Video work determination method and device, computer equipment, storage medium and program product

Abstract

The invention relates to a video work determination method and device, computer equipment, a storage medium and a program product, and relates to the technical field of artificial intelligence. The method comprises the following steps: acquiring at least two different versions of material contents uploaded for the same video material; combining the material content of each version in the at least two video materials to generate candidate video works of at least two versions; labeling each candidate video work based on a large language model, and generating a video tag corresponding to each time anchor point in the candidate video work; displaying the putting effect score information of each video tag; the putting effect score information is used for the user to select a target video work for putting from the candidate video works. By adopting the method, the video production efficiency and the video delivery effectiveness can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, computer equipment, readable storage medium and computer program product for determining a video work. Background Art

[0002] Video advertising is a common form of digital marketing. Video producers will produce videos based on the characteristics of the brand's products to be marketed, and then place the videos on relevant channels for marketing.

[0003] In the prior art, manual annotation is often performed on entire manually produced videos to form video tags, which provide limited personalized expression for the video content. After the video is produced, it is released to the corresponding channel. After the video is released, user feedback is collected to evaluate the effectiveness of the video advertisement.

[0004] However, as video advertising becomes increasingly important in digital marketing, brands' demand for personalized video content continues to rise. The manual video production method and manual labeling method in existing technologies are both inefficient. The labels for the entire video cannot fully reflect the stage characteristics of the video, and the evaluation process of the delivery effect of the video work is relatively lagging, resulting in low video production efficiency and low effectiveness of video delivery in the current video production process. Summary of the invention

[0005] Based on this, it is necessary to provide a method, device, computer equipment, computer-readable storage medium and computer program product for determining video works that can improve video production efficiency and improve the effectiveness of video delivery in response to the above technical problems.

[0006] In a first aspect, an embodiment of the present application provides a method for determining a video work. The method comprises:

[0007] Obtain at least two different versions of the uploaded video material;

[0008] Combining the material contents of each version of at least two of the video materials to generate at least two versions of candidate video works;

[0009] Annotate each of the candidate video works based on the large language model to generate a video tag corresponding to each time anchor point in the candidate video work; the video tag is used to characterize the characteristics of the video content at each time anchor point in the video work;

[0010] The delivery effect scoring information of each of the video tags is displayed; the delivery effect scoring information is used to characterize the user interaction corresponding to the video work associated with the video tag after delivery, and the delivery effect scoring information is used for the user to select the target video work for delivery from the candidate video works.

[0011] In one embodiment, combining the content of each version of at least two video materials to generate at least two versions of candidate video works includes:

[0012] Combining the material contents of each version of at least two of the video materials to obtain at least two versions of combined video works;

[0013] The corresponding publishing frequency of each of the combined video works is displayed; the publishing frequency is the frequency with which the user has published the same video work, and the same video work is a work with the same video clips and the same order of the video clips as the combined video work among the published video works; the publishing frequency is used for the user to select the at least two versions of candidate video works from each of the combined video works.

[0014] In one embodiment, the tagging of each candidate video work based on the large language model to generate a video tag corresponding to each time anchor point in the candidate video work includes:

[0015] Reasoning each of the candidate video works based on the large language model to obtain a story line corresponding to each of the candidate video works; the story line includes the time anchor points of each story stage in the video work;

[0016] Annotate the image frames corresponding to the candidate video works at the time anchor points to generate the video tags corresponding to the image frames.

[0017] In one embodiment, before the step of displaying the delivery effect score information of each video tag, the method further includes:

[0018] Obtaining user interaction data obtained after the video work is released on at least two release channels; the user interaction data includes at least two evaluation index values ​​corresponding to each of the video tags in the video work after release;

[0019] Based on the weights corresponding to the at least two evaluation index values, a weighted sum is performed on the at least two evaluation index values ​​corresponding to each of the video tags, and based on the result of the weighted sum, the delivery effect score information of each of the video tags is updated.

[0020] In one embodiment, the tagging of each candidate video work based on the large language model to generate a video tag corresponding to each time anchor point in the candidate video work includes:

[0021] In response to a manual labeling operation input for the candidate video work, determining a first label corresponding to each of the candidate video works based on each to-be-selected label recorded in a preset label tree;

[0022] Annotate each of the candidate video works based on the large language model to generate a second label;

[0023] The first tag and the second tag are used as video tags corresponding to each time anchor point.

[0024] In one embodiment, the step of labeling each of the candidate video works based on the large language model to generate a second label includes:

[0025] Identify the video promotion type corresponding to the candidate video work;

[0026] In the case where the video promotion type is a target type, the large language model is trained for a labeling function based on a data set corresponding to a label labeling rule matching the target type to obtain a pre-trained large language model; the target type is a type with a customized label labeling rule;

[0027] Each of the candidate video works is labeled according to the pre-trained large language model to generate a second label corresponding to each of the candidate video works.

[0028] In one embodiment, the method for determining a video work provided in the embodiment of the present application further includes:

[0029] Based on a preset label extraction rule, extracting a preset number of labels to be verified from each of the second labels;

[0030] In response to a manual verification operation input for the label to be verified, the label to be verified is verified to obtain a verification result; the verification result is used to characterize the marking accuracy of the second label.

[0031] In one embodiment, the method for determining a video work provided in the embodiment of the present application further includes:

[0032] Performing data cleaning on at least two types of video labels to obtain labels after data cleaning;

[0033] Based on the attribution relationship between each material in a preset material range and the label after the data cleaning, the label after the data cleaning is screened to obtain the screened label;

[0034] According to the tag type and the tag level in the tag tree, at least two types of tags after data cleaning are combined to obtain a combined tag; and according to the delivery effect scoring information corresponding to each type of the video tags, the combined delivery effect scoring information corresponding to the combined tag is determined;

[0035] The combined delivery effect score information corresponding to each of the combined tags is displayed; the combined delivery effect score information is used for the user to determine the target video work from the candidate video works.

