Short video promotion optimization method and system

Abstract

This application discloses a method and system for optimizing short video promotion, relating to the field of short video promotion optimization technology. It aims to address the difficulty of existing short video promotion methods in accurately conveying merchant intent and reaching target users when dealing with subtle elements in video content that may lead to different user interpretations. The method includes: acquiring content data of the video to be promoted; identifying multi-meaning elements in the content data to determine at least one potential multi-meaning element in the video to be promoted; determining the contextualized misinterpretation risk assessment result of the potential multi-meaning element based on the potential multi-meaning element and a preset user context model; making a decision between a content modification strategy and an audience selection strategy based on the contextualized misinterpretation risk assessment result, pre-stored promotion constraints, and a content modification cost model to generate an initial intervention strategy; executing the initial intervention strategy; and optimizing the promotion process based on user feedback data obtained after execution.

Description

Technical Field

[0001] This application relates to the field of short video promotion and optimization technology, and in particular to a short video promotion and optimization method and system. Background Technology

[0002] As a mainstream promotional medium, the effectiveness of short videos largely depends on whether the content can accurately convey the merchant's intent and reach the target audience. However, existing promotional methods often struggle to handle subtle elements in video content that may elicit different interpretations from users. These ambiguous elements can lead to biased understandings of the video's core promotional message, thus impacting promotional effectiveness. Especially given the ever-changing viewing contexts and personal preferences of users, the same video clip may be interpreted differently by different users, making it difficult to effectively optimize the promotional system. Summary of the Invention

[0003] This application provides a method and system for optimizing short video promotion, aiming to solve the problems of existing short video promotion methods in accurately conveying merchants' intentions and reaching target users when dealing with subtle elements in video content that may lead to different interpretations by users, and in effectively deciding on content modification strategies and audience screening strategies when facing high-risk misinterpretation segments.

[0004] Firstly, to address the aforementioned technical problems, this invention provides a short video promotion optimization method, comprising: acquiring content data of a video to be promoted, the content data including visual data, auditory data, and text data; identifying multi-meaning elements in the content data to determine at least one potential multi-meaning element in the video to be promoted, the potential multi-meaning element being used to characterize visual or auditory segments that may cause users to deviate from the core promotional intent; determining a contextualized misinterpretation risk assessment result for the potential multi-meaning element based on the potential multi-meaning element and a preset user context model, the contextualized misinterpretation risk assessment result being used to characterize the likelihood and direction of the potential multi-meaning element causing comprehension deviation in a specific user viewing context; making a decision between a content modification strategy and an audience selection strategy based on the contextualized misinterpretation risk assessment result, pre-stored promotion constraints, and a content modification cost model to generate an initial intervention strategy, the content modification strategy including adjustments to the potential multi-meaning elements in the video to be promoted, and the audience selection strategy including the selection of a test user group; executing the initial intervention strategy, and optimizing the promotion process based on user feedback data obtained after execution.

[0005] Secondly, this application provides a short video promotion optimization system, which includes: an acquisition unit for acquiring content data of a video to be promoted, the content data including visual data, auditory data, and text data; a first determination unit for identifying multi-meaning elements in the content data to determine at least one potential multi-meaning element in the video to be promoted, the potential multi-meaning element being used to characterize visual or auditory segments that may cause users to deviate from the core promotion intention; a second determination unit for determining the contextualized misinterpretation risk assessment result of the potential multi-meaning element based on the potential multi-meaning element and a preset user context model, the contextualized misinterpretation risk assessment result being used to characterize the possibility and direction of the potential multi-meaning element causing comprehension deviation in a specific user viewing context; a generation unit for making a decision between a content modification strategy and an audience selection strategy based on the contextualized misinterpretation risk assessment result, pre-stored promotion constraints, and a content modification cost model, to generate an initial intervention strategy, the content modification strategy including adjustments to the potential multi-meaning elements in the video to be promoted, and the audience selection strategy including the selection of a test user group; and an execution unit for executing the initial intervention strategy and optimizing the promotion process based on user feedback data obtained after execution.

[0006] This application has at least the following beneficial effects: The short video promotion optimization method disclosed in this application, by acquiring multimodal content data of the video to be promoted and identifying its multi-meaning elements, can accurately capture visual or auditory segments in the video that may cause misunderstandings among users. Based on this, combined with a pre-set user context model, a contextualized misinterpretation risk assessment is conducted on these potential multi-meaning elements, thereby quantifying the likelihood and direction of misunderstandings caused by them in a specific user viewing context. Furthermore, based on the misinterpretation risk assessment results, promotion constraints, and a content modification cost model, this application intelligently decides whether to modify the content or screen the audience to generate an initial intervention strategy. Finally, by executing the initial intervention strategy and continuously optimizing based on user feedback data, closed-loop management of the promotion process is achieved. Through the above technical solution, this application effectively solves the technical problems in the prior art of difficulty in identifying subtle polysemous elements in video content, difficulty in assessing their contextualized misinterpretation risk, and difficulty in making effective decisions between content modification and audience screening. This application can significantly improve the accuracy and effectiveness of short video promotion, reduce the risk of wasting promotional resources due to user misinterpretation, and provide merchants with more targeted and actionable promotion optimization suggestions, thereby overcoming the shortcomings of existing technologies such as poor promotional effect and high decision-making complexity, and achieving unexpected technical results. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating a short video promotion and optimization method provided in this application. Detailed Implementation

[0008] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0009] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0010] Traditional short video promotion methods often struggle to handle subtle elements within video content that can lead to differing user interpretations. These ambiguous elements can cause users to misunderstand the core promotional message, thus impacting promotional effectiveness. Especially given the ever-changing viewing contexts and personal preferences of users, the same video clip may be interpreted differently by different users, making it difficult for the promotional system to optimize effectively. If these issues are not addressed, the initial feedback data collected by the promotional system will be diluted or even misled, resulting in poor promotional performance and wasted resources.

[0011] In view of the above problems, this application provides a method for optimizing short video promotion. By introducing multi-meaning element identification, contextualized misinterpretation risk assessment, and a cost-effective intervention strategy decision-making mechanism, it effectively solves the problem of user misunderstanding caused by the polysemy of content in existing short video promotion.

[0012] The following specific embodiments will provide a detailed introduction and explanation of the short video promotion and optimization method and system provided in this application.

