatenated feature vector is mapped to a unified feature space through a fully connected layer (projection matrix) to obtain a fused feature vector (e.g., 2048-dimensional).

[0051] Subsequently, the fused feature vector is input into the subsequent fully connected layer of the suitability prediction network. Non-linear feature mapping is performed using an activation function (such as ReLU), ultimately outputting an original suitability potential score, ranging from 0 to 100. A higher score indicates a higher suitability between the creator and the current task. For example, a creator's static ability profile shows expertise in beauty-related short video creation, their dynamic ability vector has an innovation index of 0.77, and their structured task feature vector group corresponds to a beauty-related short video promotion task. After fusion, the input to the suitability prediction network yields an original suitability potential score of 85, indicating a high suitability for the current task. Conversely, if a creator excels in technology-related content creation, their static ability profile may not match the task domain, resulting in an original suitability potential score below 40, indicating a lower suitability. All creators' original suitability potential scores are stored in the matching result database, linked to the task ID and creator ID.

[0052] In S160, the specific steps for risk correction of the original fit potential score based on the current style migration innovation index and the historical originality index include: Based on the innovation requirement level set by the task issuer and the content theme vector and style aesthetic vector of the task, an innovation-oriented weight coefficient is generated; according to the formula: Final recommendation potential score = Original adaptation potential score × [Basic retention coefficient + Innovation incentive coefficient × (Innovation-oriented weight coefficient × Current style migration innovation index + (1 - Innovation-oriented weight coefficient) × Historical originality index)], the final recommendation potential score is calculated.

[0053] Specifically, for step S160, the generation logic of the innovation-oriented weight coefficients is as follows: First, the innovation requirement level (conservative, moderate, radical) explicitly set by the task publisher in S110 is used as a discrete input and transformed into a 16-dimensional initial level feature vector through an embedding layer. Simultaneously, the content theme vector and style aesthetic vector are extracted from the structured task feature vector group generated in S120, and their cosine similarity is calculated to obtain a 1-dimensional style theme difference feature. Then, the 16-dimensional initial level feature vector and the 1-dimensional difference feature are concatenated into a 17-dimensional joint feature, which is input into a fully connected network containing a ReLU activation function and a Sigmoid output layer, mapping the output to an innovation-oriented weight coefficient between 0 and 1. This generation logic combines the publisher's subjective intent with the objective features of the task and completely decouples from the constrained target vector, avoiding circular dependencies. The parameters of the fully connected network are iteratively optimized end-to-end along with the fitness prediction network.

[0054] Subsequently, the final score is calculated according to the preset formula: "Final Recommended Potential Score = Original Adaptation Potential Score × [Basic Retention Coefficient + Innovation Incentive Coefficient × (Innovation-Oriented Weight Coefficient × Current Style Migration Innovation Index + (1 - Innovation-Oriented Weight Coefficient) × Historical Originality Index)]". The basic retention coefficient (e.g., 0.6) and innovation incentive coefficient (e.g., 0.4) are preset hyperparameters to ensure the final score is not systematically suppressed. The calculation result is rounded to one decimal place, ranging from 0 to 100. For example, if a creator's original adaptation potential score is 85, the generated innovation-oriented weight coefficient is 0.7, the current style migration innovation index is 0.77, and the historical originality index is 0.8, then the weighted innovation value = 0.7 × 0.77 + 0.3 × 0.8 = 0.779, and the final recommended potential score = 85 × [0.6 + 0.4 × 0.779] = 85 × 0.9116 ≈ 77.5 points.

[0055] After the calculation is completed, all participating creators are sorted from high to low according to their final recommendation potential score, generating a sorted recommendation list. The list includes the creator ID, core information of the static ability profile, core indicators of the dynamic ability vector (innovation index), original adaptation potential score, and final recommendation potential score. At the same time, auxiliary information such as the creator's delivery efficiency and satisfaction with past tasks are also marked. This list is pushed to the task publisher so that they can select suitable creators to undertake the task.

[0056] Reference Figure 3 In S170, the steps of comparing the similarity with the features of the reference material set and the preset high-risk copyright database, and triggering graded guidance prompts based on the comparison results, specifically include S310-S350: S310, calculate the first real-time similarity between the feature vector of the work content and the feature vector of the reference material set; S320, calculates the second real-time similarity between the feature vector of the work content and the feature vector of the works in the high-risk copyright library; S330: When the first real-time similarity exceeds the first preset threshold, a prompt message is triggered to guide the creator to adjust the creation to avoid excessive imitation of the reference material. S340, when the second real-time similarity exceeds the second preset threshold which is higher than the first preset threshold, a warning message is triggered to alert the creator that the current content is at risk of infringement; S350, when both the first real-time similarity and the second real-time similarity are lower than the first preset threshold, it is determined that the current work content is in a safe original state, the current feature vector is recorded to the safe creation log, and the background monitoring continues at the normal frequency.

