AI-based hot topic material generation method and device, electronic equipment, medium and product

Abstract

The application discloses an AI-based hot topic material generation method and device, electronic equipment, medium and product. The method comprises the following steps: performing parallel search on hot topic keywords in multiple data sources to obtain search data; performing hot topic mining on the search data by using a large language model to obtain a hot topic event summary; analyzing the hot topic event summary and the search data by using the large language model to obtain multiple target attention units for generating materials; and generating hot topic material data according to the hot topic event summary, a preset business scenario and the multiple target attention units. The method can improve the efficiency and quality of hot topic material generation.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to an AI-based method, apparatus, electronic device, readable storage medium, and computer program product for generating trending topics. Background Technology

[0002] In today's era of digital content creation, the demand for high-quality content is growing in areas such as social media operations, short video content production, and brand marketing. Current technology typically involves content creators manually browsing trending topics on platforms like Weibo and Douyin, selecting high-profile events based on personal experience, then manually analyzing event elements, discussion focuses, and relationships between key figures, manually crafting creative angles, and finally writing text, creating images, or video clips. However, this entirely manual approach has significant drawbacks: inefficient acquisition of trending topics, strong subjectivity and easy omission of key information, difficulty in ensuring creative quality through manual research, and inability to meet the demands of mass production. Ultimately, this results in low efficiency, inconsistent quality, and difficulty in achieving large-scale production. Summary of the Invention

[0003] In view of the above problems, this application provides an AI-based method, apparatus, electronic device, readable storage medium, and computer program product for generating hotspot materials, which can solve the problems of low efficiency, unstable quality, and inability to generate materials in batches in existing materials.

[0004] Firstly, this application provides an AI-based method for generating trending topics, including: The search data is obtained by performing parallel searches on multiple data sources based on trending keywords. Hotspot mining is performed on the retrieved data using a large language model to obtain a summary of hot events; The hot topic event summary and the retrieved data are parsed using a large language model to obtain multiple target attention units for generating materials; Based on the hot topic event summary, the preset business scenario, and the multiple target attention units, hot topic material data is generated.

[0005] In the above technical solution, the method can efficiently obtain comprehensive hot topic information through parallel retrieval of multiple data sources, and automatically complete hot topic mining, content parsing and attention unit extraction with the help of a large language model, thereby significantly reducing the cost of manual operation and significantly improving the generation efficiency of hot topic materials. At the same time, the method can also reduce the subjective bias and information omission of manual analysis, ensure the creative quality and content standardization of materials, and thus quickly adapt to different business scenarios to produce high-quality hot topic materials in batches, thereby effectively meeting the needs of large-scale and high-efficiency content creation.

[0006] In some implementations, the step of performing parallel searches on multiple data sources based on hot keywords to obtain search data includes: Generate cache keys based on trending keywords; Check whether there is matching non-expired cached data in the memory cache and local storage based on the cache key; When a matching non-expired cached data is found in the memory cache and local storage based on the cache key, the non-expired cached data matching the cache key is retrieved as the search data; When no matching non-expired cached data exists, multiple data collection tasks are constructed based on the hot keywords, each corresponding to one of the multiple data sources. Based on the multiple data acquisition tasks, data is acquired from the multiple data sources in parallel to obtain multiple data acquisition results that correspond one-to-one with the multiple data sources; wherein, the multiple data acquisition results constitute the retrieval data.

[0007] In the above technical solution, the method can reduce duplicate data collection, reduce system resource consumption, and improve the response speed of hot data acquisition through a cache reuse mechanism. At the same time, it can also adopt parallel collection from multiple data sources to ensure the comprehensiveness and integrity of the retrieved data, while taking into account both data collection efficiency and quality.

[0008] In some implementations, the step of performing hotspot mining on the retrieved data using a large language model to obtain a summary of hot events includes: Extract the first key field data from the search data; wherein, the first key field data includes the search results list, the list of popular comments from netizens, and the data source identifier; Based on preset semantic analysis prompts, a large language model is invoked to extract entity recognition results from the first key field data; wherein, the entity recognition results include at least key events, key figures, time and place, sentiment tendencies, and points of contention; The word frequency statistics were performed on the list of popular comments from netizens, and the word frequency statistics results were obtained. Based on the word frequency statistics, several highly discussed and trending topics were extracted from the list of popular online comments. Based on the entity recognition results and the multiple highly discussed hot topics, core figure information is extracted; wherein, the core figure information includes core figure pairs, the relationship type between the core figure pairs, the relationship features between the core figure pairs, and the feature information of each core figure in the core figure pair; Based on the multiple highly discussed trending topics, the entity recognition results, and the key figures' information, a summary of trending events is generated.

[0009] In the above technical solution, the method can accurately mine the core hot information in the retrieved data, comprehensively extract the key elements of the event and the focus of user discussion, and deeply analyze the relationships and emotional tendencies of the people involved, thereby generating a high-quality summary of hot events that is logically clear, informationally complete, and in line with public opinion.

[0010] In some implementations, the step of extracting multiple highly discussed trending topics from the list of popular online comments based on the word frequency statistics results includes: Based on the word frequency statistics, discussion topics in the list of popular online comments are extracted to obtain multiple initial candidate topics; Based on the first key field data, calculate multiple interaction indicators that correspond one-to-one with the multiple initial candidate topics; wherein, the interaction indicators include at least one of the following: comprehensive popularity score, cross-platform coverage, topic growth rate, and emotional polarization index; Based on the multiple interaction metrics, the multiple initial candidate topics were filtered to obtain multiple hot topics with high discussion levels.

[0011] In the above technical solution, the method can accurately filter out real hot topics with high popularity and high dissemination value from the content discussed by netizens, avoiding the bias caused by single word frequency statistics; at the same time, it comprehensively evaluates the value of topics through multi-dimensional interactive indicators, ensuring that the extracted hot topics have high discussion, strong dissemination and high creative value.

[0012] In some implementations, the step of calculating multiple interaction metrics corresponding one-to-one with the multiple initial candidate topics based on the first key field data includes: Based on the list of popular user comments, obtain multiple interaction data points corresponding one-to-one with the multiple initial candidate topics; wherein, the interaction data is used to reflect the degree of user participation in the multiple initial candidate topics; and / or Calculate multiple comprehensive popularity scores corresponding one-to-one with the multiple initial candidate topics based on the multiple interaction data; and / or Each popular comment in the list of popular comments is categorized by sentiment to obtain a list of sentiment types corresponding to the list of popular comments; wherein, the list of sentiment types includes the sentiment type corresponding to each popular comment in the list of popular comments, and the sentiment type is one of positive, neutral, and negative; and / or Based on the list of popular online comments and the list of sentiment types, calculate multiple sentiment polarization indices that correspond one-to-one with the multiple initial candidate topics; and / or Calculate multiple cross-platform coverages corresponding one-to-one with the multiple initial candidate topics based on the data source identifier; and / or Calculate the growth rate of multiple topics that correspond one-to-one with the multiple initial candidate topics based on the search results list; The interaction metrics include the overall popularity score, the cross-platform coverage, the topic growth rate, and the emotional polarization index.