[0013] Reference Figure 1 This application provides a method for optimizing short video promotion, which may include the following steps:

[0014] S1. Obtain the content data of the video to be promoted.

[0015] Content data refers to the various information carriers that make up a short video, which can include visual data, auditory data, and text data. Visual data usually refers to the sequence of image frames in the video, auditory data refers to the accompanying audio, and text data may include video titles, descriptions, subtitles, or embedded text.

[0016] For example, this application can use a video processing module to parse uploaded short video files, extract each frame of the video as visual data, extract the audio track of the video as auditory data, and identify the video title, description, subtitles, etc. as text data.

[0017] In practical applications, this application can acquire this data in various ways. For example, visual data can be processed using image recognition technology to extract keyframes or object information from the video; auditory data can be converted into text using speech recognition technology, and non-speech features such as background music and sound effects can be extracted; text data can be obtained directly from the video's metadata.

[0018] S2. Perform multi-meaning element identification on the content data to determine at least one potential multi-meaning element in the video to be promoted.

[0019] Potentially multi-meaning elements refer to visual or auditory segments in video content that may be interpreted differently by different user groups. These elements may cause misunderstandings about the video's core promotional intent in specific contexts. For example, a visual symbol with special meaning in a particular cultural context, or background music with multiple interpretations in different contexts.

[0020] For example, in a video promoting healthy foods, a quickly flashing visual symbol associated with unhealthy foods, or background music that easily evokes other products, could be identified as a potential multi-meaning element. There are various methods for identifying potential multi-meaning elements.

[0021] As an optional implementation, this application can identify the main visual object by performing object recognition on visual data, perform speech conversion and audio feature extraction on auditory data to determine speech text and non-speech audio features, and extract keywords and entities from text data to determine the core promotional concept. Subsequently, the first semantic correlation degree between the main visual object, non-speech audio features, and speech text and the core promotional concept is calculated, and the second semantic correlation degree between the main visual object, non-speech audio features, and speech text and at least one non-core concept is also calculated. Finally, based on the comparison between the first and second semantic correlation degrees, potential multi-meaning elements are identified. For example, if a visual object has a low correlation degree with the core promotional concept but a high correlation degree with a non-core concept, then that visual object may be a potential multi-meaning element.

[0022] As another possible implementation, this application performs high-resolution acoustic feature analysis on auditory data to identify specific acoustic cues below a conventional volume threshold. When a specific acoustic cue is identified, a semantic focus guidance signal is generated, which contains semantic direction information and temporal information related to the specific acoustic cue. Based on the semantic focus guidance signal, refined visual feature extraction is performed on the visual data corresponding to the temporal information to obtain enhanced visual features related to the semantic direction information. Finally, based on the specific acoustic cue and the enhanced visual features, potentially multi-meaning elements are identified. For example, a low background music might inadvertently draw the user's attention to a secondary element in the image, thus deviating from the core promotional message.

[0023] S3. Based on the potential multi-meaning elements and the preset user context model, determine the contextualized misinterpretation risk assessment results of the potential multi-meaning elements.

[0024] The user context model is a system used to simulate and predict how users will understand video content in specific viewing contexts. This model comprehensively considers factors such as users' historical preferences, viewing environment, and cultural background to assess the likelihood and direction of misinterpretations caused by potentially multi-meaningful elements. The contextualized misinterpretation risk assessment result is a quantitative indicator calculated based on potentially multi-meaningful elements and the user context model. It characterizes the likelihood of a comprehension bias caused by potentially multi-meaningful elements in a specific user viewing context and the specific direction of that bias. For example, the assessment result might indicate that a segment might be interpreted as humorous by younger users, but as disrespectful by older users.

[0025] For example, a visual element that is considered humorous in one cultural context may be considered offensive in another.

[0026] First, this application can determine the timing and duration of potential multi-meaning elements appearing in the video to be promoted. Second, this application can obtain historical preference data from simulated users, which is used to characterize the simulated users' interest in specific content types. Finally, this application inputs the timing of appearance, duration, historical preference data, and pre-stored misinterpretation association rules into a user context model to calculate the contextualized misinterpretation risk assessment result. The misinterpretation association rules are used to define the mapping relationship between content features, user context, and comprehension bias types. For example, if a potential multi-meaning element appears for a long time in the video, and the simulated user's historical preference data shows that they may have a negative interpretation of the element, then its misinterpretation risk assessment result will be correspondingly higher.

[0027] S4. Based on the contextualized misinterpretation risk assessment results, pre-existing promotion constraints, and content modification cost model, make a decision between content modification strategy and audience screening strategy to generate an initial intervention strategy.

[0028] Promotional constraints refer to various restrictions that must be followed during short video promotion, such as promotion budget, promotion cycle, and brand image requirements. The content modification cost model is a model used to estimate the cost and time required to modify video content, which can include labor costs, technical costs, and time costs. Content modification strategies refer to intervention measures that eliminate or reduce the risk of misinterpretation by adjusting potentially multi-meaning elements in the video content. Content modification strategies include adjusting potentially multi-meaning elements in the video to be promoted. Audience screening strategies include selecting a test user group. Audience screening strategies refer to intervention measures that avoid the risk of misinterpretation by adjusting the target user group without modifying the video content.

[0029] For example, if a misinterpretation risk assessment indicates that a particular segment has an extremely high risk of misinterpretation, and the cost of modification is low, then a content modification strategy might be prioritized. The decision-making process may include: simulating the content modification strategy using a content modification cost model to estimate the cost and time required to implement the strategy; estimating the first expected promotional benefit after implementing the content modification strategy, and the second expected promotional benefit after implementing the audience screening strategy; and, based on cost, time, the first expected promotional benefit, and the second expected promotional benefit, combined with promotional constraints, selecting the strategy with the highest expected net benefit as the initial intervention strategy. For example, if modifying video content requires a significant investment of time and money, but the expected increase in benefits is limited, while precise audience screening can yield a higher net benefit, then the audience screening strategy would be chosen.

[0030] S5. Implement initial intervention strategies and optimize the promotion process based on user feedback data obtained after implementation.

[0031] The initial intervention strategy is the preferred intervention plan determined based on the contextualized misinterpretation risk assessment results, promotion constraints, and content modification cost model. This can be a content modification strategy or an audience screening strategy. User feedback data refers to various behavioral and evaluation data generated by users after watching videos, such as completion rate, likes, comments, and shares.

