Media generation recommendation system based on hotspot information

Abstract

The invention discloses a media generation recommendation system based on hot spot information, relates to the technical field of content generation, and solves the technical problems of insufficient accuracy of hot spot recognition and insufficient suitability of audience positioning. According to the method, invalid data elimination, abnormal value filtering, normalization processing and full-link optimization of the aging attenuation factor are innovatively introduced through the popularity scoring model, compared with an existing evaluation mode of only counting the surface interaction amount, the method can effectively filter amount brushing data and extreme value interference, and the hotspot score is dynamically adjusted through the aging factor, so that the evaluation efficiency is improved. According to the method, false hotspots, potential hotspots and core hotspots are accurately distinguished, decline-period hotspots or false hotspots are prevented from being included in a content production range, the accuracy and timeliness of hotspot identification are improved, three-stage screening logic of core tag hierarchical matching, attribute tag enhanced matching and audience tag accurate matching is designed, the suitability of audiences with hotspots and contents is improved, and the accuracy and timeliness of hotspot identification are improved. And precise positioning of thousands of people and thousands of surfaces is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of content generation, in particular to a media generation recommendation system based on hot information. BACKGROUND

[0002] With the rapid development of new media technology, hot information shows the characteristics of fast transmission speed, scattered sources, various forms and personalized audience needs. Currently, the generation and recommendation of media content gradually rely on the mining of hot data and the matching of audience needs.

[0003] However, the existing technology still has the following industry status: it relies on a single channel, the data types are not fully covered, and the collected data lacks standardized cleaning processes, resulting in a large amount of repeated data, garbage data and invalid data, which affects the accuracy of subsequent hot spot identification; Most of the hotness evaluation only simply counts the like, forward and comment data, without filtering invalid data and outliers, and without considering the timeliness decay characteristics of hot spots, which may misjudge false hot spots and expired hot spots as high-value hot spots, affecting the pertinence of subsequent content production; Most of the matching is based on single-dimensional user interest labels, lacks consideration of temporary user interests and rejection preferences, and does not combine regional, time-sensitive, emotional and other attribute labels of hot spots for hierarchical screening, resulting in low accuracy of audience positioning and mismatch between content and user needs. SUMMARY

[0004] To overcome the shortcomings of the prior art, the present application provides a media generation recommendation system based on hot information, which solves the problems of insufficient accuracy of hot spot identification and lack of adaptability of audience positioning.

[0005] To achieve the above purpose, the present application realizes the following technical scheme: a media generation recommendation system based on hot information, comprising: A multi-source hot spot data acquisition module is used to acquire different types of hot spot data, user behavior data and environment data through multiple channels, and to preprocess the acquired hot spot data by removing duplicates, filtering, standardizing and desensitizing to obtain preprocessed hot spot data, which is then transmitted to a hot spot identification processing module; A hot spot identification processing module is used to identify hot spots from the preprocessed hot spot data, assign a unique identifier to each identified hot spot, construct a hotness scoring model based on the like, forward and comment data of the hot spot to calculate a dynamic score, compare the dynamic score with a comparison threshold to filter potential hot spots, combine a classification threshold to filter core hot spots, analyze the attributes of the potential and core hot spots and generate multi-dimensional labels, and transmit the multi-dimensional labels to an audience positioning analysis module; The audience positioning analysis module is used for collecting static data and dynamic data of the user, extracting core interest tags, temporary interest tags and exclusion tags of the user, generating an audience screening result through multi-dimensional tag and user interest tag hierarchical matching screening, and transmitting the audience screening result to the multi-modal content generation module; The multi-modal content generation module is used for determining a target audience group and a content theme direction based on the audience screening result and the multi-dimensional tags, determining a content form according to the attribute tags, determining a content style according to the sentiment tendency attribute, performing quality evaluation on the generated media content, and transmitting the media content meeting the quality requirement to the recommendation information output module. The recommendation information output module is used for displaying the obtained media content to corresponding operating personnel.