[0013] In the above technical solution, the method can quantify the real dissemination value and public opinion attributes of candidate topics from multiple dimensions such as user interaction, sentiment, dissemination scope, and changes in popularity, thereby accurately calculating comprehensive and objective interaction indicators and avoiding the one-sidedness of single-dimensional evaluation.

[0014] In some implementations, the step of parsing the hot topic event summary and the retrieved data using a large language model to obtain multiple target attention units for generating materials includes: Extract key information from the summary of the hot events; wherein, the key information includes the event name, key figures, controversial topics, points of emotional resonance, and frequently mentioned topics in netizens' comments; Based on the key information in the summary, an attention unit is constructed to mine cue word templates; Based on the key information in the summary and the attention unit mining prompt word template, generate attention unit mining prompt word text; By mining cue word text using a large language model and the attention units, multiple initial attention units are generated; wherein, the fields in the initial attention units include attention unit type, core comparison points, discussion level, list of controversial topics, brief description, key figures, and comparison figures; Each of the multiple initial attention units is subjected to a quality assessment, resulting in multiple assessment results corresponding to each of the multiple initial attention units; wherein, the assessment results include discussion level, completeness assessment result, and comparison point clarity assessment result; Based on the multiple evaluation results, the multiple initial attention units are preprocessed to obtain multiple target attention units.

[0015] In the above technical solution, the method can automatically generate diverse and highly communicable content attention units based on hot topic core information, ensure the standardization of attention unit generation through standardized prompt word templates, and select high-quality attention units by combining multi-dimensional quality assessment to enhance the attractiveness, controversy and resonance of attention units.

[0016] In some implementations, constructing an attention unit mining cue word template based on the summary key information includes: Based on the key information in the summary, determine the types of hot events and the types of priority attention units; Based on the preferred attention unit type and the hot event type, construct a structured attention unit mining prompt word template; The types of trending events include popular TV dramas or real-life events involving people. The priority attention unit types include one or more of the following: image similarity type, multi-dimensional comparison type, and emotional narrative type; The attention unit mining prompt template includes instructions, a summary of hot events, an attention unit type definition, an attention unit type selection strategy, attention unit quality requirements, special scenario processing rules, output format requirements, and requirements for enhancing online appeal and attractiveness.

[0017] In the above technical solution, the method can combine the type of hot events and the type of priority attention units to construct a structured and targeted template for mining attention unit prompts, clarify the various specifications and requirements for the generation of attention units, ensure that the attention units generated subsequently are in line with the hot event attributes and the priority type orientation, and at the same time ensure the standardization, internet appeal and attractiveness of the attention unit generation.

[0018] In some implementations, the preprocessing of the plurality of initial attention units based on the plurality of evaluation results to obtain a plurality of target attention units includes: Based on the multiple evaluation results, the multiple initial attention units are screened to obtain multiple candidate attention units; Determine whether the total number of the plurality of candidate attention units is less than a preset attention unit number threshold; When the total number of the multiple candidate attention units is not less than the preset attention unit number threshold, the multiple candidate attention units are deduplicated to obtain multiple attention units to be verified. Perform diversity verification on the plurality of attention units to be verified, and obtain a plurality of diversity verification results corresponding one-to-one with the plurality of attention units to be verified. Based on the multiple diversity verification results, the attention units that pass the verification are selected from the multiple attention units to be verified, resulting in multiple attention units with high discussion. The attention units that have already been used for material generation among the multiple high-discussion attention units are filtered to obtain multiple target attention units.

[0019] In the above technical solution, the method can accurately select high-quality, diverse and unused target attention units through multiple rounds of screening, deduplication, diversity verification and used attention unit filtering. This ensures the high quality and high discussion of the target attention units, while avoiding the repetition and monotony of attention units, thus ensuring the diversity and innovation of subsequent material creation.

[0020] In some implementations, generating hot topic material data based on the hot topic event summary, a preset business scenario, and the multiple target attention units includes: Obtain information about multiple attention units that correspond one-to-one with the multiple target attention units; Based on the hot topic event summary, preset business scenarios, preset dimension templates, and the information of the multiple attention units, generate multiple material generation prompts that correspond one-to-one with the multiple target attention units; Based on the multiple materials, prompt words are generated, and multiple material generation tasks are constructed that correspond one-to-one with the multiple target attention units; The multiple material generation tasks are executed in parallel to obtain multiple initial material data corresponding one-to-one with the multiple target attention units; The multiple initial material data are verified one by one to obtain multiple material verification results corresponding to the multiple initial material data; wherein, the material verification results include format verification results and quality verification results; Based on the multiple material verification results, select all the initial material data that passed the verification from the multiple initial material data to obtain a candidate material data set; The candidate material data set is deduplicated to obtain hotspot material data.

[0021] In the above technical solution, the method can combine hot topic cores, business scenarios and high-quality target attention units, and efficiently generate initial materials through standardized prompt words and parallel tasks. After format and quality verification and deduplication, it ensures that the hot topic material data meets business needs, is formatted in a standardized manner, meets quality standards and is free of duplication, so as to achieve the batch and efficient production of high-quality hot topic materials and adapt to large-scale content creation scenarios.

[0022] In some embodiments, the method further includes: Extract the second key field from the hot material data; wherein, the second key field includes at least one or more of the following: title, left-side character data object, right-side character data object, and brand embedded content; both the left-side character data object and the right-side character data object contain a field order array and a red-marked array; Based on the data object in the second key field, construct a DOM structure for rendering the visual comparison chart; Receive editing instructions from the user regarding the DOM structure; The DOM structure is edited according to the editing instructions to obtain the target DOM structure; wherein, the editing instructions include at least one of text style editing instructions, image replacement instructions, and layout adjustment instructions; Render the target DOM structure and output a visual comparison chart.

[0023] In the above technical solution, the method can perform structured field parsing on the generated hot material data and automatically construct the DOM structure, thereby supporting users to flexibly perform visual editing operations such as text style, image replacement, and layout adjustment. After the editing is completed, a visual comparison chart is directly rendered and output, thereby realizing online visual editing and real-time preview display of hot materials, and improving the efficiency of material post-layout optimization and output.