[0032] Specifically, if the initial intervention strategy is a content modification strategy, this application generates targeted video modification suggestions based on the type of potentially multi-meaning elements and the direction of deviation in the contextualized misinterpretation risk assessment results, and provides these suggestions to the user. For example, it suggests replacing a potentially misleading scene with a clearer expression, or adjusting the background music to avoid ambiguity. If the initial intervention strategy is an audience screening strategy, users with a lower probability of misunderstanding potentially multi-meaning elements than a preset threshold are selected from the initial target user group based on the direction of deviation in the contextualized misinterpretation risk assessment results. These users form a test user group, and the videos to be promoted are prioritized for this test user group. For example, if a segment might be misunderstood by users of a specific age group, users of that age group are excluded from the test user group. During the promotion process, user feedback data is continuously collected, including completion rate, interaction behavior data, and semantics of comments. The accuracy of the contextualized misinterpretation risk assessment results is verified based on the user feedback data. Based on the verification results, the target user group or recommendation algorithm parameters for subsequent promotions are dynamically adjusted. For example, if a segment assessed as low-risk is found to actually generate a large number of negative comments, the user context model or misinterpretation association rules will be adjusted, and a more cautious strategy will be adopted for similar content in subsequent promotions.

[0033] The short video promotion optimization method proposed in this application effectively solves the problem of user misunderstanding caused by the polysemy of content in existing short video promotions by introducing multi-meaning element identification, contextualized misinterpretation risk assessment, and a cost-effective intervention strategy decision-making mechanism. Traditional methods often focus on the identification of explicit and general features, making it difficult to capture the subtle "implicit information" or "polysemous elements" in video content that may lead to different interpretations among different user groups. This application can accurately locate these potential sources of misinterpretation by identifying multi-meaning elements in content data. In addition, this application further combines user context models to conduct contextualized misinterpretation risk assessment of potential multi-meaning elements, enabling the system to understand the specific impact of these elements in different user viewing contexts, thereby avoiding the problem of diluted or misguided initial feedback data caused by "misinterpretation" in traditional methods.

[0034] More importantly, after identifying high-risk misinterpretation segments, this application goes beyond simply adjusting recommendation algorithm parameters. It delves into intervention suggestions regarding the "content itself" and provides a decision-making mechanism between content modification strategies and audience selection strategies. By comprehensively considering content modification cost models, expected promotional benefits, and promotional constraints, this application can intelligently make the optimal choice between modifying video content and adjusting the target user group, thereby maximizing promotional effectiveness. The complexity of this decision-making lies in its balance between the cost and effectiveness of content modification, and the efficiency and coverage of precise targeting, enabling the promotion system to respond more flexibly and efficiently to various promotional challenges. Compared to existing technologies, this application not only identifies potential misinterpretation risks but also provides practical intervention strategies and continuously optimizes them based on user feedback, significantly improving the accuracy and efficiency of short video promotion and bringing higher promotional revenue to merchants.

[0035] In some embodiments, the above-mentioned multi-meaning element identification of content data to determine at least one potential multi-meaning element in the video to be promoted includes: object identification of visual data to determine the main visual object; speech conversion and audio feature extraction of auditory data to determine speech text and non-speech audio features; keyword and entity extraction of text data to determine the core promotion concept; calculating the first semantic correlation between the main visual object, non-speech audio features, and speech text and the core promotion concept; calculating the second semantic correlation between the main visual object, non-speech audio features, and speech text and at least one non-core concept; and identifying potential multi-meaning elements based on the comparison between the first and second semantic correlations.

[0036] Object recognition in visual data refers to using computer vision techniques, such as deep learning models (YOLO, Faster R-CNN, etc.), to analyze the image content in video frames, automatically detect and identify specific objects, people, scenes, etc., appearing in the video, and using them as the main visual objects. Its purpose is to extract the core content elements of the video from a visual perspective.

[0037] Speech-to-speech conversion and audio feature extraction of auditory data refers to converting the audio content in a video into analyzable speech text using Automatic Speech Recognition (ASR) technology, while simultaneously extracting acoustic features from non-speech components (such as background music and sound effects), including pitch, volume, timbre, and rhythm. The aim is to comprehensively capture the auditory information of the video, including verbal expression and non-verbal emotions or atmosphere.

[0038] Keyword and entity extraction from text data refers to using Natural Language Processing (NLP) technology to identify keywords and named entities (such as brand names, product names, and personal names) that best represent the video's theme and promotional intent from text information such as video titles, descriptions, and subtitles, and defining them as core promotional concepts. Its purpose is to clarify the video's core promotional intent.

[0039] Calculating the first semantic relevance of the main visual objects, non-speech audio features, and speech text to the core promotional concept involves constructing a multimodal semantic model, such as a multimodal fusion model based on the Transformer architecture, to evaluate the semantic relevance or similarity between the identified main visual objects, non-speech audio features, and speech text and the core promotional concept. The higher the relevance, the more the element aligns with the core promotional intent.

[0040] Calculating the second semantic relevance of the main visual object, non-speech audio features, and speech text to at least one non-core concept refers to using a multimodal semantic model to evaluate the semantic relevance between the above elements and pre-defined non-core concepts that may cause ambiguity or deviate from the core promotional intent. These non-core concepts can be words similar to the core concept but with subtle differences in meaning, or words that are easily misunderstood in a specific context.

[0041] Based on the comparison between primary and secondary semantic relevance, potential multi-meaning elements are identified when a visual object, auditory feature, or spoken text has a low primary semantic relevance to the core promotional concept but a high secondary semantic relevance to a non-core concept, or when both relevances are high but the secondary semantic relevance is significantly higher than the primary semantic relevance. This indicates that the element might be misinterpreted by users as a non-core concept in the video, thus deviating from the promotional intent.

[0042] This application's solution addresses the accuracy limitations of traditional methods in identifying potentially multi-meaning elements by performing multi-dimensional and refined semantic analysis of short video content. Specifically, firstly, a multimodal representation of the video content is constructed through comprehensive feature extraction from visual, auditory, and textual data. Secondly, by calculating the semantic relevance of these multimodal features to both core and non-core promotional concepts, the system can quantitatively assess the alignment of each element in the video with the promotional intent and the risk of misinterpretation. When an element has a low relevance to the core promotional concept but a high relevance to a non-core concept, the system can accurately identify it as a potentially multi-meaning element. This comparative analysis mechanism allows the system to effectively distinguish segments that appear related to the core concept but may actually be misunderstood by users in specific contexts, thus providing a solid foundation for subsequent misinterpretation risk assessment and intervention strategy development.