[0057] Specifically, for step S170, a real-time monitoring system for task execution is built, which is linked with the creator's creation tools (built-in creation tools on the platform or authorized third-party creation tools). During the process of the creator executing the recommended task, the feature vector of the corresponding work content is extracted in real time, with the extraction frequency set to once every 5 minutes to ensure the real-time monitoring.

[0058] First, feature vectors of the work content are extracted. Based on the type of work content (text, image, video, audio), the corresponding feature extraction model (BERT, CNN, etc.) is used to transform it into a 512-dimensional feature vector. Then, the first real-time similarity is calculated: the feature vector of the work content is compared with the feature vector of the S110 reference material set (which has been extracted and stored in advance) using cosine similarity calculation to obtain the first real-time similarity, with a value ranging from 0 to 1. This is used to determine whether there is excessive imitation of the reference material.

[0059] Simultaneously, a second real-time similarity is calculated: a pre-built high-risk copyright database is constructed based on the Faiss vector retrieval engine. This database uses an IVF_PQ index structure to store feature vectors of officially licensed works, past infringement cases, and popular online works to support millisecond-level retrieval. Weekly, the database automatically pulls newly added popular works and historical infringement cases from third-party copyright databases via API, extracts 512-dimensional feature vectors, and incrementally updates the index. The Faiss engine quickly retrieves works from the high-risk copyright database that have the most similar feature vectors to the work content, calculates the cosine similarity between the two, and takes the maximum value as the second real-time similarity, ranging from 0 to 1, to determine whether there is an infringement risk.

[0060] Two preset thresholds are used. The first preset threshold is set to 0.7 (which can be flexibly adjusted according to the task type; 0.6 for text tasks and 0.7 for video and image tasks) to determine whether there is excessive imitation of reference materials. The second preset threshold is set to 0.85 (higher than the first preset threshold to ensure the accuracy of infringement warnings; all tasks have the same standard) to determine whether there is any infringement risk. Both thresholds have been tested with a large amount of historical data and can effectively balance creative freedom and risk control.

[0061] When the real-time monitoring system detects that the first real-time similarity exceeds the first preset threshold, it immediately triggers a prompt message. The prompt message is pushed to the creator simultaneously through three methods: pop-up window in the creation tool, platform message, and SMS. The prompt content is clear and specific, such as "The current work content has a high similarity to the reference material. Please adjust the creation content to avoid excessive imitation. You can try changing the scene, adjusting the copywriting, or the camera angle." At the same time, it provides reference adjustment suggestions to help creators quickly optimize their content.

[0062] When the detected real-time similarity exceeds the second preset threshold, an early warning is immediately triggered. This warning is sent not only to the creator but also simultaneously to platform reviewers and the task publisher. The warning clearly indicates the infringement risk level (high risk), for example, "The current work content has a high similarity to a work in the high-risk copyright library, posing an infringement risk. Please modify the content immediately; otherwise, you will be unable to submit the task, and your account credit score will be affected." Simultaneously, the system automatically suspends the creator's task submission privileges until the creator modifies the content and passes the similarity test. Reviewers simultaneously conduct a manual review; once the infringement risk is confirmed, submission privileges are restored. If a creator triggers multiple warnings and fails to effectively modify the content, the system will restrict their ability to accept subsequent tasks.

[0063] When the real-time monitoring system detects that both the first real-time similarity and the second real-time similarity are lower than the first preset threshold, it determines that the current work content is in a safe and original state. The system automatically records the feature vector of the current work content to the safe creation log for traceability, and displays a green pass sign indicating "good creation status" on the creator's creation tool interface. No intervention is made, and background monitoring continues at the normal frequency (e.g., once every 5 minutes).

[0064] Reference Figure 4 Furthermore, the task distribution method also includes S410-S430: S410: After the task is completed, collect the actual effect data of the task. The actual effect data includes the publisher's satisfaction rating and the originality judgment result of the finished product. S420 combines the final recommended potential score with the actual performance data into training samples; S430 uses training samples to iteratively optimize the fitness prediction network to reduce the prediction error of the original fitness potential score.