[0036] In a second aspect, the present application also provides a device for determining a video work. The device includes:

[0037] A material content acquisition module, used to acquire at least two different versions of material content uploaded for the same video material;

[0038] A combining module, used for combining the material contents of each version of at least two video materials to generate at least two versions of candidate video works;

[0039] A labeling module, used to label each of the candidate video works based on a large language model, and generate a video tag corresponding to each time anchor point in the candidate video work; the video tag is used to characterize the characteristics of the video content at each time anchor point in the video work;

[0040] The scoring information display module is used to display the delivery effect scoring information of each video tag; the delivery effect scoring information is used to characterize the user interaction corresponding to the video work associated with the video tag after delivery, and the delivery effect scoring information is used for users to select a target video work for delivery from each candidate video work.

[0041] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0042] Obtain at least two different versions of the uploaded video material;

[0043] Combining the material contents of each version of at least two of the video materials to generate at least two versions of candidate video works;

[0044] Annotate each of the candidate video works based on the large language model to generate a video tag corresponding to each time anchor point in the candidate video work; the video tag is used to characterize the characteristics of the video content at each time anchor point in the video work;

[0045] The delivery effect scoring information of each of the video tags is displayed; the delivery effect scoring information is used to characterize the user interaction corresponding to the video work associated with the video tag after delivery, and the delivery effect scoring information is used for the user to select the target video work for delivery from the candidate video works.

[0046] In a fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0047] Obtain at least two different versions of the uploaded video material;

[0048] Combining the material contents of each version of at least two of the video materials to generate at least two versions of candidate video works;

[0049] Annotate each of the candidate video works based on the large language model to generate a video tag corresponding to each time anchor point in the candidate video work; the video tag is used to characterize the characteristics of the video content at each time anchor point in the video work;

[0050] The delivery effect scoring information of each of the video tags is displayed; the delivery effect scoring information is used to characterize the user interaction corresponding to the video work associated with the video tag after delivery, and the delivery effect scoring information is used for the user to select the target video work for delivery from the candidate video works.

[0051] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0052] Obtain at least two different versions of the uploaded video material;

[0053] Combining the material contents of each version of at least two of the video materials to generate at least two versions of candidate video works;

[0054] Annotate each of the candidate video works based on the large language model to generate a video tag corresponding to each time anchor point in the candidate video work; the video tag is used to characterize the characteristics of the video content at each time anchor point in the video work;

[0055] The delivery effect scoring information of each of the video tags is displayed; the delivery effect scoring information is used to characterize the user interaction corresponding to the video work associated with the video tag after delivery, and the delivery effect scoring information is used for the user to select the target video work for delivery from the candidate video works.

[0056] The above-mentioned method, apparatus, computer equipment, storage medium and computer program product for determining video works first obtain different versions of material content by uploading video materials, then perform multi-variable combination of each version of the material content to generate multiple candidate video works, and then annotate the candidate video works based on the large language model to generate video tags, and finally display the delivery effect score information of each video tag for users to select target video works; video production is performed by multi-variable combination, thereby improving the flexibility of video production, and video work annotation is performed by a large language model, thereby improving the richness and accuracy of tags, and by scoring video tags based on the interactive effects of the released videos, users are provided with more dimensional work selection basis, and the delivery effect is displayed in the video production process, thereby improving the efficiency of the video work production process and the availability of video works. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0058] Figure 1 is a flowchart of a method for determining a video work in one embodiment;

[0059] Figure 2 A material source and analysis flow chart of a method for determining a video work in one embodiment;

[0060] Figure 3 A label combination flow chart of a method for determining a video work in one embodiment;

[0061] Figure 4 is a flowchart of a method for determining a video work in another embodiment;

[0062] Figure 5 is a structural block diagram of a device for determining a video work in one embodiment;

[0063] Figure 6 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0065] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0066] It should be noted that the generative artificial intelligence services provided by the large language model involved in this application comply with the requirements of laws and regulations and respect social morality and public order and good customs. The information and data used by the large language model involved in this application are all authorized by the user or fully authorized by all parties.

[0067] In one embodiment, Figure 1 As shown, a method for determining a video work is provided. This embodiment uses the method applied to a terminal as an example for illustration. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0068] S101, obtaining at least two different versions of material content uploaded for the same video material.

[0069] Among them, video materials can be video clips and elements. Elements can be icons (logos), subtitles, background music (BGM), video special effects or transition effects and other multimedia elements that can be used for video production; material content is the specific optional content in each video material.

[0070] Exemplarily, when the video material is a video clip, the material content may include clip 1, clip 2 and clip 3; when the video material is a subtitle, the material content may include subtitle A, subtitle B and subtitle C; when the video material is background music, the material content may include music a, music b and music c.

[0071] S102: combining the content of each version of at least two video materials to generate at least two versions of candidate video works.

[0072] The combination method in step S102 is also called a multi-variable combination method, and the candidate video works refer to multiple versions of video works obtained through multi-variable combination.

[0073] For example, it is necessary to produce a video work that conforms to the "beginning, climax, and ending" framework, which includes three stages: beginning, climax, and ending. For the beginning stage, the video material is a video clip, and the uploaded material content is clip 1, clip 2, and clip 3. For the climax stage, when the video material is a video clip and subtitles, the uploaded material content is clip 4, subtitle A, subtitle B, and subtitle C. For the ending stage, the video material is a video clip and background music, and the uploaded material content is video 5, music a, music b, and music c. At this time, by exhaustively enumerating the optional video materials and combining the possible optional situations, multiple different versions of video works can be obtained.

[0074] Specifically, there are 3 optional clips in the beginning stage, 3 optional subtitles in the climax stage, and 3 optional background music in the ending stage. Theoretically, there are 3×3×3=27 versions of the video work obtained after multi-variable combination; for example, one of the video works can be expressed as: "A video consisting of video 1, video 4 showing subtitle B, and video 5 playing music c".