[0043] In some embodiments, this application further proposes the following steps for identifying multi-meaning elements in content data to determine at least one potential multi-meaning element in a video to be promoted: performing high-resolution acoustic feature analysis on auditory data to identify specific acoustic cues below a conventional volume threshold; when a specific acoustic cue is identified, generating a semantic focus guidance signal, the semantic focus guidance signal containing semantic direction information and time information related to the specific acoustic cue; performing refined visual feature extraction on the visual data corresponding to the time information based on the semantic focus guidance signal to obtain enhanced visual features related to the semantic direction information; and identifying potential multi-meaning elements based on the specific acoustic cue and the enhanced visual features.

[0044] Specifically, high-resolution acoustic feature analysis of auditory data refers to using advanced audio processing techniques, such as wavelet transform, Mel-frequency cepstral coefficient (MFCC) analysis, or deep learning acoustic models, to perform detailed spectral and temporal analysis of the auditory data from short videos. The aim is to capture acoustic features that may not be apparent at normal volumes but carry specific semantic or emotional information. These features may include faint background noise, whispers, subtle changes in specific sound effects, etc.

[0045] Among these, specific acoustic cues below the conventional volume threshold can be understood as acoustic segments identified through high-resolution acoustic feature analysis whose loudness or energy is below the level easily perceived by the human ear or conventional audio processing systems, but whose presence is crucial for understanding video content or evoking specific associations. For example, a faint instrumental sound with specific cultural connotations suddenly appearing in the background music of a video, or a sigh that is not easily detected in a dialogue.

[0046] When a specific acoustic cue is identified, a semantic focus guidance signal is generated. This signal contains semantic orientation information and temporal information related to the specific acoustic cue. Semantic orientation information refers to the potential meaning or associated concept that the acoustic cue may imply; for example, a faint alarm sound may indicate "danger" or "warning." The temporal information precisely records the time and duration of the acoustic cue's appearance in the video. The purpose of this signal is to expand the focus of analysis from purely auditory data to visual content closely related to the auditory cue.

[0047] Furthermore, based on the semantic focus guidance signal, refined visual feature extraction is performed on the visual data corresponding to the time information to obtain enhanced visual features related to the semantic directional information. This means that the system can accurately locate visual segments in the video that occur simultaneously with specific acoustic cues, based on the time information provided in the semantic focus guidance signal. Subsequently, more in-depth and detailed visual feature extraction is performed on these visual segments, such as recognizing micro-expressions, gestures, detailed textures or color changes of specific objects in the image. These visual features are closely related to the semantic directional information and can jointly constitute a potential multi-meaning element.

[0048] Therefore, potential multi-meaning elements can be identified based on specific acoustic cues and enhanced visual features. By combining specific acoustic cues that are not easily detected in auditory data with refined visual features associated with them in visual data, potential multi-meaning elements that may cause misunderstandings by users can be identified more comprehensively and accurately.

[0049] This application's solution, through high-resolution acoustic feature analysis of auditory data, overcomes the limitations of traditional methods on volume thresholds, capturing specific acoustic cues that may be overlooked but possess significant semantic value. Because these cues are often imperceptible, they are more likely to trigger unconscious comprehension biases when viewed by the user. By generating semantic focus guidance signals, the semantic direction and temporal information of these acoustic cues are precisely transmitted to the visual analysis module, enabling targeted and refined feature extraction from the visual data. This deep fusion analysis of auditory and visual information allows the system to identify potential multi-modal elements composed of multiple modal information that are difficult to discover through single-modal analysis. For example, a faint background sound may, together with a subtle detail of an object in the image, constitute a metaphor with dual meaning; this solution reveals these metaphors through this collaborative analysis mechanism.

[0050] In some embodiments, the step of determining the contextualized misreading risk assessment result of potential multi-meaning elements based on potential multi-meaning elements and a preset user context model in the above-mentioned short video promotion optimization method can be further refined into the following operations: determining the appearance time and duration of potential multi-meaning elements in the video to be promoted; obtaining historical preference data of simulated users, which is used to characterize the simulated users' interest in specific content types; inputting the appearance time, duration, historical preference data, and pre-stored misreading association rules into the user context model to calculate the contextualized misreading risk assessment result, whereby the misreading association rules are used to define the mapping relationship between content features, user context, and comprehension deviation types.

[0051] Determining the timing and duration of potential multi-meaning elements in the video to be promoted involves performing time-series analysis on the video content to accurately identify the start and end times of these elements within the video stream and calculate their duration. For example, video content analysis techniques can be used to identify and annotate visual or auditory segments, thereby obtaining their precise location information on the timeline.

[0052] Obtaining historical preference data from simulated users refers to collecting and analyzing their past viewing behaviors, interaction records (such as likes, comments, and shares), subscribed content types, and demographic information. This data is used to construct user interest profiles to characterize the simulated user's interest tendencies towards specific content types or topics. For example, a user might show a high interest in technology videos but a lower interest in entertainment videos.

[0053] Pre-stored misinterpretation association rules can be understood as a set of predefined rules or rules trained through machine learning, aiming to establish a mapping relationship between content features, user context, and types of comprehension bias. Specifically, these rules can be defined as "if content contains specific visual elements and the user has a certain historical preference, it may trigger a certain type of comprehension bias." For example, a rule might state that when a "specific gesture" appears in a video and the user's historical preferences show sensitivity to content from a "specific cultural background," that gesture may be misinterpreted as having a negative meaning.

[0054] This application's approach first determines the specific time points and durations of potential multi-meaning elements appearing in the videos to be promoted, providing a precise temporal context for subsequent contextual analysis. Simultaneously, by acquiring historical preference data from simulated users, it gains a deeper understanding of users' interests in different content types, thus providing a personalized user perspective for assessing the likelihood of potential multi-meaning elements causing comprehension bias in specific user contexts. Furthermore, this temporal information, user preference data, and pre-stored misreading association rules are input into a user context model, enabling the model to comprehensively consider the characteristics of the content itself, individual user viewing habits, and known misreading patterns. Therefore, the user context model can calculate a more accurate contextualized misreading risk assessment result, which not only quantifies the probability of comprehension bias but also indicates the specific direction of the bias, thus providing a solid data foundation for the development of subsequent intervention strategies.