[0024] Secondly, this application provides an AI-based hotspot material generation device, comprising: The retrieval unit is used to perform parallel searches across multiple data sources based on hot keywords to obtain retrieval data. The hotspot mining unit is used to perform hotspot mining on the retrieved data using a large language model to obtain a summary of hot events. The parsing unit is used to parse the hot topic event summary and the retrieved data through a large language model to obtain multiple target attention units for generating materials; The generation unit is used to generate hot topic material data based on the hot topic event summary, the preset business scenario, and the multiple target attention units.

[0025] In the above technical solution, the device can efficiently obtain comprehensive hot topic information through parallel retrieval of multiple data sources, and automatically complete hot topic mining, content parsing and attention unit extraction with the help of a large language model, thereby significantly reducing the cost of manual operation and significantly improving the generation efficiency of hot topic materials. At the same time, the method can also reduce the subjective bias and information omission of manual analysis, ensure the creative quality and content standardization of materials, and thus quickly adapt to different business scenarios to produce high-quality hot topic materials in batches, thereby effectively meeting the needs of large-scale and high-efficiency content creation.

[0026] Thirdly, this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor runs the computer program to cause the electronic device to perform the AI-based hotspot material generation method described in any one of the first aspects.

[0027] Fourthly, this application provides a readable storage medium storing a computer program, which, when executed by a processor, performs the AI-based hotspot material generation method described in any one of the first aspects.

[0028] Fifthly, this application provides a computer program product, which includes a computer program that, when executed by a processor, performs the AI-based hotspot material generation method described in any one of the first aspects.

[0029] The beneficial effects of this application are as follows: Firstly, it significantly improves the comprehensiveness and real-time nature of hot topic information collection. Through multi-source parallel search, it comprehensively grasps the dissemination and discussion intensity of hot topics, avoiding the omission of key information. Secondly, it can leverage structured attention units to mine prompt word templates, guiding large language models to automatically generate high-discussion-level material creation directions, thereby greatly improving the automation and quality of creative direction mining. Thirdly, it can generate multiple material variations with consistent formats in parallel through a batch generation engine, significantly improving the efficiency and quality of material generation. Fourthly, it can extract multi-dimensional features of individuals, generating rich and highly discussed comparative content, thereby improving the depth and accuracy of multi-dimensional comparative analysis. Fifthly, it supports real-time preview of materials and various editing operations, enhancing real-time preview and visual editing capabilities and simplifying the editing process. Finally, through caching optimization and streaming response processing, it can significantly improve system performance and response speed, enhancing user experience. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart illustrating the AI-based hotspot material generation method in some embodiments of this application; Figure 2 This is a schematic diagram of a multi-source parallel search process in some embodiments of this application; Figure 3 This is a schematic diagram of an attention unit for mining cue words in some embodiments of this application; Figure 4 This is a schematic diagram of an attention unit mining process in some embodiments of this application; Figure 5 This is a schematic diagram illustrating an example of a material generation process in some embodiments of this application; Figure 6 This is a schematic diagram of the architecture of a hotspot material generation system that performs an AI-based hotspot material generation method in some embodiments of this application; Figure 7This is a schematic diagram of the structure of an AI-based hotspot material generation device in some embodiments of this application; Figure 8 This is a schematic diagram of the structure of an electronic device in some embodiments of this application. Detailed Implementation

[0032] The embodiments of the technical solution of this application will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of this application and are therefore merely examples, and should not be used to limit the scope of protection of this application.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0034] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more (including two), similarly, "multiple sets" refers to two or more sets (including two sets), and "multiple pieces" refers to two or more pieces (including two pieces) unless otherwise explicitly defined.

[0035] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0036] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0037] The current field of generating trending content suffers from pain points such as reliance on manual operation, incomplete acquisition of trending topics, low efficiency in creative mining, inability to produce content in batches and inconsistent quality, and a lack of convenient post-editing capabilities.

[0038] Based on this, this application proposes an AI-based method for generating trending content. The core concept of this method lies in connecting key steps such as multi-source parallel retrieval, deep mining, target attention unit selection, batch content generation, and visual editing through a large language model, forming an automated, end-to-end trending content generation pipeline. This replaces traditional manual processes, effectively addressing pain points in each step of traditional methods, and thus efficiently, efficiently, and on a large scale generating trending content adapted to specific business scenarios, thereby significantly improving the overall efficiency of content creation.

[0039] like Figure 1 As shown, some embodiments of this application provide an AI-based method for generating trending topics, which includes: S100. Perform parallel searches on multiple data sources based on hot keywords to obtain search data.

[0040] In this embodiment, hot keywords are search keywords used to locate current popular events, people and topics on the Internet, and are used to accurately match hot content data sources on various platforms.

[0041] For example, trending keywords can be celebrity names, movie and TV show titles, names of trending social events, names of popular online topics, etc.

[0042] In this embodiment, the method can be pre-connected to multiple trending data sources such as Weibo, Douyin, and Maoyan's popular dramas.

[0043] In this embodiment, in addition to Weibo, Douyin, and Maoyan's trending dramas, it can also be extended to access new trending data sources such as Xiaohongshu, Bilibili, and Zhihu to cover more channels of public opinion dissemination; at the same time, web crawlers, RSS subscriptions and other methods can be used to collect trending information.

[0044] In this embodiment, after completing the unified adaptation of the interface, the method can initiate parallel searches across multiple data sources based on hot keywords, thereby aggregating the results from various platforms to obtain the final search data.

[0045] In this embodiment, the retrieved data is the original data from multiple platforms obtained by parallel retrieval and aggregation. Specifically, it includes structured original data such as hot content items, published information, interactive data, popular comments and reviews from netizens, and data source identifiers from each platform.

[0046] S200. Hotspot mining is performed on the retrieved data using a large language model to obtain a summary of hot events.

[0047] In this embodiment, the method can parse the search data obtained by parallel retrieval and extract core content such as search results, popular comments from netizens, and data source identifiers.

[0048] In this embodiment, the method can invoke a large language model through preset semantic analysis prompts to complete entity recognition of key events, key figures, time and place, emotional tendencies, and points of contention.

[0049] In this embodiment, the large language model is an artificial intelligence language model with semantic understanding, entity recognition, sentiment analysis and topic extraction capabilities. GPT-4, Claude, Doubao and other large language models that meet the requirements can be selected to perform in-depth analysis and mining of hot topic-related information in the retrieved data.

[0050] In this embodiment, in addition to the mainstream GPT series and Claude series models, open source Llama series, GooglePaLM series, domestic Wenxin Yiyan, Tongyi Qianwen and other large language models can also be selected. After fine-tuning for business scenarios, they can all realize the understanding of hot information and the generation of materials.