[0075] S103: annotate each candidate video work based on the large language model to generate a video tag corresponding to each time anchor point in the candidate video work.

[0076] The large language model refers to a neural network model trained through deep learning, which can generate, understand and convert text, images and decorations. In this step, the large language model is used to label the characteristics of the materials at different times in the input video works to form video tags. The video tags are used to characterize the characteristics of the video content at each time anchor point in the video works.

[0077] S104, displaying the delivery effect score information of each video tag.

[0078] Among them, the delivery effect scoring information is used to characterize the user interaction corresponding to the video work associated with the video tag after delivery. Specifically, the user interaction situation can be the interaction score of a certain tag. For example, at the time anchor point corresponding to the "milk" tag, the user has a large number of order confirmation interaction operations. The interaction score of the "milk" tag can be set to a larger score to indicate that the user interaction effect is better at this time point.

[0079] Furthermore, in addition to displaying the delivery effect score information of each video tag, the candidate video works can also be sorted according to the size of the interaction score corresponding to the delivery effect score information, for example, the preferred video works with higher scores are placed at the front of the video list.

[0080] Among them, the delivery effect rating information is used for users to select target video works for delivery from various candidate video works. Specifically, users can select video works that meet their needs based on the interactive rating values ​​or rankings of various candidate video works in the delivery effect rating information. The video production software can respond to the selection operation input by the user for the required video work, select the target video work from various candidate video works, and deliver the target video work to the corresponding channel. Naturally, the video work delivered to the corresponding channel can be used to update the delivery effect rating information of the video tag.

[0081] In the above-mentioned method for determining video works, firstly, different versions of material contents are obtained by uploading video materials, then a plurality of candidate video works are generated by multi-variable combination of the material contents of each version, and then the candidate video works are annotated based on the large language model to generate video tags, and finally the delivery effect score information of each video tag is displayed for users to select target video works; video production is performed by multi-variable combination, which improves the flexibility of video production, and video works are annotated by a large language model, which improves the richness and accuracy of labels, and video tags are scored by the interactive effects of the released videos, which provides users with more dimensional work selection basis, and the delivery effect is displayed in the video production process, which improves the efficiency of the video work production process and the availability of video works.

[0082] In one embodiment, the content of each version of at least two video materials is combined to generate at least two versions of candidate video works, including: combining the content of each version of at least two video materials to obtain at least two versions of combined video works; and displaying the corresponding release frequency of each combined video work.

[0083] The publishing frequency refers to the frequency with which the user has published the same video work, and the same video work refers to a work in the published video works that has the same video clips and the same order of video clips as the combined video work; the publishing frequency is used for the user to select at least two versions of candidate video works from each combined video work.

[0084] For example, if a work using a certain framework has been published by the user in the historical publishing records, the work will be displayed at the front in the combined video works. In addition, the combined video works can also be sorted based on the user's selection habits, so that the user can choose works according to his or her personal preferences.

[0085] For example, to record various elements in the video creation process, the creator only needs to use the video production software to create the pure version of the video clip and elements (such as logos, subtitles, background music, special effects and transitions, etc.). Both video clips and elements are allowed to upload multiple versions. After cross-product combination, multiple versions of the finished video will be produced, and different combinations will be evaluated by release frequency or personal preference through multivariate testing to obtain the best version combination. In this way, the element differences of each video version can help more accurately identify key tags, and can also help creators analyze which specific features and element combinations perform best in different situations, greatly improving the reproducibility of the creation effect.

[0086] In this embodiment, the content of each version of at least two video materials is first combined to generate multiple candidate video works, and then the publishing frequency of the published works is recorded to facilitate users to select the combined video works. By combining the user's historical publishing data, the accuracy and personalization of video work selection are improved, the user's video production experience is optimized, and the efficiency of video publishing is improved.

[0087] In one embodiment, each candidate video work is annotated based on a large language model to generate a video label corresponding to each time anchor point in the candidate video work, including: reasoning about each candidate video work based on the large language model to obtain a story line corresponding to each candidate video work; annotating an image frame in the corresponding candidate video work for each time anchor point to generate a video label corresponding to each image frame.

[0088] The storyline includes the time anchor points of each story stage in the video work. Generally, the time anchor point can be a moment, multiple moments or a certain time period in the video, and accordingly, the image frame can also be a set of image frames at a certain moment and image frames in a certain time period.

[0089] For example, a large language model is used to infer the storyline of the entire clip of the candidate video work. Each screen feature needs to be deduced through the context of the storyline to determine its validity and rationality. The screen features promoted by the storyline will be marked as important, which will affect their weight in the analysis process. At the same time, through the in-depth sentiment analysis (text reasoning) of the traces left by user participation and interaction (such as forwarding, liking, and commenting) in the advertising video landing page, the influence relationship with the screen feature label is inferred to ensure that the video content can better resonate with the audience. This process not only optimizes the video structure, but also provides a scientific basis for subsequent content creation.

[0090] In this embodiment, firstly, the corresponding story line is obtained by inferring each candidate video work based on the large language model, and then the image frames of each time anchor point in the candidate video work are annotated to generate video tags; through the reasoning ability of the large language model and the image frame annotation technology, the structured and comprehensible nature of the video content is improved, which facilitates the use of the annotated video tags to guide the production of products.

[0091] In one embodiment, before the step of displaying the delivery effect score information of each video tag, the method further includes:

[0092] Obtain user interaction data after the video work is released on at least two distribution channels;

[0093] Based on the weights corresponding to the at least two evaluation index values, a weighted sum is performed on the at least two evaluation index values ​​corresponding to each video tag, and based on the result of the weighted sum, the delivery effect score information of each video tag is updated.

[0094] The user interaction data includes at least two evaluation index values ​​corresponding to each video tag in the video work after delivery. The evaluation index value can be the value of indicators such as the material customer acquisition score, material quality score, and material 3s play rate score.