[0055] In some embodiments, the above-mentioned decision-making between content modification strategy and audience screening strategy based on contextualized misreading risk assessment results, pre-stored promotion constraints, and content modification cost model to generate an initial intervention strategy specifically includes: simulating the content modification strategy using the content modification cost model to estimate the cost and time required to implement the content modification strategy; estimating the first expected promotional benefit after implementing the content modification strategy and the second expected promotional benefit after implementing the audience screening strategy; and selecting the strategy with the larger expected net benefit as the initial intervention strategy based on cost, time, the first expected promotional benefit, and the second expected promotional benefit, combined with the promotion constraints.

[0056] Specifically, when making decisions, the first step is to simulate the content modification strategy using a content modification cost model. This model can be a pre-trained machine learning model whose inputs include the length and complexity of the video to be promoted, the number and types of potentially multi-meaningful elements, and the required degree of modification. The output is the cost and time required to execute the content modification strategy. Costs can include human resources, software tool costs, and time costs, while time refers to the work cycle required to complete the modification. In this way, the actual investment in the content modification strategy can be quantitatively estimated.

[0057] Furthermore, it is necessary to estimate the potential benefits of the two intervention strategies. The first expected promotional benefit after implementing the content modification strategy refers to the potential increase in user completion rate, interaction, and conversion rate, resulting from the video content becoming clearer and more aligned with the promotional intent after the adjustment of potentially multi-meaning elements. The second expected promotional benefit after implementing the audience screening strategy refers to the potential increase in completion rate, more active interaction, and higher conversion rate, resulting from pushing the video to a specific user group with a lower likelihood of misunderstanding, due to the higher match between the target users and the content. These expected benefits can be estimated based on historical promotional data, user behavior models, and an assessment of the degree of reduction in the risk of misinterpretation.

[0058] Ultimately, a decision is made based on estimated costs, time, the first expected promotional return, and the second expected promotional return, combined with pre-existing promotional constraints. These constraints may include budget limitations, time window requirements, and brand image maintenance. The decision-making process involves calculating the expected net return for each strategy, i.e., expected promotional return minus execution costs. Assuming all promotional constraints are met, the strategy with the greater expected net return is selected as the initial intervention strategy. For example, if the expected net return of the content modification strategy is A, and the expected net return of the audience screening strategy is B, and A > B, then the content modification strategy is chosen; conversely, if B > A, then the audience screening strategy is chosen.

[0059] This application's solution elevates the decision-making process from a simple assessment of technical feasibility to a consideration of maximizing economic benefits by introducing cost and time estimations for content modification strategies and projected promotional returns for two intervention strategies. It is precisely because of the quantitative analysis of various key indicators that the system can comprehensively evaluate the input-output ratio of different intervention strategies. By calculating the expected net return and comparing it with promotional constraints, it ensures that the selected initial intervention strategy not only effectively reduces the risk of misinterpretation but also brings the greatest economic return to short video promotion within limited resources and time. This decision-making mechanism based on quantitative evaluation avoids the problems of resource waste and poor promotional effects that may result from blind selection or empirical decision-making.

[0060] In some embodiments, when the initial intervention strategy generated by the decision is a content modification strategy, the initial intervention strategy is executed, including: generating targeted video modification suggestions based on the type of potential multi-meaning elements and the direction of deviation in the contextualized misinterpretation risk assessment results; and providing the video modification suggestions to the user.

[0061] Specifically, the types of potentially multi-meaning elements can be understood as the specific content forms that lead to comprehension biases. For example, these could be visually ambiguous images, auditory ambiguous speech, or textually ambiguous words or phrases. The direction of bias in the contextualized misinterpretation risk assessment results refers to the specific tendency of potentially multi-meaning elements to cause comprehension biases in a particular user's viewing context. For example, it might cause users to misunderstand the promotional content as having negative connotations, being sarcastic, or being on topics completely unrelated to the core promotional intent.

[0062] The generation of targeted video modification suggestions refers to the system automatically or semi-automatically proposing specific and actionable content adjustment plans based on the specific type of identified potentially multi-meaning elements and the direction of any potential misinterpretation. For example, if the potentially multi-meaning element is a visually blurred image, and the misinterpretation is that it might be misunderstood as indecent content, the modification suggestions might include replacing the image, adding text descriptions, or partially blurring the image. If the potentially multi-meaning element is auditory ambiguity, and the misinterpretation is that it might be misunderstood as other product information, the modification suggestions might include re-recording the audio, adding subtitles, or adjusting background music to highlight the core information. These suggestions aim to directly eliminate or reduce the risk of misinterpretation caused by potentially multi-meaning elements.

[0063] In practical applications, providing video modification suggestions to users means transmitting the generated modification suggestions to content creators or operators through the short video platform's content management interface, API interface, or customized reports. Users can then edit and adjust the video content accordingly based on these suggestions to ensure that the video accurately conveys the promotional intent.

[0064] This application's solution generates highly targeted video modification suggestions by deeply analyzing the specific types of potentially multi-meaning elements and the possible deviations they may cause in specific contexts. Because these suggestions are tailored based on detailed risk assessments, content creators can accurately understand the problems and take the most effective measures to adjust the content. This mechanism avoids blind or indiscriminate modifications, ensuring that every content adjustment directly addresses the risk of misinterpretation, thereby significantly improving the efficiency and accuracy of content modification strategies.

[0065] In some embodiments, when the initial intervention strategy generated by the decision is an audience screening strategy, the initial intervention strategy is executed, including: selecting users from the initial target user group whose misunderstanding of potential multi-meaning elements is less than a preset threshold based on the deviation direction in the contextualized misreading risk assessment results, to form a test user group; and prioritizing the push of the video to be promoted to the test user group.