[0006] As a further scheme of the present application, the process of constructing the hotness score model by the hot spot recognition processing module includes: The like quantity, forwarding quantity and comment quantity corresponding to the hot spot i are obtained, and i=1, 2, …, j, wherein j represents the category of the hot spot, invalid data is removed to obtain effective like quantity, effective forwarding quantity and effective comment quantity, and outliers are removed, and the filtered effective indicators are normalized to obtain normalized like quantity S norm , forwarding quantity C norm and comment quantity L norm ; According to the formula Score 基础 =W s xS norm +W c xC norm +W z xL norm , the comprehensive score Score 基础 is calculated, the dynamic score is calculated in combination with the timeliness factor, and the timeliness factor formula is T=e -k ×t, t represents the time since the hot content is published, k is the attenuation coefficient, and e is the natural constant.

[0007] As a further scheme of the present application, the multi-dimensional tag generated by the hot spot recognition processing module is as follows: The hot spot with the dynamic score Score 动态 greater than the comparison threshold value is marked as a potential hot spot, and the hot spot with the dynamic score Score 动态 less than the comparison threshold value is marked as a false hot spot, all potential hot spots are obtained, and the dynamic score corresponding to the potential hot spots is compared with the classification threshold value, and the specific value of the classification threshold value is set by the operating personnel; The potential hot spot greater than the classification threshold value is marked as a core hot spot, and the hot spot attributes corresponding to the potential hot spot and the core hot spot are analyzed to generate multi-dimensional tags, including core tags, attribute tags and audience tags.

[0008] As a further aspect of the present invention, the hotspot attributes parsed by the hotspot identification and processing module include: Lifecycle attributes are determined through time series analysis, and sentiment attributes are extracted through sentiment analysis algorithms. Core tags include theme and keywords, attribute tags include region, timeliness, and sentiment, and audience tags include potential interest groups.

[0009] As a further aspect of the present invention, the static data of the audience positioning analysis module is the information filled in by the user during registration, and the dynamic data is the user behavior captured in real time; the core interest tags are extracted based on the user's long-term behavior, the temporary interest tags are extracted based on the user's short-term behavior, and the rejection tags are extracted based on the user's negative behavior.

[0010] As a further aspect of the present invention, the hierarchical matching of multi-dimensional tags and user interest tags in the audience positioning analysis module includes core tag matching, specifically: The core tags in the multi-dimensional tags are matched with the user's core interest tags or temporary interest tags. The matching rules include that the hot core tags are completely consistent with the user tags, the hot core tags are a sub-subset of the user tags, and the word vector cosine similarity is ≥0.7. If any of the matching rules are met, the tags are included, and the core tag filtering results are generated.

[0011] As a further aspect of the present invention, the specific method for generating audience screening results is as follows: Based on the core tag screening results, a second screening is performed on the comprehensive attribute tags, which are then filtered according to regional adaptation, timeliness adaptation, and emotional adaptation. If the three criteria of regional, timeliness, and emotional adaptation are all met or there is no conflict, the target audience can be entered into the attribute adaptation audience pool, and the attribute tag screening results will be generated. Finally, audience tags are used for filtering. The specific filtering method is based on preset matching rules, including positive matching, negative exclusion, and frequency control. Positive matching means that the hot topic audience tag and the user tag have ≥1 intersection. Negative exclusion means that if the user's exclusion tag has an intersection with any hot topic tag, it will be removed. Frequency control means that the same user receives ≤3 push notifications on the same topic within 24 hours. If the number exceeds this, it will not be included in this filtering. The audience filtering results are generated.

[0012] As a further aspect of the present invention, the rule for the multimodal content generation module to determine the content format is as follows: If the attribute tags indicate that the audience is geographically widespread and has high requirements for timeliness, the content should be presented in a combination of text and images and updated in real time; if the audience is concentrated in a specific region and their emotional inclination is light and entertaining, the content should be presented in video format, incorporating local characteristics and humorous expressions; the rules for determining the content style are as follows: when the emotional inclination is positive, the content style should be lively and inspirational; when the emotional inclination is serious and earnest, the content style should be rigorous and professional.

[0013] As a further aspect of the present invention, the quality evaluation of the multimodal content generation module includes five indicators: accuracy, completeness, readability, attractiveness, and adaptability. The specific evaluation rules are as follows: Accuracy: Core information is consistent with authoritative data sources and contains no factual errors, assigned 27 points, weight 30%; Completeness: Covering all key information points of the core tags without omitting any important information, it is awarded 22.5 points, with a weight of 25%; Readability: Text-based Flesch-Kincaid index ≥ 60, video-based speech speed adapted to the audience, assigned 18 points, weight 20%; Attractiveness: Aligned with audience characteristics, awarded 13.5 points, weight 15%; Suitability: Match between content format, style and tag combination ≥ 0.9, awarded 9 points, weight 10%; Calculate the weighted sum of the five indicators. Content with a score of ≥90 is high-quality and output directly. Content with a score of 75-89 is qualified and output after automatic optimization. Content with a score <75 is unqualified and returned for regeneration until the requirements are met.