[0051] In this embodiment, the method can also perform word frequency statistics on popular online comments and extract high-frequency topics.

[0052] In this embodiment, the method can comprehensively use quantitative interaction index weighting, cross-platform coverage bonus, topic growth rate determination, and sentiment polarization index analysis, combined with qualitative evaluation by a large language model, to jointly screen out hot topics with high discussion volume.

[0053] In this embodiment, the method can further identify key figures, types and characteristics of their relationships, and controversial information about their personal backgrounds.

[0054] In this embodiment, the method can ultimately integrate all analysis results to generate a standardized, structured summary of hot events.

[0055] In this embodiment, the hot topic summary is a standardized structured data that integrates the results of retrieval data analysis, entity recognition results, highly discussed hot topics and key figures information. It contains key elements of the event, public opinion focus, and relationships between figures, and is used to provide standardized and complete input data for subsequent attention unit mining.

[0056] In this embodiment, in the field of content creation, attention units can also be called hooks, which are concretized as a type of structured data object. A hook is a commonly used term in the industry, referring to a content entry point that can quickly attract the audience's attention.

[0057] S300: The hot topic event summary and retrieval data are parsed using a large language model to obtain multiple target attention units for generating materials.

[0058] In this embodiment, the method can use the target attention unit to refer to professional terms in the field of content creation and concretize it into a structured data object.

[0059] In this embodiment, the method can perform in-depth analysis of hot event summaries and multi-source retrieval data using a large language model to obtain the basic content of the target attention unit.

[0060] In this embodiment, the method can perform multiple rounds of preprocessing operations on the parsed content, including quality assessment, deduplication, and diversity verification. After preprocessing, it outputs multiple target attention units with high discussion potential and strong propagation attributes.

[0061] In this embodiment, the method can set a fixed and standardized field structure for the target attention unit, which can accurately guide the direction of batch creation of subsequent materials.

[0062] For example, regarding the trending event of "So-and-so announcing their relationship," the target attention unit generated by this method is represented as a standardized structured data example as follows: {"id":"hook-1234567890-0-abc123", "type":"multiDimension", "contrastPoint": "Comparison of the growth paths of so-and-so vs. so-and-so's girlfriend", "discussionLevel":"high", "topics":["Family background and personal assets","Educational style and intergenerational inheritance"], "description": "By comparing the two individuals' upbringing and educational experiences, this sparked online discussions about the impact of one's family of origin, generating significant buzz and discussion." "corePerson":"XXX", "contrastPerson":"So-and-so's girlfriend", "createdAt":1234567890, "selected":true }

[0063] This structured data can be directly fed into the material batch generation engine to complete the automated creation of corresponding trending materials.

[0064] S400 generates hot topic material data based on hot topic event summaries, preset business scenarios, and multiple target attention units.

[0065] In this embodiment, the method can configure exclusive dimension templates and automatically switch dimension fields according to different business scenarios, such as real public figures, fictional characters in film and television, and major social events. For fields lacking public information, fixed text can be used for filling or equivalent dimension replacement, and reasonable error tolerance rules can be set.

[0066] In this embodiment, the method can generate prompt words based on each target attention unit combined with a summary of hot events, and construct exclusive material that includes format requirements, number of dimensions, word count specifications, and title rules.

[0067] In this embodiment, the method can simultaneously construct generation tasks for multiple attention units and call large language models in parallel to generate materials.

[0068] In this embodiment, the method can perform JSON format verification, field integrity verification, and content quality verification on the output content of a large language model.

[0069] In this embodiment, the method can also automatically regenerate content that fails the verification.

[0070] In this embodiment, the method can perform similarity deduplication on verified materials to ensure the differentiation of material variants.

[0071] In this embodiment, the method can ultimately output standardized, compliant, and diverse hot topic material data.

[0072] In this embodiment, in addition to the large language model directly generating complete structured materials in one go, a phased generation path can also be adopted. For example, the large language model can first generate a high-level outline and creation script for the materials, and then use specialized generators such as template systems, diffusion models, and GAN generation networks to transform the outline content into the final standardized materials, thereby further improving the overall structuring level and content consistency of the materials.

[0073] In the above embodiments, the method can efficiently obtain comprehensive hot topic information through parallel retrieval of multiple data sources, and automatically complete hot topic mining, content parsing and attention unit extraction with the help of a large language model, thereby significantly reducing the cost of manual operation and significantly improving the generation efficiency of hot topic materials. At the same time, the method can also reduce the subjective bias and information omission of manual analysis, ensure the creative quality and content standardization of materials, and thus quickly adapt to different business scenarios to produce high-quality hot topic materials in batches, thereby effectively meeting the needs of large-scale and high-efficiency content creation.

[0074] In some embodiments, after step S400, the method may further include: S500, Extract the second key field from the hot material data; wherein, the second key field includes at least one or more of the following: title, left-side character data object, right-side character data object, and brand embedded content; both the left-side character data object and the right-side character data object contain a field order array and a red-marked array.

[0075] In this embodiment, the method can extract key fields such as title, full-dimensional information of the characters on the left and right sides, and brand placement copy from the structured material data output by the large language model, and use them as the second key field.

[0076] For example, this method can extract the key fields title, left, right, and brand from the parsed JSON material data, and perform data integrity verification on the extracted key fields.

[0077] In this embodiment, the method can retain the built-in field sorting identifier array and the key highlight mark array of the person object.

[0078] For example, this method can arrange fields in an ordered manner according to the order specified by the _fieldOrder array (which specifies the order in which fields are displayed).

[0079] In this embodiment, the method can provide raw data support for ordered DOM rendering and highlighting of key fields.

[0080] For example, this method can mark the fields that need to be highlighted in red based on the _highlights array (used to mark which content needs to be highlighted in red).

[0081] In this embodiment, the method can also perform normalization processing on multi-line text content containing line breaks.

[0082] In this embodiment, in addition to using the streaming incremental parsing method of regular expressions combined with state machines, the ANTLR parser can also be used for structured parsing, or the JSON can be parsed uniformly after the large language model has completely output the content.

[0083] S600: Based on the data object in the second key field, construct the DOM structure for rendering the visual comparison chart.

[0084] In this embodiment, the method can automatically create an overall container for the comparison chart, a title container, and independent containers for the information of the left and right figures based on the extracted key fields.

[0085] In this embodiment, the method can strictly follow the field order array to render each person dimension content sequentially.

[0086] In this embodiment, the method can bind corresponding highlighting styles to specified fields in the red-marked array.

[0087] In this embodiment, the method can automatically complete the overall page layout and the DOM structure of element hierarchy.