[0066] Specifically, when the system decides to generate an initial intervention strategy of audience screening based on the contextualized misinterpretation risk assessment results, pre-stored promotion constraints, and content modification cost model, it needs to conduct refined screening of target users. The deviation direction in the contextualized misinterpretation risk assessment results refers to the specific type or tendency of comprehension deviation that potential multi-meaning elements may cause in a particular user's viewing context, such as interpreting something positive as negative, or misunderstanding a specific concept. The initial target user group refers to all potential viewers who might be interested in the short video, as pre-defined by the system before any screening. Users whose probability of comprehension deviation of potential multi-meaning elements is below a preset threshold are those whose probability of misunderstanding the potential multi-meaning element is predicted to be lower than a pre-set standard by analyzing their historical behavior, preferences, demographic characteristics, and the deviation direction in the contextualized misinterpretation risk assessment results. Users screened in this way constitute a test user group, which is considered to have a more accurate understanding of the video's core promotional intent. Subsequently, the video to be promoted will be prioritized for this test user group, meaning these users will receive the video recommendation earlier or more frequently than other users.

[0067] This application's solution leverages the bias direction information from contextualized misinterpretation risk assessment results to refine the initial target user group during audience screening. Specifically, instead of blindly selecting a subset of users for testing, the system proactively identifies and excludes users highly sensitive or susceptible to specific bias directions that may arise from potentially multi-meaning elements. This ensures that the selected test user group has a low probability of misunderstanding potentially multi-meaning elements, effectively reducing the risk of negative feedback or poor promotional results due to misunderstandings in the early stages of video promotion. By prioritizing the delivery of the video to this optimized test user group, user feedback aligned with the core promotional intent can be collected more accurately, providing a more reliable data foundation for subsequent promotional optimization.

[0068] In some embodiments, the above-mentioned optimization of the promotion process based on user feedback data obtained after execution includes: continuously collecting user feedback data during the promotion process, including viewing completion rate, interaction behavior data, and semantics of comment content; verifying the accuracy of the contextualized misinterpretation risk assessment results based on the user feedback data; and dynamically adjusting the target user group or recommendation algorithm parameters for subsequent promotions based on the verification results.

[0069] Specifically, during the promotion of short videos, the system is configured to continuously collect user feedback data. User feedback data is a key indicator for measuring users' understanding and acceptance of video content, and it can include, but is not limited to, completion rate, interaction behavior data, and semantic analysis of comments. Completion rate refers to the user's progress in watching the video; for example, whether the user watched the entire video or at what point they stopped watching. Interaction behavior data can be understood as the various interactive actions users take while watching the video, such as liking, sharing, saving, forwarding, and clicking links. These actions directly reflect users' interest in and engagement with the video content. Semantic analysis of comments involves performing natural language processing and sentiment analysis on the text content left by users in the comment section to extract information such as users' specific opinions, questions, emotional tendencies, and whether there are any misunderstandings regarding the video content. By continuously collecting this multi-dimensional, real-time feedback data, a comprehensive understanding of users' actual reactions to the promoted videos can be achieved.

[0070] Furthermore, the accuracy of the previously determined contextualized misinterpretation risk assessment results can be verified based on the collected user feedback data. For example, if the contextualized misinterpretation risk assessment results predict that a potentially multi-meaning element may cause comprehension bias in a specific context, and actual user feedback data (such as numerous questions in comments or a significant drop in completion rates for specific segments) confirms this bias, then the assessment results are accurate. Conversely, if the assessment results predict a risk, but actual feedback shows that users understand it well, or if the assessment results do not predict a risk but comprehension bias actually occurs, then the assessment results have a certain degree of bias. In this way, the effectiveness of the contextualized misinterpretation risk assessment model can be calibrated in real time.

[0071] Based on this, the target user group or recommendation algorithm parameters for subsequent promotions can be dynamically adjusted according to the verification results. For example, if the verification results show that a potentially multi-meaning element is indeed prone to misinterpretation among a specific user group, the target user group for subsequent promotions can be adjusted. The video can be prioritized for users whose likelihood of misunderstanding the element is below a preset threshold, or promotion to this easily misinterpreted user group can be temporarily avoided. Simultaneously, recommendation algorithm parameters can be adjusted, for example, reducing the recommendation weight of the video under a specific user profile, or increasing the recommendation of explanatory or supplementary videos related to the video content to guide users to form a correct understanding. This dynamic adjustment mechanism allows the promotion strategy to be flexibly optimized based on actual results, thereby improving overall promotion efficiency and user experience.

[0072] This application's solution effectively addresses the problems of insufficient feedback utilization and lack of targeted optimization that may exist in traditional promotion optimization processes by introducing a closed-loop optimization mechanism based on user feedback. Specifically, by continuously and comprehensively collecting user feedback data such as completion rate, interaction behavior data, and semantic data of comments during the promotion process, the system can obtain rich information about users' actual understanding and acceptance of the promotional video. It is precisely because of this multi-dimensional feedback data that the system can objectively and in real-time verify the accuracy of the previous contextualized misinterpretation risk assessment results. When the verification results show that the assessment is biased, the system can promptly identify the shortcomings of the assessment model or the initial intervention strategy. On this basis, by dynamically adjusting the target user group or recommendation algorithm parameters for subsequent promotions, this application's solution can directly transform user feedback into actionable optimization measures, thereby avoiding the waste of promotional resources and ensuring that the promotional strategy can continuously adapt to user needs and market changes.

[0073] In some embodiments, the method further includes: acquiring social opinion data related to potential multi-meaning elements from external information sources in real time; analyzing the sentiment tendency of the social opinion data and the changes in its correlation strength with preset concepts; dynamically adjusting the semantic representation of potential multi-meaning elements based on the changes in sentiment tendency and correlation strength to update the semantic association basis in the user context model; the determination of the contextualized misreading risk assessment result of potential multi-meaning elements based on the potential multi-meaning elements and the preset user context model includes: determining the appearance time and duration of potential multi-meaning elements in the video to be promoted; acquiring historical preference data of simulated users, which is used to characterize the simulated users' interest in specific content types; calculating a first misreading risk assessment result based on the appearance time, duration, historical preference data, and pre-stored misreading association rules; calculating a second misreading risk assessment result based on the sentiment tendency and correlation strength changes of social opinion data related to potential multi-meaning elements, which is used to characterize new comprehension bias risks caused by the external social environment; and weightedly combining the first misreading risk assessment result and the second misreading risk assessment result to obtain a contextualized misreading risk assessment result.