[0014] This invention provides a media generation and recommendation system based on trending information. Compared with existing technologies, it has the following advantages: This invention integrates multiple channels such as news websites, social media, and government / enterprise announcements, covering multimodal data including text, video, audio, and mixed text and images. It also designs a full-process data cleaning mechanism for deduplication, filtering, standardization, and desensitization. Compared with existing single-channel collection and simple cleaning technologies, this invention reduces the proportion of duplicate and invalid data, improves the accuracy and usability of preprocessed data, and provides high-quality data support for subsequent hotspot identification.

[0015] This invention's popularity scoring model innovatively introduces end-to-end optimization, including invalid data removal, outlier filtering, normalization, and timeliness decay factors. Compared to existing evaluation methods that only count surface interaction volume, it can effectively filter out inflated data and extreme value interference. Furthermore, by dynamically adjusting the hot topic score through timeliness factors, it accurately distinguishes between fake hot topics, potential hot topics, and core hot topics, avoiding the inclusion of declining or fake hot topics in content production, thus improving the accuracy and timeliness of hot topic identification.

[0016] This invention designs a three-level filtering logic: core tag hierarchical matching, attribute tag enhanced matching, and audience tag precise matching. It combines user core interest tags, temporary interest tags, and exclusion tags, while incorporating regional adaptation, timeliness adaptation, emotional adaptation, and push frequency control. Compared with existing simple interest tag matching technology, it can accurately target the audience with interest fit, scenario adaptation, and identity matching, avoid ineffective pushes, improve the adaptability of the audience to hot topics and content, and achieve precise positioning for each individual.

[0017] This invention is driven by multi-dimensional tags of trending topics and audience screening results, dynamically adapting the content format and style, and establishing five quantitative quality evaluation indicators: accuracy, completeness, readability, attractiveness, and adaptability, forming a closed-loop mechanism. Compared with existing technologies that generate fixed templates without quantitative quality control, the content format is more in line with the scenario, the style is more in line with audience preferences, and the quality is more controllable, effectively meeting the personalized needs of different trending topics and different audiences. Attached Figure Description

[0018] Figure 1 A system block diagram for the media generation recommendation system of this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] First Embodiment Please see Figure 1 This application provides a media generation and recommendation system based on trending information, including: a multi-source trending data acquisition module, a trending information identification and processing module, an audience positioning and analysis module, a multimodal content generation module, and a recommendation information output module, and combined with... Figure 1 It can be seen that the information between the above functional modules is transmitted in one direction only.

[0021] The multi-source hotspot data acquisition module is used to collect different types of hotspot data through multiple channels, including news websites, social media, forums, industry platforms, and government / enterprise announcements. The types of hotspot data include text, video, audio, and mixed text and image content. It also collects user behavior data and environmental data. The module performs data cleaning on the collected hotspot data, including deduplication, filtering, and standardization. Deduplication involves filtering duplicate content using MD5 hash values ​​and text similarity algorithms. Filtering removes spam and invalid data and uses a keyword blacklist to initially screen out risky data. Standardization unifies the data format, extracts core metadata, and de-identifies the cleaned hotspot data to obtain pre-processed hotspot data, which is then transmitted to the hotspot identification and processing module.