[0095] like Figure 2 As shown in the figure, the sources of label combination analysis include new media channels and information streaming media platforms. The analysis results are obtained based on not only the image features or manual label information of the video content itself, but also the specific data performance of the video in the advertising market. Figure 2 In the analysis, the material effect data refers to the user interaction effect of each material in the video after the video is released, such as click-through rate, conversion rate, and viewing time; the material dimension basic data refers to the background information and contextual data that need to be considered when analyzing the material; for example, whether the release channel is social media or search engine; the age, gender, geographic location and other information of the users; the specific time of release, such as weekdays or holidays.

[0096] Considering the performance differences of data in different channels after the advertising creative is released, this embodiment integrates the creative data of multiple channels for comprehensive analysis. By aggregating the data of each platform, potential misjudgments can be identified in a timely manner (for example, in an advertising video that needs to be watched for 15 seconds before it can be skipped, its play rate weight is weakened, and the weight of clicking on the ad after the skipping time is strengthened - after watching the ad for a long time, you are still interested in the video content), ensuring the accuracy of the advertising creative data in reflecting the label combination weight.

[0097] The calculation and update process of the delivery effect scoring information can be as follows: the analysis adopts the material scoring method, taking the advertising promotion of video materials on the information flow advertising platform as an example.

[0098] The scoring formula can be: material customer acquisition score × weight 1 + material quality score × weight 2 + material 3s play rate × weight 3; among them, the material customer acquisition score is a score function reflecting the customer acquisition volume; the material quality score is a score function reflecting the retention rate, such as the proportion of users who continue to watch the content within a certain period of time (usually the next day, next week, etc.); the material 3s play rate score is a score function reflecting the 3s play rate, such as the proportion of users who watched the video for at least 3 seconds after the first play to the total number of users who played the video. In addition, the evaluation indicator can also be the average daily customer acquisition score on the effective day.

[0099] In addition, you can also score and calculate the indicators of customer acquisition, average daily customer acquisition, retention rate, and 3s playback rate; normalize them to a value between 0 and 1 and then multiply them by 10 to keep two decimal places to become a number within 10; for example, the normalization method can be Min-Max Scaling, and the formula can be:

[0100]

[0101] In addition, in order to ensure that the advertising content always meets the market trend, the present invention implements a real-time update mechanism. The historical tags dynamically adjust the hot weights based on the frequency of occurrence and the interval duration. New tags participate in PID competition. The real-time tag library is combined with the real-time delivery data, and is continuously updated and displayed in the element tag combination in real time. Through this mechanism, it is possible to quickly respond to changes in user behavior and market demand, thereby maximizing the advertising effect and continuously optimizing the creative strategy.

[0102] In this embodiment, user interaction data obtained after the video work is released on at least two delivery channels is first obtained, and then based on the weights corresponding to the at least two evaluation index values, a weighted sum is performed on the at least two evaluation index values ​​corresponding to each video tag, and then the delivery effect scoring information of each video tag is updated according to the result of the weighted sum, and finally the updated delivery effect scoring information of each video tag is displayed; through the comprehensive evaluation of user interaction data and weighted summation, the accuracy and timeliness of the delivery effect scoring are improved.

[0103] In one embodiment, each candidate video work is labeled based on a large language model to generate a video label corresponding to each time anchor point in the candidate video work, including: in response to a manual labeling operation input for the candidate video work, determining a first label corresponding to each candidate video work based on each candidate label recorded in a preset label tree; labeling each candidate video work based on the large language model to generate a second label; and using the first label and the second label as the video label corresponding to each time anchor point.

[0104] Among them, the first label is the manual label, and the second label is the AI ​​label.

[0105] Manual tags refer to personalized tags that creators mark video content with. The sources of tags include custom tags (any content can be tagged) and preset tag trees (usually general tags of an enterprise, covering all video content creation, advertising placement, content acceptance - where users go after clicking on an ad, data classification, and other analysis scenarios);

[0106] AI labeling means that without considering the manually labeled content, only from the content expression of the video itself, deep learning algorithms (such as convolutional neural network CNN), ASR (Automatic Speech Recognition) audio recognition, OCR (Optical Character Recognition) text recognition, VGGFace facial recognition and other technologies are used to perform frame-level analysis on the video content, extract key phrases (audio text, narration & dialogue text, screen text, etc.), character information (identity image, emotional expression, character behavior and interaction, etc.), scene events (specific events or scenes related to text & characters, such as people, vehicles, etc.), camera expression (camera push, pull, pan, special effects, sound effects and other expression information) in the video, and mark the corresponding labels. This type of tag does not require a pre-set tag tree. It is marked after LLM (Large Language Model) reasoning. When there are similar tags but different words, they need to be abstracted and merged. Therefore, the AI ​​tags still have a hierarchical structure, and as the number of videos accumulates and time passes, the latest version of the tag tree will be automatically updated (the popularity is determined by the frequency of tag appearance and the length of the interval, and tags that no longer appear gradually weaken to dormancy). After the latest tag is added, it will automatically participate in the PID (Personal Identification) competition. The heat weight of the tag is reflected at the analysis level.

[0107] The difference between manual labeling and AI labeling is that manual labeling requires human participation, and the relationship between the video content itself and the label is difficult to verify, resulting in wrong labeling, random labeling, and missing labeling. The depth of analysis depends on the sophistication of the label tree. The update of the label tree requires the participation of multiple parties (director design, pitcher, client - for example, APP, etc.), and the timeliness is low. The more sophisticated the labeling, the more labels there are, and the greater the pressure on label management.

[0108] This embodiment combines manual labels and AI labels. Manual labels can provide more metaphorical expressions or directional guidance in addition to content (such as before advertising is placed), and AI labels can supplement the content expressed in the picture in addition to manual labels. The combination of the two enables the uniqueness of manual labels and the scale advantage of AI labels to complement each other, thereby improving the flexibility and accuracy of the overall labeling system.