[0074] Specifically, acquiring real-time social opinion data related to potentially multi-meaning elements from external information sources refers to the system using web crawlers, API interfaces, and other technical means to collect text or multimedia data such as user discussions, comments, and news reports related to potentially multi-meaning elements in the video to be promoted from public channels such as social media platforms (e.g., Weibo, WeChat, Douyin comment sections), news websites, forums, and blogs. This data can reflect the public's real-time views and emotions regarding specific concepts, events, or expressions.

[0075] Analyzing the sentiment trends in public opinion data and the changes in the strength of their association with pre-set concepts can be understood as using Natural Language Processing (NLP) technology to perform sentiment analysis on the acquired public opinion data, determining whether it is positive, negative, or neutral. Simultaneously, through techniques such as word vector models and topic models, the semantic association strength between potential multi-meaning elements and various pre-set concepts (including core and non-core concepts) in the current public opinion environment is calculated, and the dynamic changes in these association strengths are monitored. For example, a word may generate new negative associations due to social events at a specific time, and its association strength with negative concepts will significantly increase.

[0076] In practical applications, dynamically adjusting the semantic representation of potential multi-meaning elements based on changes in sentiment and association strength to update the semantic association basis in the user context model means that when a significant change in public sentiment towards a potential multi-meaning element is detected, or when its association strength with a specific concept shifts significantly, the system will correspondingly modify the semantic vector or association weight of that potential multi-meaning element in the user context model. For example, if the association between the word "apple" and the concept of "monopoly" rises sharply due to negative news within a certain period, the semantic representation of "apple" in the user context model will be adjusted to give more consideration to the negative association of "monopoly" when assessing the risk of misinterpretation.

[0077] Furthermore, in determining the contextualized misinterpretation risk assessment results for potential multi-meaning elements, a first misinterpretation risk assessment result is first calculated based on the time of occurrence, duration, historical preference data, and pre-existing misinterpretation association rules. This first misinterpretation risk assessment result is mainly calculated based on existing, relatively stable user context models and rules within the system, reflecting the possibility of comprehension bias caused by potential multi-meaning elements in normal contexts. Based on this, a second misinterpretation risk assessment result is calculated according to the sentiment tendency and association strength changes of social opinion data related to potential multi-meaning elements. The second misinterpretation risk assessment result is specifically used to characterize new or dynamically changing comprehension bias risks caused by external social environments (such as sudden events, hot topics, etc.). Finally, the first and second misinterpretation risk assessment results are weighted and synthesized to obtain the final contextualized misinterpretation risk assessment result. This weighted synthesis can assign different weights to the first and second misinterpretation risk assessment results according to actual needs and the importance of different risk sources, to obtain a more comprehensive and accurate integrated risk assessment.

[0078] This application's solution effectively overcomes the limitations of traditional static user context models for misinterpretation risk assessment by introducing real-time social opinion data. Specifically, once a potentially multi-meaning element is identified, the system not only considers internal factors such as its appearance time in the video, duration, and simulated user historical preferences, but also, by acquiring social opinion data related to that element from external information sources in real time, captures the element's latest semantic changes and sentiment tendencies in the current social context. For example, a word or image may be given new meanings or evoke specific emotions due to social events at a particular time, changes that are difficult for static models to predict. By analyzing the changes in sentiment tendencies and correlation strength in social opinion data, the system can dynamically adjust the semantic representation of potentially multi-meaning elements, thereby updating the semantic association basis in the user context model and enabling the model to adapt to the constantly changing external environment. Therefore, when calculating the contextualized misinterpretation risk assessment result, the system can weightedly synthesize the first misinterpretation risk assessment result based on the internal model and the second misinterpretation risk assessment result based on external social opinion. This dual assessment mechanism ensures the comprehensiveness of risk assessment, taking into account both individual user preferences and inherent content characteristics, as well as the real-time impact of the broader social environment on content comprehension. This enables more accurate prediction of potential comprehension biases, especially when facing new types of misinterpretation risks caused by external environments, where this solution can provide more timely and accurate early warnings.

[0079] This application also discloses a short video promotion optimization system, comprising: an acquisition unit for acquiring content data of a video to be promoted, the content data including visual data, auditory data, and text data; a first determination unit for identifying multi-meaning elements in the content data to determine at least one potential multi-meaning element in the video to be promoted, the potential multi-meaning element being used to characterize visual or auditory segments that may cause users to deviate from the core promotion intention; a second determination unit for determining the contextualized misinterpretation risk assessment result of the potential multi-meaning element based on the potential multi-meaning element and a preset user context model, the contextualized misinterpretation risk assessment result being used to characterize the possibility and direction of the potential multi-meaning element causing comprehension deviation in a specific user viewing context; a generation unit for making a decision between a content modification strategy and an audience selection strategy based on the contextualized misinterpretation risk assessment result, pre-stored promotion constraints, and a content modification cost model, to generate an initial intervention strategy, the content modification strategy including adjustments to the potential multi-meaning elements in the video to be promoted, and the audience selection strategy including the selection of a test user group; and an execution unit for executing the initial intervention strategy and optimizing the promotion process based on user feedback data obtained after execution.

[0080] This system, through its modular design, achieves intelligent optimization of the entire short video promotion process. The acquisition unit is responsible for comprehensively collecting video content information; the first and second determination units work together to accurately identify potential misinterpretation risks in the video and conduct contextual assessments; the generation unit, based on the assessment results and multiple constraints, intelligently decides on the optimal initial intervention strategy; finally, the execution unit is responsible for implementing this strategy and continuously optimizing the promotion process. This systematic approach effectively solves the problem of user misunderstanding caused by the ambiguity of content in traditional promotion, significantly improving the accuracy and efficiency of promotion.

[0081] The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for short video promotion optimization, characterized in that, include: Obtain the content data of the video to be promoted, including visual data, auditory data, and text data; The content data is subjected to multi-meaning element identification to determine at least one potential multi-meaning element in the video to be promoted. The potential multi-meaning element is used to characterize visual or auditory segments that may cause users to deviate from the core promotional intent. Based on the potential multi-meaning elements and the preset user context model, the contextualized misreading risk assessment result of the potential multi-meaning elements is determined. The contextualized misreading risk assessment result is used to characterize the possibility and direction of the comprehension deviation caused by the potential multi-meaning elements in a specific user viewing context. Based on the contextualized misinterpretation risk assessment results, pre-stored promotion constraints, and content modification cost model, a decision is made between content modification strategy and audience screening strategy to generate an initial intervention strategy. The content modification strategy includes adjusting the potential multi-meaning elements in the video to be promoted, and the audience screening strategy includes selecting a test user group. Implement the initial intervention strategy and optimize the promotion process based on user feedback data obtained after implementation.