[0088] In this embodiment, the DOM structure is a complete page element framework used to carry comparison images, titles, character information, and highlighted fields.

[0089] For example, this method can create a DOM container and title element for the comparison chart, as well as content containers on the left and right sides, thereby building the basic page structure for the visual comparison chart.

[0090] S700 receives editing commands input by the user regarding the DOM structure.

[0091] In this embodiment, the method can monitor user interaction behavior on the front-end page in real time.

[0092] In this embodiment, the method can capture user-initiated text editing operations such as font size, color, and bolding.

[0093] For example, this method can listen for the user's text selection operation and receive the corresponding text style editing instructions.

[0094] In this embodiment, the method can capture user-initiated local image uploads and online image retrieval and replacement operations.

[0095] For example, this method can monitor a user's image upload operation and receive corresponding image replacement and editing instructions.

[0096] In this embodiment, the method can capture user-initiated element spacing and alignment adjustment operations.

[0097] For example, this method can listen for user layout adjustment operations and receive corresponding layout adjustment and editing instructions.

[0098] By implementing this method, all of the above operations can be uniformly identified as corresponding editing commands.

[0099] S800. Edit the DOM structure according to the editing instructions to obtain the target DOM structure; wherein, the editing instructions include at least one of the following: text style editing instructions, image replacement instructions, and layout adjustment instructions.

[0100] In this embodiment, the method can parse the operation parameters corresponding to various editing commands.

[0101] In this embodiment, the method can modify the style attributes, image resource address, margin, padding, and alignment of the corresponding DOM element in real time.

[0102] For example, this method can support modifying font size, color, and bolding style when editing text styles.

[0103] For example, this method can support local image upload, image search by keyword, and image cropping using Canvas during image replacement.

[0104] For example, this method can update the URL address of the corresponding source image after image processing is completed.

[0105] For example, this method can support modification of element spacing and alignment when adjusting the layout.

[0106] In this embodiment, the method can preserve the original page structure hierarchy and field order.

[0107] S900: Render based on the target DOM structure and output a visual comparison chart.

[0108] In this embodiment, the method can render the edited target DOM structure in real time, achieving a WYSIWYG page preview.

[0109] In this embodiment, the method can export the complete DOM structure as an image through canvas conversion.

[0110] For example, this method can use the html2canvas library to convert the DOM structure into a Canvas object, then export the converted Canvas object as a PNG image and provide an image download link.

[0111] In this embodiment, the method can support one-click copying of source text content.

[0112] For example, this method can copy the edited material data to the system clipboard with one click.

[0113] In this embodiment, the method can realize a closed-loop business process of previewing, editing, outputting images, and reusing.

[0114] In this embodiment, the method can update the DOM structure immediately after each editing operation to achieve real-time preview.

[0115] In this embodiment, in addition to using the rendering method of DOM combined with Canvas, visualization rendering can also be achieved based on SVG and WebGL technologies, or third-party open-source visualization libraries such as D3.js and ECharts can be connected to complete the rendering of material comparison charts.

[0116] In the above embodiments, the method can perform structured field parsing on the generated hotspot material data and automatically construct the DOM structure, thereby supporting users to flexibly perform visual editing operations such as text style, image replacement, and layout adjustment, and directly render and output a visual comparison chart after editing, thereby realizing online visual editing and real-time preview display of hotspot materials, and improving the efficiency of material post-layout optimization and output.

[0117] In some embodiments, step S100 may include: S110. Generate cache keys based on trending keywords.

[0118] In this embodiment, the method can generate a unique identifier by performing MD5 hash on the input hot keywords and include the data source API version prefix in the cache key to realize a standardized cache key generation rule that isolates by version and by keyword.

[0119] For example, this method can generate a cache key based on keywords, in the format of "search_keyword MD5 hash value".

[0120] For example, the method also supports cache key version management, adding a version number to the cache key in the format "v1_search_keyword MD5 hash".

[0121] S120. Check whether there is matching non-expired cached data in the memory cache and local storage based on the cache key.

[0122] In this embodiment, the method can adopt a two-level cache verification mechanism with memory cache as the priority and local storage as the fallback. The generated cache keys are matched sequentially, and the cache data is verified to be within the preset validity period by using the timestamp.

[0123] For example, this method can first check the memory cache, and if the memory cache is not hit, then check the local storage, and determine whether the data has expired by the cache timestamp.

[0124] S130. When a matching non-expired cached data is found in the memory cache and local storage based on the cache key, the non-expired cached data matching the cache key is retrieved as the retrieval data.

[0125] In this embodiment, when the cache is hit and has not expired, the method can directly read the cached data and load it synchronously into the memory cache, saving repeated interface requests and directly reusing it as the retrieval data for this parallel retrieval.

[0126] S140. When there is no matching non-expired cached data, construct multiple data collection tasks corresponding one-to-one with multiple data sources based on hot keywords.

[0127] In this embodiment, when the cache misses or expires, the method can independently create keyword retrieval and collection tasks for each data source such as Weibo and Douyin, adapting to the request parameters, request headers and interface specifications of each platform interface.

[0128] S150. Based on multiple data acquisition tasks, data is acquired from multiple data sources in parallel to obtain multiple data acquisition results that correspond one-to-one with the multiple data sources; among them, the multiple data acquisition results constitute the retrieval data.

[0129] In this embodiment, the method can execute multiple data source collection tasks simultaneously in a parallel and asynchronous manner, configure an independent timeout and an incremental retry mechanism for each task, and ensure that the failure of a single data source request does not affect the execution of other tasks. After completion, the results are independently collected according to the data source and aggregated to form complete retrieval data.

[0130] For example, this method can use Promise.all (a front-end asynchronous programming API used to execute multiple asynchronous tasks in parallel, wait for all tasks to complete and return the result in a unified manner to achieve multi-task concurrent processing) to initiate data collection tasks on multiple platforms in parallel, set an independent 5-second timeout for each task, and retry up to 3 times after the task fails, with retry intervals of 1 second, 2 seconds and 3 seconds respectively.

[0131] In this embodiment, when a single data source task fails, only an error log is recorded without interrupting the execution of other normal tasks. Finally, the execution results of each data source are stored independently and integrated to obtain complete retrieval data.

[0132] In the above embodiments, the method can reduce duplicate data collection, reduce system resource consumption, and improve the response speed of hot data acquisition through a cache reuse mechanism; at the same time, it can also adopt parallel collection from multiple data sources to ensure the comprehensiveness and integrity of the retrieved data, while taking into account both data collection efficiency and quality.

[0133] In some embodiments, before accessing the hotspot data source in step S100, the method further includes: S101. Initialize the data source configuration.