[0109] In one embodiment, each candidate video work is labeled based on a large language model to generate a second label, including: identifying a video promotion type corresponding to the candidate video work; when the video promotion type is a target type, based on a data set corresponding to a label labeling rule matching the target type, training the labeling function of the large language model to obtain a pre-trained large language model; labeling each candidate video work according to the pre-trained large language model to generate a second label corresponding to each candidate video work.

[0110] Among them, there are two types of AI tags. One is multimodal content understanding based on the LLM large model, which understands and extracts labels from the picture features and text information in the video content itself. There is no need to preset the label tree in advance, and the entire process is handled by the LLM large model. This method has low time cost (no training required), but the reasoning cost is extremely high; the other is to manually define the label content and feature range, and provide label features (usually video clips or pictures with label features) for algorithm model training (usually at least 20+ samples are required for each label). The trained model can infer and identify the features of the picture when the video is uploaded and automatically label it. This method has high time cost, but the reasoning cost is very low.

[0111] Among them, the target type is a type with customized label annotation rules, specifically, the target type can be the type of APP promotion products. Exemplarily, in addition to this, for specific APP promotion products, specific APP screens (such as APP homepage, personal page, voice room page) and public content involved in the video content (such as MOBA game and gun battle game screens, or social media application media information release screens) are also collected, and the open source Labelimg graphic image annotation tool is used for annotation training. The model uses the yolo series. When specific screen features that have a direct impact on the promotion effect of video content are identified, the video content is marked to address the problem of inaccurate recognition of the LLM large model that has not been trained for this specific content.

[0112] In this embodiment, the video promotion type corresponding to the candidate video work is first identified, and then, when the video promotion type is the target type, the labeling function of the large language model is trained based on the data set corresponding to the label labeling rule matching the target type, thereby obtaining a pre-trained large language model, and finally, each candidate video work is labeled according to the pre-trained large language model to generate a corresponding second label; through accurate video promotion type identification and large language model training, the accuracy of label generation is improved, which facilitates the labeling and production of video works using the second label.

[0113] In one embodiment, the method for determining the video work provided by the embodiment of the present application also includes: based on a preset label extraction rule, extracting a preset number of labels to be verified from each second label; in response to a manual verification operation input for the labels to be verified, verifying the labels to be verified and obtaining a verification result.

[0114] The verification result is used to characterize the marking accuracy of the second label.

[0115] Exemplarily, after training, manual verification is performed and 400 samples are sampled. The sampling process may include: (1) randomly sampling 1,000 video data from the set of second labels; (2) saving the labeled picture frames from the above video data; (3) removing duplicate pictures from the video sample set based on picture frame similarity, with no more than 30 pictures in each category, to obtain a sample set. The sampling results may include the number and total number of labels that have been verified to be correct in each category, as well as the accuracy of the verification. For example, in a sampling result, the overall accuracy is about 93%, and then the bad cases (samples that do not reach 80%) are continuously corrected according to the passing line of 80% accuracy to maintain and optimize their accuracy.

[0116] In this embodiment, firstly, based on the preset label extraction rules, a preset number of labels to be verified are extracted from each second label, and then in response to the manual verification operation input for the labels to be verified, the labels to be verified are verified to obtain the verification results, and finally the verification results are used to characterize the labeling accuracy of the second labels. By combining systematic label extraction with manual verification, the accuracy and reliability of label labeling are improved.

[0117] In one embodiment, the method for determining a video work provided by the embodiment of the present application further includes:

[0118] Performing data cleaning on at least two types of video labels to obtain labels after data cleaning;

[0119] Based on the attribution relationship between each material in the preset material range and the label after data cleaning, the label after data cleaning is filtered to obtain the filtered label;

[0120] According to the tag type and the tag level in the tag tree, at least two types of tags after data cleaning are combined to obtain a combined tag; and according to the delivery effect scoring information corresponding to each type of video tag, the combined delivery effect scoring information corresponding to the combined tag is determined;

[0121] The combined delivery effect score information corresponding to each combined label is displayed; the combined delivery effect score information is used for users to determine the target video work from various candidate video works.

[0122] This embodiment provides an efficient combined tag analysis mechanism, which comprehensively calculates the key feature tags of the story line, the high conversion feature tags of creative data, and the multivariate test law tags to form the best tag combination. This analysis tool can be quickly provided to creators so that they can clarify the key elements and structures during the production process and improve the accuracy and effectiveness of the creative direction.

[0123] The label screening and combination process is as follows Figure 3 As shown, before combining various types of labels, the data needs to be screened first, and then the labels of the screened high-quality data are combined. Because the data dimensions involved in the combination analysis of labels are large and the data volume is large, the data analysis efficiency is low. Therefore, the data needs to be cleaned and screened to improve the data analysis efficiency.

[0124] The logic of cleaning can be to determine whether the values ​​of the interactive data indicators corresponding to each tag meet the preset requirements; you can filter a single tag or filter batch tags. For example, the indicator requirements for filtering a single tag can be: single plan day consumption > x, single plan day total customer acquisition > x, single plan day retention > x, single plan day exposure > x, single plan day 3s playback volume > x, single plan day total playback volume > x; the indicator requirements for filtering batch tags can be: total customer acquisition within the statistical period > x, retention within the statistical period > x, exposure within the statistical period > x, 3s playback volume within the statistical period > x, total playback volume within the statistical period > x, effective delivery days within the statistical period > x. Among them, "x" is the project business value, which can be defined by the project.

[0125] In addition, cleaning can also be based on other methods, such as screening labels based on dimensions such as material type, media type, operating system, and delivery target, or screening labels based on interference factors such as account starting factors, bids, and target groups.