[0022] The hotspot identification and processing module is used to identify hotspots based on the preprocessed hotspot data. It retrieves all hotspots within the current time period and assigns them a unique label, denoted as i, where i = 1, 2, ..., j, and j represents the type of hotspot. Then, it constructs a popularity scoring model based on the number of likes, shares, and comments for each hotspot, and the specific construction method is as follows: Obtain the number of likes, shares, and comments corresponding to trending topic i, and remove invalid data to obtain the corresponding valid metrics, specifically valid likes, valid shares, and valid comments. Simultaneously, apply the 3σ principle to filter outliers from the obtained valid metrics. Specifically, calculate the mean μ and standard deviation σ of the metrics, remove outliers where "metric value > μ + 3σ", and normalize the valid metrics according to the formula. Calculate the normalized value X of the effective index norm And X norm Including likes S norm , number of reposts C norm and the number of comments L norm , where X 有效 X represents the effective value after filtering. min X represents the minimum value of the indicator. max This indicates the maximum value of the indicator; Next, the normalized values ​​are weighted and summed according to the formula Score. 基础 =W s xS norm +W c xC norm +W z xL norm The overall score is calculated. 基础 Next, dynamic analysis is conducted by combining the time factor, according to the formula T=e -k×t, where t represents the time since the trending content was published, k is the decay coefficient, and e is the natural constant with a value of 2.71, combined with the comprehensive score. 基础 Perform the calculation according to the formula Score 动态 =Score 基础 The dynamic score is calculated by multiplying T by 10. 动态 At the same time, the obtained dynamic score will be... 动态 The score is compared with a corresponding comparison threshold, the specific value of which is set by the operator, to generate the dynamic score. 动态 Hotspots exceeding the comparison threshold are marked as potential hotspots, and a dynamic score is assigned. 动态 Hotspots with values ​​below the comparison threshold are marked as false hotspots; Next, all potential hotspots are acquired, and their corresponding dynamic scores are compared with classification thresholds. The specific values ​​of the classification thresholds are set by the operators. Potential hotspots with scores exceeding the classification thresholds are marked as core hotspots. At the same time, the hotspot attributes corresponding to potential and core hotspots are analyzed. The stage of the hotspot is determined through time series analysis to obtain life cycle attributes. The public's sentiment towards the hotspots is extracted through sentiment analysis algorithms to obtain sentiment tendency attributes, and multi-dimensional tags are generated, including core tags, attribute tags, and audience tags. Core tags include topics and keywords, attribute tags include regions, timeliness, and sentiment, and audience tags include potential interest groups. These tags are then transmitted to the audience positioning analysis module.

[0023] Second Embodiment As a second embodiment of the present invention, it is implemented based on the first embodiment, and the difference from the first embodiment is as follows: The audience targeting and analysis module is used to accurately target and analyze the audience based on the acquired multi-dimensional tags. It collects user data from different users, including static and dynamic data. Specifically, static data represents the information users fill in during registration, while dynamic data represents real-time captured user behavior. Next, it extracts user interest tags based on the user data. These interest tags include core interest tags, temporary interest tags, and aversion tags. Core interest tags are extracted based on long-term user behavior, temporary interest tags are extracted based on short-term user behavior, and aversion tags are extracted based on negative user behavior. The multi-dimensional tags are then matched with the user interest tags, and the specific matching method is as follows: The core tags in the multi-dimensional tags are matched with the core interest tags or temporary interest tags in the user interest tags. According to the preset matching rules, including full matching where the hot core tags are completely consistent with the user tags, matching where the hot core tags are a sub-subset of the user tags, and semantic relevance matching determined by word vector cosine similarity ≥ 0.7, users who meet any of the matching rules can be included. Users who meet the core tags are filtered and core tag filtering results are generated. Based on the core tag screening results, a second screening is performed on the comprehensive attribute tags. The specific screening method is to screen according to the regional adaptation, timeliness adaptation, and emotional adaptation respectively. If the three aspects of regional, timeliness, and emotional all meet the adaptation or have no conflict, they can enter the attribute adaptation audience pool and generate attribute tag screening results. Finally, audience tags are used for filtering. The specific filtering method is based on preset matching rules, including positive matching, negative exclusion, and frequency control. Positive matching means that there is at least one intersection between the hot topic audience tag and the user tag. Negative exclusion means that if the user's exclusion tag has an intersection with any hot topic tag, it will be removed. Frequency control means that the same user receives the same topic hot topic push ≤3 times within 24 hours. If the number exceeds this, it will not be included in this filtering. The audience filtering results are generated and transmitted to the multimodal content generation module.