[0134] In this embodiment, the method maintains independent configuration information for each hot data source such as Weibo, Douyin, and Maoyan's popular dramas. The configuration information includes API address, authentication information (such as APIKey, Token, etc.), and request parameters (such as request headers, timeout, etc.).

[0135] In this embodiment, the method can store the configuration information of all data sources in a configuration file in JSON or YAML format, and automatically load all configurations when the method is implemented to establish a data source configuration mapping table, providing basic configuration support for subsequent data source calls.

[0136] S102. Perform a health check on the data source.

[0137] In this embodiment, the method can send requests to the health check endpoints of each data source every 5 minutes to detect the service availability of the data source in real time.

[0138] In this embodiment, the method can synchronously record the availability or unavailability status of each data source, and update and maintain the data source status table.

[0139] In this embodiment, the method can automatically disable the data source when it is detected that the data source is unavailable, thereby avoiding invalid requests from causing subsequent process abnormalities.

[0140] In this embodiment, the method can monitor the health status of multiple data sources through periodic polling.

[0141] S103, Unified data interface adaptation.

[0142] In this embodiment, the method can predefine a unified data return format specification and implement an adaptation conversion function independently for each data source to convert the original response data of different platforms into a unified format.

[0143] In this embodiment, the method can also simultaneously handle differences in data sources, internal endpoints, rate limiting authentication anomalies, and pagination logic, ensuring stable and complete data collection, and supporting automatic API version adaptation and compatibility processing.

[0144] In this embodiment, the method can perform unified mapping and conversion at the adapter layer to address the differences in fields, structures, and versions of APIs from different platforms.

[0145] In this embodiment, the method can maintain independent sub-adapters for different API endpoints of the same data source to complete the structure conversion.

[0146] For example, when an HTTP 429 rate limiting exception is triggered, an exponential backoff retry strategy of 1 second, 2 seconds, and 4 seconds is executed; When an HTTP 401 / 403 authentication failure occurs, the token will be automatically refreshed and the request will be resent.

[0147] In this embodiment, the logic for different pagination methods is uniformly encapsulated to provide a consistent pagination interface to the outside world.

[0148] In this embodiment, the method declares the supported API version range in the data source configuration file and identifies the actual version number through the response header or response body.

[0149] In this embodiment, the method can maintain a version compatibility matrix and automatically match the corresponding adapter according to the version number; when a new version that is not adapted is encountered, it automatically downgrades to use the most recently compatible version adapter and records the alarm.

[0150] In this embodiment, the method triggers a cache key prefix update when a version change is detected, thus avoiding contamination between old and new data.

[0151] S104, Multi-source Parallel Search.

[0152] In this embodiment, the method can receive hot keywords input by the user, construct data collection tasks corresponding to multiple platforms, first check whether there are valid results in the cache, and if so, return directly.

[0153] For example, this method can use Promise.all (a front-end asynchronous programming API used to execute multiple asynchronous tasks in parallel, wait for all tasks to complete and return the result in a unified manner to achieve multi-task concurrent processing) to implement multi-platform concurrent requests, and configure independent timeouts and tiered retry mechanisms for each task.

[0154] In this embodiment, if a request fails on a single platform, only the error is recorded, which does not affect the processing of results on other platforms, and the data on each platform is displayed independently.

[0155] In this embodiment, the search results obtained in parallel are deduplicated using the Jaccard similarity algorithm (an algorithm used to calculate the degree of overlap and similarity between two items), and sorted according to multiple dimensions such as relevance, popularity, and time. The search progress is displayed in real time with a streaming response. The final results are synchronously stored in memory cache and local storage (localStorage), and a 1-hour cache expiration time is set. The cache is automatically cleaned up after expiration, which improves the response efficiency of repeated searches.

[0156] In this embodiment, search results are cached for 1 hour, attention unit results are cached for 24 hours, and material cache is permanently valid.

[0157] In this embodiment, the cache is compressed and stored, supporting preheating and preloading to improve system response speed.

[0158] Please refer to Figure 2 , Figure 2 A schematic diagram of a multi-source parallel search process is shown.

[0159] In this embodiment, in addition to adopting a two-level caching strategy that combines memory caching with localStorage, Redis can also be connected to achieve distributed caching deployment, and CDN can also be used to accelerate the edge caching of static material resources.

[0160] In some embodiments, step S200 may include: S210. Extract the first key field data from the search data; wherein, the first key field data includes the search results list, the list of popular comments from netizens, and the data source identifier.

[0161] In this embodiment, the method can perform structured parsing processing on the raw data returned by multi-source parallel retrieval.

[0162] In this embodiment, the method can uniformly extract a list of search results from the original data. The list includes the title, content, publication time, and metrics such as likes, comments, and reposts.

[0163] For example, this method can parse trending event data and extract a list of search results containing title, content, publication time, number of likes, and number of comments.

[0164] In this embodiment, the method can extract a list of popular comments from the original data. The list includes comment content and related interaction data.

[0165] For example, this method can parse trending event data and extract a list of popular comments from netizens, including comment content, number of likes, and number of replies.

[0166] In this embodiment, the method can extract identification information from the raw data to distinguish the platform source.

[0167] For example, this method can parse trending event data and extract data source identifiers corresponding to platforms such as Weibo and Douyin.

[0168] In this embodiment, the method can receive hotspot event data output from multi-channel parallel search.

[0169] S220. Based on the preset semantic analysis prompts, call the large language model to extract the entity recognition results from the first key field data; wherein, the entity recognition results include at least key events, key figures, time and place, sentiment tendencies and points of contention.

[0170] In this embodiment, the method can pre-set a fixed semantic analysis prompt word template, inject the content of the key search fields into the template, and then call the large language model to automatically complete event summarization, person entity extraction, spatiotemporal information positioning, overall sentiment judgment, and extraction of public opinion controversy points.

[0171] In this embodiment, the method can construct semantic analysis prompts in a specified format, call the large language model API to complete semantic understanding, and parse the model's returned results to extract the corresponding entity information.

[0172] For example, the semantic analysis prompt word format constructed by this method is as follows: Analyze the following trending events and extract key information: Event content: ${hotspotContent} Please extract: 1. Key events 2. Key figures 3. Time and place 4. Emotional inclination 5. Points of contention.

[0173] HotspotContent refers to trending content.

[0174] S230. Perform word frequency statistics on the list of popular comments from netizens to obtain the word frequency statistics results.

[0175] In this embodiment, the method can segment and filter stop words in all netizens' comment texts, count the frequency of high-frequency words and phrases, and generate word frequency distribution statistics that can be used for topic mining.

[0176] In this embodiment, the method can extract a list of popular comments from netizens, count high-frequency words and phrases, and analyze the distribution of sentiment tendencies in the popular comments.