[0126] Finally, you can also verify the effect after screening. For example, use the data from a certain stage to derive the ranking of material scores and label combination scores, and then calculate whether the proportion of material customer acquisition under the topic10 label combination is consistent with the label ranking. If they are consistent, it means that the cleaning result is good.

[0127] After data cleaning is completed, it is necessary to filter out labels for attribution and then combine the labels, such as Figure 3 As shown, the combination process needs to be carried out in a classified manner. For example, for manual labels, they need to be combined hierarchically based on the hierarchy of the label tree, for example, each first-level label and each leaf label needs to be combined separately.

[0128] In this embodiment, data cleaning is first performed on at least two types of video tags to obtain cleaned tags, and then the cleaned tags are screened based on the attribution relationship between each material in a preset material range and the tags after data cleaning to obtain the screened tags, and finally the screened tags are combined according to the tag type and the tag level in the tag tree, and the delivery effect scoring information corresponding to the combined tags is determined; this embodiment improves the accuracy of tags and the combination efficiency of the tag combination process through the orderly processing of data cleaning and tag screening.

[0129] In another embodiment, Figure 4 As shown, a method for determining a video work is provided, comprising the following steps:

[0130] S401, obtaining at least two different versions of material content uploaded for the same video material.

[0131] S402, combining the content of each version of at least two video materials to obtain at least two versions of combined video works.

[0132] S403, displaying the release frequencies corresponding to the combined video works, so that the user can select at least two versions of candidate video works from the combined video works.

[0133] S404, inferring each candidate video work based on the large language model to obtain a story line corresponding to each candidate video work.

[0134] S405 , annotating image frames of each time anchor point in the corresponding candidate video work, and generating video tags corresponding to each image frame.

[0135] S406, obtaining user interaction data after the video work is released in at least two release channels.

[0136] S407, based on the weights corresponding to the at least two evaluation index values, weighted sum is performed on the at least two evaluation index values ​​corresponding to each video tag, and based on the result of the weighted sum, the delivery effect score information of each video tag is updated.

[0137] S408, displaying the delivery effect score information of each video tag.

[0138] It should be noted that the specific limitations of the above steps can be found in the specific limitations of the method for determining a video work above, which will not be repeated here.

[0139] As video advertising becomes increasingly important in digital marketing, brands are increasingly demanding personalized video content. However, existing technologies are clearly insufficient in meeting this demand, especially in the application of personalized tags. Although existing tag analysis technologies can provide a certain degree of personalized expression for video content, these unique tags are often treated as special data and filtered out in large-scale data analysis, making it difficult to form a systematic video production methodology.

[0140] Some advertising optimization tools currently on the market, such as label analysis tools on media platforms, all use rule-based label generation technology and combined analysis methods. These solutions guide the creative direction of videos through picture feature analysis and data statistical models. However, they only process surface information on the screen, such as whether a character appears, gender, and the type of object. They lack an in-depth analysis of the inherent picture features of the video and the user's viewing psychology, and cannot help creators accurately determine when the character appears to best fit the content expression, or what objects appearing in the screen can better promote the development of the plot.

[0141] In addition, some platforms try to use sentiment analysis tools to identify emotional elements and optimize creativity, but their ability to capture and analyze key screen elements is still limited, making it difficult to accurately match audience needs. At the same time, there are solutions on the market that rely on manual updates of personalized tags, which are insufficient in terms of real-time and efficiency. Manual updates require a large number of samples for training, and it is impossible to exhaust all key information for generalized words and fuzzy words. Therefore, there are obvious defects in quickly responding to changing user behaviors and media trends.

[0142] Based on this, an embodiment of the present application provides a method for determining video works, also known as an application method of label combination analysis in video creative guidance, which breaks through the limitations of the existing technology and realizes efficient and accurate analysis and guidance in the process of video creative production and optimization. It adopts CNN (Convolutional Neural Network) technology and LLM (Large Language Model) prompt enhancement strategy, identifies and generates personalized labels, and combines multiple elements to generate multiple versions for multivariate testing, so as to find the best combination that best matches the user motivation rules, and continuously optimizes the conversion rate of advertising creativity through data-driven methods.

[0143] Specifically, existing technologies face four major deficiencies in label analysis and guidance:

[0144] (1) The contradiction between individuality and rules: Although individualized labels (including labels based on the result of image content recognition and labels subjectively marked by the creator for the video) can intuitively reflect the creativity of the video content, these unique features are often regarded as special data and filtered out in large-scale analysis, making it difficult to transform them into mature methodologies. This overly unique feature makes it impossible to form a replicable formula, such as clear rules for which props to use in which occasion, thus limiting the large-scale promotion and application of individualized content.

[0145] (2) Limited surface data analysis: Data-assisted label analysis after advertising is often superficial, lacking in-depth exploration of the relationship between the visual features of the video content and the motivations of user behavior. This leads to one-sided system analysis results, which still rely on the personal experience and understanding of manual analysts, involve luck, and cannot effectively use data to verify conclusions, making it difficult to promote its analysis methods to the team or industry level.

[0146] (3) Information loss during the creative process: During the video production process, even if there is data after the release, this information usually only serves as a directional guide. It is difficult to clearly identify which screen elements are the key to attracting the audience and the best time for these elements to appear during video playback. This lack of information seriously affects the effectiveness of creative guidance.

[0147] (4) Real-time and dynamic update challenges: User behavior and media gameplay trends are constantly evolving. Relying on regular manual updates and optimization of personalized tags is insufficient, making it impossible to keep video content in sync with user needs, thus missing out on traffic dividends.

[0148] In response to the above-mentioned defects, the method for determining video works provided in the embodiments of the present application involves the following contents: (1) distinguishing and combining manual labels and AI labels; (2) recording the creative process and conducting multivariate testing; (3) LLM reasoning story line and refining key features; (4) comprehensive analysis of all-channel creative data to avoid misjudgment; (5) combining label results to assist creators in analysis and guide creation; (6) real-time matching of creative labels and delivery data, and dynamic updating.