[0024] The multimodal content generation module is used to generate and recommend media content based on the obtained audience screening results and multi-dimensional tags. First, the target audience is determined based on the audience screening results, and then the content theme direction is clarified by combining the core tags in the multi-dimensional tags. The content format is determined based on the attribute tags in the multi-dimensional tags. If the attribute tags show that the audience is geographically widespread and has high requirements for timeliness, the content will be presented in a combination of text and images and updated in real time. If the audience is concentrated in a specific region and their emotional inclination is towards light entertainment, video format can be used, incorporating local characteristic elements and interesting expressions. In terms of content style, the emotional orientation is taken into account. If the emotional orientation is positive, the content style will be lively and inspirational; if the emotional orientation is serious and earnest, the content style will be rigorous and professional. After media content is generated, it undergoes a quality assessment. Using preset content quality assessment indicators, the content is evaluated and analyzed in terms of accuracy, completeness, readability, attractiveness, and suitability. For accuracy, the core information is analyzed to ensure consistency with authoritative data sources (cross-validation). If no factual errors are found, a passing score of 27 is assigned, with a weight of 30%. For completeness, all key information points of the core tags are analyzed. If no important information is omitted, a passing score of 22.5 is assigned, with a weight of 25%. For readability, text-based content... If the sch index is ≥60 and the Flesch index = 206.835 - (1.015 × average word length) - (84.6 × average sentence length), and the video's speaking speed is suitable for the audience, then a passing threshold of 18 points is assigned and a weight of 20% is given. For attractiveness, it is analyzed whether it fits the audience characteristics. If it fits, a passing threshold of 13.5 points is assigned and a weight of 15% is given. For suitability, it is analyzed whether the content format, style and tag combination match. If it is ≥0.9, a passing threshold of 9 points is assigned and a weight of 10% is given. The obtained values ​​are then weighted and summed to calculate a comprehensive evaluation score. Content with a comprehensive evaluation score of ≥90 is considered high-quality and is directly output. Content with a comprehensive evaluation score of 75-89 is considered qualified and is automatically optimized before being output. Content with a comprehensive evaluation score <75 is considered unqualified and is returned to the generation module for regeneration until it meets the requirements. Finally, the media content that meets the quality requirements is transmitted to the recommendation information output module.

[0025] The recommendation information output module is used to display the acquired media content to the corresponding operators.

[0026] Third Embodiment As a third embodiment of the present invention, the focus is on combining the implementation processes of the first and second embodiments.

[0027] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0028] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A media generation recommendation system based on hotspot information, characterized by, The application comprises the following: A multi-source hot spot data collection module is configured to collect different types of hot spot data, user behavior data and environmental data through multiple channels, to perform deduplication, filtering and standardization processing on the collected hot spot data, to obtain pre-processed hot spot data through desensitization, and to transmit the pre-processed hot spot data to a hot spot identification processing module; The hot spot identification processing module is configured to perform hot spot identification on the pre-processed hot spot data, to assign a unique identifier to each identified hot spot, to construct a hotness scoring model based on the like count, the forwarding count and the comment count of the hot spot to calculate a dynamic score, to compare the dynamic score with a comparison threshold to screen potential hot spots, to combine the comparison threshold to screen core hot spots, to analyze the attributes of the potential hot spots and the core hot spots and to generate multi-dimensional labels, and to transmit the multi-dimensional labels to an audience positioning analysis module; The audience positioning analysis module is configured to collect static data and dynamic data of users, to extract core interest labels, temporary interest labels and exclusion labels of the users, to perform hierarchical matching between the multi-dimensional labels and the user interest labels to screen and generate audience screening results, and to transmit the audience screening results to a multi-modal content generation module; The multi-modal content generation module is configured to determine target audience groups and content theme directions based on the audience screening results and the multi-dimensional labels, to determine content forms based on the attribute labels, to determine content styles based on the emotional tendency attributes, to perform quality evaluation on generated media content, and to transmit media content meeting quality requirements to a recommended information output module; The recommended information output module is configured to display the obtained media content to corresponding operators. 2.The media generation recommendation system based on hotspot information of claim 1, wherein, The process of constructing the hotness scoring model by the hot spot identification processing module comprises: Obtaining the like amount, forwarding amount and comment amount corresponding to the hotspot i, and i = 1, 2, …, j, wherein j represents the category of the hotspot, eliminating invalid data to obtain effective like amount, effective forwarding amount and effective comment amount, eliminating outliers, and normalizing the filtered effective indicators to obtain normalized like amount S norm , forwarding amount C norm , and comment amount L norm ; According to the formula Score 基础 = W s x S norm + W c x C norm + W z x L norm , the comprehensive score Score 基础 is calculated, the dynamic score is calculated in combination with the time factor, and the time factor formula is T = e -k x t, t represents the time since the hot content is published, k is the attenuation coefficient, and e is the natural constant. 3.The media generation recommendation system based on hotspot information of claim 1, wherein, The hot spot identification processing module generates multi-dimensional labels in the following manner: The dynamic score Score 动态 The hotspot greater than the comparison threshold is marked as a potential hotspot, and the dynamic score Score 动态 The hotspot less than the comparison threshold is marked as a false hotspot, all potential hotspots are obtained, and the corresponding dynamic score of the potential hotspots is compared with a classification threshold, and the specific value of the classification threshold is set by an operator. Hot spots greater than the classification threshold are marked as core hot spots, and the attributes of the potential hot spots and the core hot spots are analyzed to generate multi-dimensional labels, including core labels, attribute labels and audience labels.