[0177] S240. Based on word frequency statistics, extract several highly discussed trending topics from the list of popular online comments.

[0178] In this embodiment, the method initially screens candidate topics based on word frequency statistics, and then combines multi-dimensional quantitative indicators and a large language model for comprehensive evaluation. This allows the selection of hot topics with high popularity, high growth rate, cross-platform dissemination, and high controversy. The multi-dimensional quantitative indicators and the large language model include interactive indicators such as comprehensive popularity score, cross-platform coverage, topic growth rate, and emotional polarization index.

[0179] For example, this method can extract frequently controversial topics and emotional resonance points, prioritizing the extraction of topics that appear repeatedly in hot comments and are likely to trigger discussion.

[0180] To avoid biases caused by relying solely on word frequency statistics and to ensure that the extracted trending topics have high discussion volume, strong dissemination potential, and high creative value, step S240 may also include: S241. Based on the word frequency statistics, extract the discussion topics from the list of popular comments from netizens to obtain multiple initial candidate topics.

[0181] In this embodiment, the method can summarize the mainstream discussion directions in netizens' hot comments based on word frequency clustering, and extract and summarize multiple semantically independent initial candidate discussion topics.

[0182] S242. Calculate multiple interaction indicators that correspond one-to-one with multiple initial candidate topics based on the data of the first key field; among them, the interaction indicators include at least one of the following: comprehensive popularity score, cross-platform coverage, topic growth rate, and emotional polarization index.

[0183] In this embodiment, the method can use the interaction volume, platform source, time series, and sentiment distribution of comments in the retrieved data to quantify and calculate four major indicators for each candidate topic: overall popularity, cross-platform coverage, short-term popularity growth rate, and degree of polarization between positive and negative comments.

[0184] As an optional implementation, multiple interaction metrics are calculated based on the first key field data, each corresponding to one of the multiple initial candidate topics, including: Based on the list of popular user comments, multiple interaction data points were obtained, each corresponding to one of the initial candidate topics; these interaction data reflect the degree of user participation in the multiple initial candidate topics; and / or Calculate multiple comprehensive popularity scores based on multiple interaction data points, each corresponding to one of the multiple initial candidate topics; and / or Each popular comment in the list of popular comments is categorized by sentiment, resulting in a list of sentiment types corresponding to the list of popular comments. The sentiment type list includes the sentiment type corresponding to each popular comment in the list, which can be one of the following: positive, neutral, or negative; and / or Based on the list of popular online comments and the list of sentiment types, calculate multiple sentiment polarization indices that correspond one-to-one with the multiple initial candidate topics; and / or Calculate cross-platform coverage based on data source identifiers, corresponding one-to-one with multiple initial candidate topics; and / or Calculate the growth rate of multiple topics that correspond one-to-one with multiple initial candidate topics based on the search results list; The interaction metrics include at least one of the following: overall popularity score, cross-platform coverage, topic growth rate, and emotional polarization index.

[0185] In this embodiment, a comprehensive popularity score can be calculated based on interaction data such as the number of likes, comments, and reposts for each candidate topic. The comprehensive popularity score is calculated by weighting comments, likes, and reposts; individual comments are labeled with three categories of sentiment, and the degree of polarization is calculated based on the positive and negative ratios; the number of platforms on which the topic appears is counted to obtain cross-platform coverage; and the incremental increase in new content is counted within a fixed time window to calculate the month-on-month growth rate of topic popularity, achieving a comprehensive quantitative evaluation.

[0186] For example, the method can calculate a comprehensive popularity score by weighting the number of comments at 40%, the number of likes at 30%, and the number of reposts / shares at 30%, and mark topics with a score of more than 10,000 as candidate topics with high discussion.

[0187] For example, the method can also apply a 1.5x weighting to the popularity score of topics that spread across multiple platforms such as Weibo and Douyin.

[0188] For example, the method can also count the growth rate of new discussions on a topic within one hour, and mark topics whose growth rate exceeds the historical average by two standard deviations as rapidly rising hot topics.

[0189] For example, this method can also identify topics where both positive and negative comments account for more than 30% as highly controversial topics.

[0190] S243. Based on multiple interaction indicators, multiple initial candidate topics are screened to obtain multiple hot topics with high discussion.

[0191] In this embodiment, the method can preset thresholds for each indicator and weighted scoring rules to comprehensively evaluate the initial candidate topics, eliminate topics with low popularity, low controversy, and low dissemination potential, and retain hot topics that meet the high discussion standards.

[0192] In this embodiment, the method can input the initial candidate topics and corresponding quantitative indicators into the large language model to obtain the discussion level output by the model.

[0193] In this embodiment, the method can identify topics that meet the quantitative indicators and are rated as high by the model as hot topics with high discussion.

[0194] For example, this method can perform initial screening of candidate topics based on preset weighted scoring rules, filter out low-quality topics with no dissemination value, and narrow the scope of subsequent hot topic determination.

[0195] S250. Extract core figure information based on entity recognition results and multiple highly discussed hot topics; among which, core figure information includes core figure pairs, relationship types between core figure pairs, relationship features between core figure pairs, and feature information of each core figure in a core figure pair.

[0196] In this embodiment, the method can combine the identified person entities with popular topics to form a comparison group of people, automatically identify the relationship types between people such as lovers, partners, and competitors, extract differences in characteristics such as age, background, and identity, and organize characteristic information such as occupation, education, family, and past controversies of individual people.

[0197] S260. Generate a summary of hot events based on multiple highly discussed trending topics, entity recognition results, and key figure information.

[0198] In this embodiment, the method can integrate basic event information, highly discussed topics, points of emotional controversy, key figures and relationship features, and organize them into a hot event summary with a fixed JSON structure, providing a standardized input data source for subsequent attention unit mining.

[0199] For example, this method can integrate semantic analysis, topic extraction, and key figure information into structured data to generate a summary of trending events containing event name, key figures, topic, emotional resonance points, points of controversy, popular comments, and summary content, in the following format: { "eventName":"So-and-so announces their relationship", "corePersons":["XXX","XXX's girlfriend"], "topics":["family background disparity","educational background"], "emotionalPoints":["Age difference","Regional identity"], "controversyPoints":["Identity Background","Career Development"], "hotComments":[...], "summary": "Summary of trending events" }

[0200] In the above embodiments, the method can accurately mine the core hot information in the retrieved data, comprehensively extract the key elements of the event and the focus of user discussion, and deeply analyze the relationships and emotional tendencies of the people involved, thereby generating a high-quality hot event summary that is logically clear, informationally complete, and in line with public opinion trends.