[0065] Specifically, for step S410, after the creator completes the task and submits the final video, the platform automatically triggers the actual effect data collection process. The collection period is 7 working days after the final video is submitted, ensuring the integrity and objectivity of the data.

[0066] The publisher satisfaction rating is collected through a dedicated rating portal on the platform, which is linked to a task ID. Publishers must complete the rating within 7 business days; failure to do so within the deadline is considered as default satisfaction (rating out of 10 points). Rating dimensions include content quality (0-10 points), originality (0-10 points), style matching (0-10 points), and delivery timeliness (0-10 points). Each dimension requires a specific rating and a brief evaluation. The platform automatically calculates the average score across the four dimensions as the publisher satisfaction rating (rounded to one decimal place, ranging from 0-10). The rating is then normalized (divided by 10) to obtain a normalized score of 0-1, which is used for subsequent model training.

[0067] The originality assessment of the final product is determined by a combination of the platform's copyright detection module and manual review. The copyright detection module compares the platform's copyright library with third-party copyright databases (such as CNKI and Visual China) to detect the originality rate and infringement status of the final product. The manual review is conducted by the platform's professional copyright reviewers, focusing on checking the suspected infringement parts indicated by the detection module. The final assessment results are divided into four levels: "Completely Original" (originality rate ≥ 90%, no infringement), "Basically Original" (80% ≤ originality rate < 90%, no infringement), "Suspected Infringement" (originality rate ≥ 80%, but with a small amount of unauthorized material references), and "Clear Infringement" (originality rate < 80% or with obvious infringing material), corresponding to values ​​of 1, 0.8, 0.3, and 0, respectively. The assigned values ​​are used for subsequent training sample construction. All collected actual effect data (publisher satisfaction rating, normalized score, originality assessment results and assigned values) are associated with the task ID and creator ID and stored in the effect data database.

[0068] For step S420, training samples are constructed based on the final recommendation potential score and the actual effect data from S410. These samples must meet the requirements of diversity and representativeness. Each training sample contains two parts: input features and labels. The input features are a structured task feature vector group (e.g., 2048 dimensions), a creator's static ability profile vector (e.g., 1024 dimensions), and a dynamic ability vector (e.g., 512 dimensions), which are concatenated in sequence to form an input feature vector (e.g., 3584 dimensions). The label is the comprehensive score of the actual effect, calculated as "Comprehensive score of actual effect = Normalized score of publisher satisfaction × f + Originality judgment of finished product × (1-f)", where f is the preset weight of satisfaction, ranging from 0 to 1, rounded to two decimal places. This score is used to measure the prediction accuracy of the final recommendation potential score.

[0069] The construction of training samples requires screening valid data and removing abnormal samples (such as malicious ratings by publishers or cases where the final product contains serious infringement but is incorrectly judged). The total number of samples should be no less than 1000, covering different task types (copywriting, video, images, etc.), different creator ability levels (S, A, B, C levels), and different innovation-oriented weight coefficient scenarios to ensure sample diversity. All screened training samples are divided into a training set (80%) and a validation set (20%) in an 8:2 ratio. The training set is used for iterative training of the fit prediction network, and the validation set is used to verify the training effect. All samples are stored in the model training database using encrypted storage to prevent data leakage.

[0070] For step S430, the gradient descent algorithm (Adam optimizer, learning rate set to 0.001) is used, with the "prediction error of the original fitness potential score" as the loss function, to iteratively optimize the fitness prediction network. The loss function is calculated as "loss value = |normalized value of original fitness potential score - comprehensive score of actual effect|", where the normalized value of the original fitness potential score is the original score divided by 100 (range 0-1). The smaller the loss value, the smaller the prediction error and the higher the accuracy of the fitness prediction network.

[0071] During iterative training, each iteration uses all samples from the training set for training, updating the parameters of the fitness prediction network (fully connected layer weights, attention mechanism weights, etc.). After each iteration, the average loss value is calculated using the validation set, and the iteration log is recorded. The preset number of iterations is 50 rounds. If the average loss value of the validation set changes by less than 0.0001 for three consecutive rounds, it indicates that the model has stabilized, and the iteration can be stopped early. If the loss value is still not stable after 50 iterations, the learning rate is adjusted appropriately (reduced to 0.0001), and the number of iterations is increased to 80 rounds. After optimization, the optimized fitness prediction network (including its attention mechanism weights and fully connected layer parameters) and the embedding network parameters of the dynamic capability vector replace the original network parameters for matching calculations of subsequent new tasks. Every 500 newly added valid training samples, the above iterative optimization process is repeated to continuously reduce the prediction error of the original fitness potential score and improve the matching accuracy between creators and tasks.