4. The media generation recommendation system based on hotspot information according to claim 3, wherein, The hot spot attributes analyzed by the hot spot identification processing module comprise: Life cycle attributes determined through time series analysis and emotional tendency attributes determined through emotional analysis algorithms, wherein the core labels comprise themes and keywords, the attribute labels comprise regions, time effectiveness and emotions, and the audience labels comprise potential interest groups. 5.The media generation recommendation system based on hotspot information of claim 1, wherein, The static data of the audience positioning analysis module is information filled in by users during registration, and the dynamic data is real-time captured user behavior; The core interest labels are extracted based on long-term behavior of the users, the temporary interest labels are extracted based on short-term behavior of the users, and the exclusion labels are extracted based on negative behavior of the users.

6. The media generation recommendation system based on hotspot information according to claim 1, wherein, The hierarchical matching between the multi-dimensional labels and the user interest labels in the audience positioning analysis module comprises core label matching, which specifically comprises: The core labels in the multi-dimensional labels are matched with the core interest labels or the temporary interest labels of the users, and the matching rules comprise that the hot spot core labels are completely consistent with the user labels, the hot spot core labels are subsets of the user labels, and the word vector cosine similarity is greater than or equal to 0.7, and any one of the matching rules is included to generate core label screening results.

7. The media generation recommendation system based on hotspot information according to claim 1, wherein, The specific manner of generating audience screening results is that Based on the core label screening result, the attribute label is screened again, and is screened according to the region adaptation, time adaptation and emotion adaptation, if the region, time and emotion meet the adaptation or have no conflict, it can enter the attribute adaptation audience pool, and the attribute label screening result is generated; Finally, the audience label is screened, and the specific screening method is that according to the preset matching rule, including positive matching, reverse exclusion and frequency control, wherein the positive matching means that the intersection of the hot audience label and the user label is greater than or equal to 1, the reverse exclusion means that the user exclusion label has intersection with any label, and then the user is excluded, the frequency control means that the same user receives the same theme hot spot push within 24 hours is less than or equal to 3 times, and the excess is not included in this screening, and the audience screening result is generated. 8.The media generation recommendation system based on hotspot information of claim 1, wherein, The rule for the multi-modal content generation module to determine the content form is: If the attribute label shows that the audience is widely distributed in the region and has high requirements on timeliness, the form of combination of text and picture and real-time update is adopted; if the audience is concentrated in a specific region and the emotional tendency is inclined to be relaxed and entertaining, the video form is adopted, and local characteristic elements and interesting expressions are integrated; the rule for determining the content style is that when the emotional tendency is positive, the content style is lively and inspiring; when the emotional tendency is serious, the content style is rigorous and professional. 9.The media generation recommendation system based on hotspot information of claim 1, wherein, The quality evaluation of the multi-modal content generation module includes five indexes of accuracy, integrity, readability, attractiveness and adaptability, and the specific evaluation rule is that: Accuracy: the core information is consistent with the authoritative data source cross verification, no factual error, 27 points are given, and the weight is 30%; Integrity: all key information points of the core label are covered, no important information is missed, 22.5 points are given, and the weight is 25%; Readability: the Flesch-Kincaid index of text is greater than or equal to 60, and the speech speed of video is adapted to the audience, 18 points are given, and the weight is 20%; Attractiveness: adapt to the audience characteristics, 13.5 points are given, and the weight is 15%; Adaptability: the matching degree of content form, style and label combination is greater than or equal to 0.9, 9 points are given, and the weight is 10%; The weighted sum score of the five indexes is calculated, and the score greater than or equal to 90 is the high-quality content directly output, the score of 75-89 is the qualified content automatically optimized and output, and the score less than 75 is the unqualified content returned to be regenerated until the requirements are met.

Images