[0149] The method for determining video works provided in the embodiment of the present application records the creative process information of video clips and different element combinations, and helps to refine and infer labels through A / B testing (group comparison testing) and multivariate testing to improve accuracy. In this regard, different clip contents (such as models, drop plates, flower font styles, transitions, filters, etc.) and typesetting layouts are combined and tested, and reasonable story line creation suggestions are generated through LLM reasoning. These suggestions include creative production framework formulas and sequences (such as a framework formula of "introduction (Hook) + development (Build) + climax (Climax) + ending (Conclusion)"), essential picture features to promote story expression, and key elements that match user preferences. Finally, these achievements are presented in the form of visual label combinations to achieve more intuitive and efficient data applications.

[0150] Even if the same batch of video materials show different data performances in different advertising channels (such as short video platforms, social media platforms and other channels of different manufacturers), the method for determining video works provided by the embodiment of the present application can integrate multi-channel advertising data and original video creation information to achieve information sharing of different channel archives. This feature allows for comprehensive analysis based on omni-channel creative tags when optimizing video creatives for specific channels, avoiding misjudgments caused by data information barriers of a single channel, thereby improving the pertinence of advertising creative optimization.

[0151] By dynamically updating and capturing user behavior in real time, the method for determining video works provided in the embodiment of the present application ensures that the video content is closely integrated with market trends, and quickly presents the optimal label combination of video creative frameworks with high conversion rates, thereby achieving more efficient and accurate advertising creative production and optimization, and providing scientific guidance for the next creative production.

[0152] The method for determining a video work provided in the embodiment of the present application mainly includes the following contents:

[0153] (1) AI tag generation and application: This invention uses convolutional neural networks (CNN), ASR, OCR and other technologies to analyze and refine picture features frame by frame, and uses a large language model (LLM) to understand the logic of the video storyline, and combines the identified picture features with the storyline context information to accurately capture key features. The automatically generated AI tags are sorted by PID weight, and after advertising is placed, the weights are dynamically adjusted again based on creative data. The innovation of this process is to improve the efficiency of tag generation using AI technology, as well as the accuracy of matching video content with user motivations, while protecting the effectiveness of the tag generation for verification and evaluation, ensuring that the value level of each tag in the analysis scenario is traceable.

[0154] (2) Multivariate testing and creative process recording: Record the multivariate in the creative process, test the performance of different element combinations, and integrate data from multiple channels for comprehensive analysis to avoid misjudgment and optimize the creative process and result feedback. Protect the real-time calibration of test results and label weights, and the effectiveness of data-driven labels in video creative guidance applications.

[0155] (3) Real-time dynamic update mechanism: The present invention provides a method for real-time matching of creative tags and delivery data, ensuring that the advertising content is consistent with market demand through continuous updates. This dynamic update mechanism and its algorithm implementation will be an important part of the scope of protection.

[0156] (4) Method for forming the best label combination: The present invention provides a method for forming the best label combination by combining the key feature labels of the story line, the high conversion feature labels of the creative data, and the multivariate test regularity labels, so as to ensure that the creators can quickly obtain effective creative guidance.

[0157] Compared with the above-mentioned prior art, the method for determining video works provided by the embodiment of the present application has the following significant advantages:

[0158] (1) Improve the efficiency and accuracy of label generation: Through the automatic generation of AI labels, the reliance on human resources is reduced, while the subjective bias that may exist in manual labeling is avoided. This efficiency improvement enables the rapid generation of labels that meet market needs in large-scale advertising creation, thereby accelerating the creation process.

[0159] (2) Deep analysis and matching of user emotions: Using large language models and convolutional neural network technology to deeply analyze video content, the advertising creativity can better match user emotions. This technical advantage ensures that the content can more effectively attract the target audience and improve the audience's engagement and satisfaction.

[0160] (3) Scientific creative guidance: By recording the creative process and conducting multivariate testing, comprehensive omnichannel data analysis can promptly identify potential misjudgments, and labels and their weights are dynamically calibrated in real time based on test results rather than relying solely on experience. This data-based feedback mechanism improves the success rate of creation and provides a replicable success model for subsequent projects.

[0161] (4) Effective combination strategy: The best label combination is formed by combining the key feature labels of the story line, the high conversion feature labels of creative data, and the multivariate test rule labels. This combination strategy not only helps creators clarify the creative direction, but also guides them to make more targeted designs in visual presentation, story narration, and emotional expression.

[0162] In summary, the method for determining video works provided in the embodiments of the present application surpasses the existing technology in many aspects such as efficiency, accuracy, and flexibility, and provides a comprehensive, scientific, and dynamically adaptable new solution for advertising creation.

[0163] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0164] Based on the same inventive concept, the embodiment of the present application also provides a device for determining a video work for implementing the method for determining a video work involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in the embodiments of one or more video work determination devices provided below can refer to the limitations of the video work determination method above, and will not be repeated here.

[0165] In one embodiment, Figure 5 As shown, a device for determining a video work is provided, comprising: a material content acquisition module 501, a combination module 502, a marking module 503, and a rating information display module 504, wherein:

[0166] The material content acquisition module 501 is used to acquire at least two different versions of material content uploaded for the same video material;

[0167] A combination module 502 is used to combine the content of each version of at least two video materials to generate at least two versions of candidate video works;

[0168] The tagging module 503 is used to tag each candidate video work based on the large language model, and generate a video tag corresponding to each time anchor point in the candidate video work; the video tag is used to characterize the characteristics of the video content at each time anchor point in the video work;

[0169] The scoring information display module 504 is used to display the delivery effect scoring information of each video tag; the delivery effect scoring information is used to characterize the user interaction corresponding to the video work associated with the video tag after delivery, and the delivery effect scoring information is used for users to select the target video work for delivery from various candidate video works.