[0072] Furthermore, the task distribution method also includes: When all creators’ final recommendation potential scores are lower than the preset recommendation threshold, the gap between the structured task feature vector group and the available creator ability profile is analyzed to identify core ability gaps. Based on the core ability gaps, corresponding learning materials and micro-tasks are matched in the resource library to generate and push adaptive ability improvement paths to relevant creators.

[0073] Specifically, a preset recommendation threshold is set, which is determined based on historical task matching data statistics. For example, a value of 30 points (0-100 points) is used. This threshold has been tested and can effectively distinguish between suitable and unsuitable creators. If the final recommendation potential score of all creators is lower than 30 points, it means that the current available creators' abilities cannot meet the task requirements, and the core ability gap needs to be identified.

[0074] Subsequently, a feature comparison module was built. Through a pre-set feature mapping layer, the structured task feature vector group, static ability profile vector, and dynamic ability vector were uniformly mapped to a common feature space of the same dimension. Within the common feature space, the vector difference between the task feature vector and the creator feature vector was calculated. The larger the difference, the greater the gap between the creator's ability and the task requirements. By performing feature attribution analysis on this vector difference, the feature dimension causing the largest difference was located and mapped back to the corresponding semantic label space. The label dimension with the most significant gap was extracted, which is the core ability gap. The core ability gap needs to be specific and quantifiable to avoid vague descriptions. For example, the structured task feature vector group of a certain task requires "proficiency in Douyin short video editing, ability to transfer and innovate in ancient style, and originality rate ≥ 85%". However, among all available creators' static ability profiles, creators who are proficient in ancient style account for less than 10%, and the style transfer and innovation index in the dynamic ability vectors are all below 0.4, with an average originality rate of 75%. Therefore, the core ability gap is "ability to transfer and innovate in ancient style, ability to edit Douyin short videos, and originality ability (needs to be improved to 85% or more)".

[0075] At the same time, the causes of core competency gaps are analyzed, such as "insufficient reserves of creators in the ancient style" and "insufficient training in style transfer and innovation for existing creators," to provide a basis for generating subsequent competency improvement paths. The core competency gap analysis results are stored in the competency gap database and associated with task IDs.

[0076] Based on the identified core competency gaps, a resource library of learning materials and micro-tasks is built. The resource library is categorized and stored according to creative fields and competency types, covering learning materials (tutorial videos, text tutorials, case studies, and live courses) and micro-tasks (small creative tasks with difficulty levels ranging from low to high). Each resource is associated with a corresponding competency dimension. For example, the learning materials for "traditional Chinese style transfer and innovation competency" are "traditional Chinese style short video editing tutorials and case studies," and the corresponding micro-tasks are "imitating traditional Chinese style copywriting and editing short video clips in the traditional Chinese style."

[0077] The system uses a feature matching algorithm to select the most suitable learning materials and micro-tasks from the resource library based on core competency gaps, generating an adaptive competency improvement path. The improvement path is designed with a gradient of "basic learning - practical exercises - comprehensive improvement," with each stage's learning materials and micro-tasks corresponding one-to-one, and the difficulty gradually increasing. For example, the improvement path for "traditional Chinese style transfer and innovation ability" is as follows: Stage 1 (basic learning): watch tutorials on traditional Chinese style short video style analysis and learn techniques for adjusting the color tone of traditional Chinese style; Stage 2 (practical exercises): complete micro-task 1 (imitate a piece of traditional Chinese style copywriting) and micro-task 2 (edit a 10-second traditional Chinese style short video clip); Stage 3 (comprehensive improvement): complete micro-task 3 (create a complete short video in traditional Chinese style based on reference materials). Each micro-task has a corresponding score and feedback after completion, helping creators understand their shortcomings.

[0078] Once the improvement path is generated, it will be pushed to relevant creators (i.e., all available creators participating in this task matching) via platform messages and emails. It will also be displayed in the "Skills Enhancement" section of the creator's account to track their learning progress and micro-task completion. If a creator completes all micro-tasks within the specified time and achieves the required score, the system will update their static skill profile and dynamic skill vector, increasing their priority for subsequent task matching. Furthermore, the difficulty and content of the improvement path will be adjusted in real time based on the creator's learning progress and feedback to ensure effective improvement.