2. The method of claim 1, wherein, The step of identifying multiple meaning elements in the content data to determine at least one potential multiple meaning element in the video to be promoted includes: The visual data is used for object recognition to identify the main visual object; The auditory data is subjected to speech conversion and audio feature extraction to determine the speech text and non-speech audio features; Keyword and entity extraction is performed on the text data to determine the core promotional concept; Calculate the first semantic correlation between the main visual object, the non-speech audio features, the speech text, and the core promotion concept; Calculate the second semantic association degree between the main visual object, the non-speech audio features, and the speech text and at least one non-core concept; Based on the comparison between the first semantic relevance and the second semantic relevance, the potential multi-meaning elements are identified.

3. The method of claim 1, wherein, The step of identifying multiple meaning elements in the content data to determine at least one potential multiple meaning element in the video to be promoted includes: High-resolution acoustic feature analysis is performed on the auditory data to identify specific acoustic cues below the conventional volume threshold; When the specific acoustic cue is identified, a semantic focus guidance signal is generated, which contains semantic pointing information and time information related to the specific acoustic cue; Based on the semantic focus guidance signal, the visual data corresponding to the time information is subjected to refined visual feature extraction to obtain enhanced visual features related to the semantic orientation information; The potential multi-meaning elements are identified based on the specific acoustic cues and the enhanced visual features.

4. The method of claim 1, wherein, The step of determining the contextualized misinterpretation risk assessment result of the potential multi-meaning elements based on the potential multi-meaning elements and the preset user context model includes: Determine the timing and duration of the appearance of the potential multi-meaning elements in the video to be promoted; Acquire historical preference data of simulated users, which is used to characterize the simulated users' interest in specific content types; The occurrence time, duration, historical preference data, and pre-stored misreading association rules are input into the user context model to calculate the contextualized misreading risk assessment result. The misreading association rules are used to define the mapping relationship between content features, user context, and comprehension bias types.

5. The method of claim 1, wherein, The step involves making a decision between content modification strategies and audience screening strategies based on the contextualized misinterpretation risk assessment results, pre-stored promotion constraints, and content modification cost models, in order to generate an initial intervention strategy, including: The content modification strategy is simulated using a content modification cost model to estimate the cost and time required to implement the content modification strategy; Estimate the first expected promotional revenue after implementing the content modification strategy, and the second expected promotional revenue after implementing the audience screening strategy; Based on the cost, the time, the first expected promotion revenue, and the second expected promotion revenue, and in conjunction with the promotion constraints, the strategy with the largest expected net revenue is selected as the initial intervention strategy.

6. The method of claim 1, wherein, When the initial intervention strategy generated by the decision is the content modification strategy, executing the initial intervention strategy includes: Based on the type of the potential multi-meaning elements and the direction of deviation in the contextualized misreading risk assessment results, targeted video modification suggestions are generated; The video modification suggestions will be provided to the user.

7. The method according to claim 1, characterized in that, When the initial intervention strategy generated by the decision is the audience screening strategy, executing the initial intervention strategy includes: Based on the deviation direction in the contextualized misreading risk assessment results, users whose misunderstanding of the potential multi-meaning elements is less than a preset threshold are selected from the initial target user group to form a test user group; The video to be promoted will be prioritized and pushed to the test user group.

8. The method of claim 1, wherein, The optimization of the promotion process based on user feedback data obtained after execution includes: During the promotion process, the user feedback data is continuously collected, including viewing completion rate, interaction behavior data, and semantic information of comment content; Based on the user feedback data, verify the accuracy of the contextualized misreading risk assessment results; Based on the verification results, the target user group or recommendation algorithm parameters for subsequent promotions will be dynamically adjusted.

9. The method of claim 1, wherein, The method further includes: Real-time acquisition of public opinion data related to the potential multi-meaning elements from external information sources; Analyze the sentiment trends in the aforementioned public opinion data and the changes in the strength of their association with pre-defined concepts; Based on the changes in the emotional tendency and the strength of the association, the semantic representation of the potential multi-meaning elements is dynamically adjusted to update the semantic association basis in the user context model; The step of determining the contextualized misinterpretation risk assessment result of the potential multi-meaning elements based on the potential multi-meaning elements and the preset user context model includes: Determine the timing and duration of the appearance of the potential multi-meaning elements in the video to be promoted; Acquire historical preference data of simulated users, which is used to characterize the simulated users' interest in specific content types; The first misread risk assessment result is calculated based on the occurrence time, duration, historical preference data, and pre-stored misread association rules. Based on the sentiment tendency and the change in the correlation strength of social opinion data related to potential multi-meaning elements, a second misreading risk assessment result is calculated. The second misreading risk assessment result is used to characterize the risk of new comprehension bias caused by the external social environment. The first misreading risk assessment result and the second misreading risk assessment result are weighted and synthesized to obtain the contextualized misreading risk assessment result.

10. A short video promotion optimization system, characterized in that, include: The acquisition unit is used to acquire content data of the video to be promoted, the content data including visual data, auditory data and text data; The first determining unit is used to identify multi-meaning elements in the content data to determine at least one potential multi-meaning element in the video to be promoted. The potential multi-meaning element is used to characterize visual or auditory segments that may cause users to deviate from the core promotional intent. The second determining unit is used to determine the contextualized misreading risk assessment result of the potential multi-meaning element based on the potential multi-meaning element and the preset user context model. The contextualized misreading risk assessment result is used to characterize the possibility and direction of the comprehension deviation caused by the potential multi-meaning element in a specific user viewing context. The generation unit is used to make a decision between the content modification strategy and the audience screening strategy based on the contextualized misreading risk assessment results, the pre-stored promotion constraints and the content modification cost model, so as to generate an initial intervention strategy. The content modification strategy includes adjusting the potential multi-meaning elements in the video to be promoted, and the audience screening strategy includes selecting the test user group. An execution unit is used to execute the initial intervention strategy and optimize the promotion process based on user feedback data obtained after execution.

Images