[0170] In one embodiment, the device for determining a video work is used to: combine the content of each version of at least two video materials to obtain at least two versions of combined video works; display the publishing frequency corresponding to each combined video work; the publishing frequency is the frequency with which the user has published the same video work, and the same video work is a work in the published video works that has the same video clips and the same order of the video clips as the combined video work; the publishing frequency is used for the user to select at least two versions of candidate video works from each combined video work.

[0171] In one embodiment, the device for determining a video work is used to: infer each candidate video work based on a large language model to obtain a story line corresponding to each candidate video work; the story line includes time anchor points of each story stage in the video work; annotate the image frames in the corresponding candidate video work for each time anchor point, and generate video tags corresponding to each image frame.

[0172] In one embodiment, the device for determining a video work is used to: obtain user interaction data obtained after the video work is delivered in at least two delivery channels; the user interaction data includes at least two evaluation index values ​​corresponding to each video tag in the video work after delivery; based on the weights corresponding to the at least two evaluation index values, perform weighted summation on the at least two evaluation index values ​​corresponding to each video tag, and update the delivery effect score information of each video tag based on the result of the weighted summation.

[0173] In one embodiment, the video work determination device is used to: in response to a manual labeling operation input for a candidate video work, determine a first label corresponding to each candidate video work based on each candidate label recorded in a preset label tree; annotate each candidate video work based on a large language model to generate a second label; and use the first label and the second label as video labels corresponding to each time anchor point.

[0174] In one embodiment, a device for determining a video work is used to: identify a video promotion type corresponding to a candidate video work; when the video promotion type is a target type, based on a data set corresponding to a label annotation rule matching the target type, train a labeling function of a large language model to obtain a pre-trained large language model; the target type is a type with a customized label annotation rule; and label each candidate video work according to the pre-trained large language model to generate a second label corresponding to each candidate video work.

[0175] Each module in the above-mentioned video work determination device can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0176] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be realized through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a method for determining a video work is implemented. The display unit of the computer device is used to form a visually visible picture, which can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device shell, or an external keyboard, touchpad or mouse.

[0177] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0178] In one embodiment, a computer device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the steps in the above method embodiments are implemented.

[0179] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0180] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0181] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0182] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0183] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for determining a video work, characterized in that: The method comprises: Obtain at least two different versions of the uploaded video material; Combining the material contents of each version of at least two of the video materials to generate at least two versions of candidate video works; Annotate each of the candidate video works based on the large language model to generate a video tag corresponding to each time anchor point in the candidate video work; the video tag is used to characterize the characteristics of the video content at each time anchor point in the video work; The delivery effect scoring information of each of the video tags is displayed; the delivery effect scoring information is used to characterize the user interaction corresponding to the video work associated with the video tag after delivery, and the delivery effect scoring information is used for the user to select the target video work for delivery from the candidate video works.

2. The method according to claim 1, characterized in that: The step of combining the material contents of each version of at least two video materials to generate at least two versions of candidate video works includes: Combining the material contents of each version of at least two of the video materials to obtain at least two versions of combined video works; The corresponding publishing frequency of each of the combined video works is displayed; the publishing frequency is the frequency with which the user has published the same video work, and the same video work is a work with the same video clips and the same order of the video clips as the combined video work among the published video works; the publishing frequency is used for the user to select the at least two versions of candidate video works from each of the combined video works.

3. The method according to claim 1, characterized in that The step of labeling each of the candidate video works based on the large language model to generate a video tag corresponding to each time anchor point in the candidate video work includes: Reasoning each of the candidate video works based on the large language model to obtain a story line corresponding to each of the candidate video works; the story line includes the time anchor points of each story stage in the video work; Annotate the image frames corresponding to the candidate video works at the time anchor points to generate the video tags corresponding to the image frames.

4. The method according to claim 1, characterized in that Before the step of displaying the delivery effect score information of each video tag, the method further includes: Obtaining user interaction data obtained after the video work is released on at least two release channels; the user interaction data includes at least two evaluation index values ​​corresponding to each of the video tags in the video work after release; Based on the weights corresponding to the at least two evaluation index values, a weighted sum is performed on the at least two evaluation index values ​​corresponding to each of the video tags, and based on the result of the weighted sum, the delivery effect score information of each of the video tags is updated.

5. The method according to claim 1, characterized in that: The step of labeling each of the candidate video works based on the large language model to generate a video tag corresponding to each time anchor point in the candidate video work includes: In response to a manual labeling operation input for the candidate video work, determining a first label corresponding to each of the candidate video works based on each to-be-selected label recorded in a preset label tree; Annotate each of the candidate video works based on the large language model to generate a second label; The first tag and the second tag are used as video tags corresponding to each time anchor point.

6. The method according to claim 5, characterized in that The step of labeling each of the candidate video works based on the large language model to generate a second label includes: Identify the video promotion type corresponding to the candidate video work; In the case where the video promotion type is a target type, the large language model is trained for a labeling function based on a data set corresponding to a label labeling rule matching the target type to obtain a pre-trained large language model; the target type is a type with a customized label labeling rule; Each of the candidate video works is labeled according to the pre-trained large language model to generate a second label corresponding to each of the candidate video works.

7. A device for determining a video work, characterized in that: The device comprises: A material content acquisition module, used to acquire at least two different versions of material content uploaded for the same video material; A combining module, used for combining the material contents of each version of at least two video materials to generate at least two versions of candidate video works; A labeling module, used to label each of the candidate video works based on a large language model, and generate a video tag corresponding to each time anchor point in the candidate video work; the video tag is used to characterize the characteristics of the video content at each time anchor point in the video work; The scoring information display module is used to display the delivery effect scoring information of each video tag; the delivery effect scoring information is used to characterize the user interaction corresponding to the video work associated with the video tag after delivery, and the delivery effect scoring information is used for users to select a target video work for delivery from each candidate video work.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images