[0079] Based on the above method embodiments, the second embodiment of this application discloses a media creation task distribution system based on multi-dimensional feature matching. The media creation task distribution system based on multi-dimensional feature matching of this application embodiment can implement any of the above-described methods for media creation task distribution based on multi-dimensional feature matching, and the specific working process of each module in the media creation task distribution system based on multi-dimensional feature matching can be referred to the corresponding process in the above method embodiments.

[0080] For ease of understanding, an example is provided below: A media creation task distribution system based on multi-dimensional feature matching includes: The task requirement receiving module is used to receive multimodal task requirements input by the task issuer. Multimodal task requirements include text briefings, reference material sets, and structured constraints. The task feature deconstruction module is used to perform joint semantic understanding and feature deconstruction on multimodal task requirements, and generate a structured task feature vector set, which includes content theme vector, style and aesthetic vector, platform audience vector and constraint target vector. The creator profile building module is used to: extract the creator's historical content and style characteristics based on the creator's historical works, and calculate the historical originality index to build a static capability profile; and evaluate the corresponding creative output submitted by the creator by issuing style transfer and innovation challenge tasks to the creator, and calculate the corresponding current style transfer and innovation index to generate a dynamic capability vector. The task matching and recommendation module is used to: calculate the original adaptation potential score based on the structured task feature vector group and the creator's static ability profile and dynamic ability vector through the adaptation prediction network; perform risk correction on the original adaptation potential score based on the current style transfer innovation index and the historical originality index to obtain the final recommendation potential score; and generate a ranked recommendation list based on the final recommendation potential scores of all creators. The execution process monitoring module is used to extract the feature vector of the corresponding work content in real time during the process of the creator performing the recommended task, and compare it with the features of the reference material set and the preset high-risk copyright library. Based on the comparison results, it triggers graded guidance prompts.

[0081] The third embodiment of this application provides a computer device, which may include a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement a media creation task distribution method based on multi-dimensional feature matching.

[0082] The memory can communicate with the processor via a communication bus, which can be an address bus, a data bus, a control bus, etc.

[0083] Additionally, the memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device.

[0084] Furthermore, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0085] The fourth embodiment of this application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as a media creation task distribution method based on multidimensional feature matching.

[0086] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0087] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A media creation task distribution method based on multi-dimensional feature matching, characterized in that, include: Receive multimodal task requirements input from the task publisher, wherein the multimodal task requirements include text briefings, reference material sets, and structured constraints; The multimodal task requirements are subjected to joint semantic understanding and feature deconstruction to generate a structured task feature vector set, which includes content theme vector, style and aesthetic vector, platform audience vector and constraint target vector. Based on the creator's historical works, extract the corresponding historical content and stylistic characteristics of the creator, calculate the historical originality index, and construct a static capability profile. By issuing style transfer innovation challenge tasks to creators, evaluating the corresponding creative outputs submitted by creators, calculating the corresponding current style transfer innovation index, and generating dynamic capability vectors; Based on the structured task feature vector group, the creator's static ability profile, and the dynamic ability vector, the original adaptation potential score is calculated through the adaptation prediction network. Based on the current style migration innovation index and the historical originality index, the original adaptation potential score is risk-corrected to obtain the final recommendation potential score, and a sorted recommendation list is generated based on the final recommendation potential scores of all creators; During the process of creators performing recommended tasks, feature vectors of the corresponding work content are extracted in real time and compared with the features of the reference material set and the preset high-risk copyright library. Based on the comparison results, graded guidance prompts are triggered.

2. The media creation task distribution method based on multi-dimensional feature matching according to claim 1, characterized in that, The specific steps for performing joint semantic understanding and feature deconstruction on the multimodal task requirements include: A pre-trained multimodal model is used to jointly analyze the text briefing and the reference material set to achieve semantic space alignment; From the aligned semantic information, we deconstruct the semantic features of four dimensions: content theme, style aesthetics, platform audience, and constraint objectives. The semantic features of the four dimensions are mapped to the corresponding vector spaces to generate the content theme vector, the style aesthetic vector, the platform audience vector, and the constraint target vector.

3. The media creation task distribution method based on multi-dimensional feature matching according to claim 1, characterized in that, The specific steps involved in issuing style transfer innovation challenges to creators, evaluating their submitted creative outputs, and calculating the corresponding current style transfer innovation index include: The style transfer innovation challenge task is issued to creators, requiring them to create a new theme based on the style characteristics of specified reference materials; Receive creative outputs submitted by creators in response to the aforementioned challenge task; Calculate the content similarity score between the creative output and the specified reference material in terms of specific content elements; The degree to which the creative output inherits the core stylistic characteristics of the specified reference material is evaluated to obtain a style conformity score; The current style migration innovation index is calculated according to the formula: Current Style Migration Innovation Index = Style Conformity Score × [1 - |Content Similarity Score - Expected Difference Baseline Value|].

4. The media creation task distribution method based on multi-dimensional feature matching according to claim 3, characterized in that, The steps for risk correction of the original fit potential score based on the current style migration innovation index and the historical originality index specifically include: Based on the level of innovation requirements set by the task issuer and the content theme vector and style aesthetic vector of the task, an innovation orientation weight coefficient is generated. The final recommendation potential score is calculated using the formula: Final Recommendation Potential Score = Original Adaptation Potential Score × [Basic Retention Coefficient + Innovation Incentive Coefficient × (Innovation Orientation Weight Coefficient × Current Style Migration Innovation Index + (1 - Innovation Orientation Weight Coefficient) × Historical Originality Index)].

5. The media creation task distribution method based on multi-dimensional feature matching according to claim 1, characterized in that, The steps of comparing the similarity between the reference material set and the preset high-risk copyright database, and triggering graded guidance prompts based on the comparison results, specifically include: Calculate the first real-time similarity between the feature vector of the work content and the feature vector of the reference material set; Calculate the second real-time similarity between the feature vector of the work content and the feature vector of the works in the high-risk copyright database; When the first real-time similarity exceeds the first preset threshold, a prompt message is triggered to guide the creator to adjust the creation to avoid excessive imitation of the reference material; When the second real-time similarity exceeds a second preset threshold that is higher than the first preset threshold, an early warning message is triggered to alert the creator that the current content is at risk of infringement. When both the first real-time similarity and the second real-time similarity are lower than the first preset threshold, it is determined that the current work content is in a safe and original state, the current feature vector is recorded in the safe creation log, and background monitoring continues at the normal frequency.

6. The media creation task distribution method based on multi-dimensional feature matching according to claim 1, characterized in that, The task distribution method further includes: After the task is completed, collect the actual effect data of the task, which includes the publisher's satisfaction score and the originality judgment result of the finished product; The final recommendation potential score is combined with the actual performance data to form a training sample; The fitness prediction network is iteratively optimized using the training samples to reduce the prediction error of the original fitness potential score.

7. The media creation task distribution method based on multi-dimensional feature matching according to claim 1, characterized in that, The task distribution method further includes: When the final recommendation potential scores of all creators are lower than the preset recommendation threshold, the gap between the structured task feature vector group and the available creator ability profile is analyzed to identify the core ability gap. Based on the core capability gaps, corresponding learning materials and micro-tasks are matched in the resource library to generate and push adaptive capability improvement paths to relevant creators.

8. A media creation task distribution system based on multi-dimensional feature matching, characterized in that, Performing the media creation task distribution method based on multidimensional feature matching as described in any one of claims 1 to 7 includes: The task requirement receiving module is used to receive multimodal task requirements input by the task publisher. The multimodal task requirements include text briefings, reference material sets, and structured constraints. The task feature deconstruction module is used to perform joint semantic understanding and feature deconstruction on the multimodal task requirements to generate a structured task feature vector group, which includes content theme vector, style aesthetic vector, platform audience vector and constraint target vector. The creator profile building module is used to: extract the creator's historical content and style characteristics based on the creator's historical works, and calculate the historical originality index to build a static capability profile; and evaluate the corresponding creative output submitted by the creator by issuing style transfer and innovation challenge tasks to the creator, and calculate the corresponding current style transfer and innovation index to generate a dynamic capability vector. The task matching and recommendation module is used to: calculate the original adaptation potential score through an adaptation prediction network based on the structured task feature vector group, the creator's static ability profile, and the dynamic ability vector; perform risk correction on the original adaptation potential score based on the current style transfer innovation index and the historical originality index to obtain the final recommendation potential score; and generate a ranked recommendation list based on the final recommendation potential scores of all creators. The execution process monitoring module is used to extract the feature vector of the corresponding work content in real time during the process of the creator performing the recommended task, and compare it with the features of the reference material set and the preset high-risk copyright library. Based on the comparison results, it triggers graded guidance prompts.

9. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the media creation task distribution method based on multidimensional feature matching as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7, which is a method for distributing media creation tasks based on multidimensional feature matching.

Images