[0201] In some embodiments, step S300 may include: S310. Extract key information from the summary of hot events; the key information includes the event name, key figures, controversial topics, points of emotional resonance, and frequently mentioned topics in netizens' comments.

[0202] In this embodiment, the structured fields of the hot topic event summary are parsed, and the event name, combination of people, controversial topics, emotional resonance direction and high-frequency hot comments are extracted as the basic elements for attention unit type determination and prompt word construction.

[0203] S320. Construct attention unit mining prompt word templates based on key information in the abstract.

[0204] In this embodiment, the event attributes and creative adaptation direction are determined based on the key information of the summary. A two-layer logic of rule detection and model generation is adopted to build a special prompt word template that includes type constraints, quality constraints and output format constraints.

[0205] To ensure that the subsequently generated attention units align with trending topics and conform to priority type guidelines, while also guaranteeing the standardization, internet appeal, and attractiveness of the attention unit generation, step S320 may further include: S321. Determine the hot topic event type and priority attention unit type based on the key information in the abstract.

[0206] In this embodiment, the method can distinguish between events related to popular TV dramas and events involving real public figures through keyword matching.

[0207] In this embodiment, the method can prioritize the determination of the attention unit type of hot events based on the hot comment keyword rule. The attention unit types include image similarity type, multi-dimensional comparison type, and emotional narrative type.

[0208] For example, when detecting words related to appearance or resemblance, this method prioritizes classifying them as image-similar attention units.

[0209] For example, when detecting words related to background or family circumstances, this method prioritizes classifying them as multi-dimensional contrastive attention units.

[0210] For example, when detecting emotional or inspirational words, the method prioritizes identifying them as emotional narrative attention units.

[0211] In this embodiment, the method can force the priority generation of character comparison attention units for film and television events.

[0212] In this embodiment, the method can analyze hot keywords, determine whether an event is a popular TV series or movie, and set special processing flags when film and television-related content is identified.

[0213] In this embodiment, the method can determine the priority attention unit type of hot events through keyword scanning.

[0214] S322. Based on the priority attention unit type and hot event type, construct a structured attention unit mining prompt word template.

[0215] In this embodiment, the types of trending events include popular TV dramas or real-life events. Priority attention unit types include one or more of the following: image similarity type, multi-dimensional contrast type, and emotional narrative type; The attention unit mining prompt template includes instructions, a summary of hot events, an attention unit type definition, an attention unit type selection strategy, attention unit quality requirements, special scenario handling rules, output format requirements, and requirements for enhancing online appeal and attractiveness.

[0216] In this embodiment, the method can construct a structured attention unit mining prompt word template for the determined hot event type (such as popular TV series, real-life events) and the priority attention unit mining type (such as image similarity type, multi-dimensional comparison). Specifically, the generation rules, type definitions, high discussion hard requirements, film and television special rules, fixed JSON output format, and online topic mining requirements can all be solidified into the prompt word template according to the hot event type and the priority attention unit mining type, thus constraining the generation boundary of the large language model.

[0217] In this embodiment, in addition to fixed structured prompt templates, prompt engineering methods such as few-shot learning and chain-of-thought can also be used. Specifically, this method can embed standard attention unit examples into the prompts to guide the model's learning and generation paradigm, or it can adopt a step-by-step decomposition approach, first completing the analysis of hot events and then creating attention units, thereby adapting to the efficiency and controllability requirements of attention unit generation in different scenarios.

[0218] S330. Generate attention unit mining prompt text based on the key information in the abstract and the attention unit mining prompt word template.

[0219] In this embodiment, the method can inject key information variables of the abstract into the template and splice them together to form a complete set of prompt words with clear constraints and a well-defined format that can be directly input into a large model.

[0220] For example, this method can inject hot event data into a prompt word template to generate complete attention unit mining prompt words.

[0221] S340. Mine cue word text through large language model and attention units to generate multiple initial attention units; among them, the fields in the initial attention units include attention unit type, core comparison point, discussion degree, list of controversial topics, brief description, key figures, and comparison figures.

[0222] In this embodiment, the large language model strictly follows the prompt word constraints to output structured initial attention units in batches, with fixed required fields such as type of carry, comparison point, discussion level, topic, and person.

[0223] For example, this method can call the Large Language Model API, receive streaming responses, parse JSON-formatted attention unit data in real time, and add a unique ID and timestamp to each attention unit.

[0224] S350. Perform quality assessment on multiple initial attention units respectively to obtain multiple assessment results corresponding to each initial attention unit; among which, the assessment results include discussion level, integrity assessment results, and comparison point clarity assessment results.

[0225] In this embodiment, the method can verify each attention unit's discussion level as high, whether the fields are complete, and whether the core comparison points are clear and actionable, thus forming a multi-dimensional quality assessment result.

[0226] For example, the attention unit discussionLevel must be high and must include the main character, the contrasting character, and the main contrast point.

[0227] S360. Based on multiple evaluation results, preprocess multiple initial attention units to obtain multiple target attention units.

[0228] In this embodiment, the method can filter, supplement, deduplicate, verify diversity, and filter historically used attention units layer by layer based on the quality assessment results, and finally output a sufficient number of usable target attention units that are diverse in type, non-repeating, and highly topical.

[0229] To ensure high quality and high level of discussion for the target attention units while avoiding repetition and monotony of attention units, step S360 may also include: S361. Based on multiple evaluation results, multiple initial attention units are screened to obtain multiple candidate attention units.

[0230] In this embodiment, the method can eliminate unqualified initial attention units that do not meet the discussion level, have missing fields, or have unclear comparison points, and retain qualified candidate attention units.

[0231] For example, this method can filter all attention units with a discussion level of "high" and mark them as selected.

[0232] S362. Determine whether the total number of multiple candidate attention units is less than the preset attention unit number threshold.

[0233] In this embodiment, the method can count the total number of qualified candidate attention units and compare it with a preset minimum threshold. If the number is insufficient, the large model will be automatically regenerated to supplement the number.

[0234] For example, if the number of high-discussion attention units is less than 10, the large language model is called again to generate them until the number requirement is met.

[0235] S363. When the total number of multiple candidate attention units is not less than the preset attention unit number threshold, the multiple candidate attention units are deduplicated to obtain multiple attention units to be verified.

[0236] In this embodiment, the method can perform deduplication based on the semantic similarity of the core comparison points (i.e., the semantic similarity of the core comparison points of the attention units), eliminating attention units with highly overlapping ideas and expressions, and preventing creative duplication.

[0237] For example, this method can identify duplicate attention units based on the similarity of core comparison points to achieve deduplication.

[0238] S364. Perform diversity verification on multiple attention units to be verified, and obtain multiple diversity verification results that correspond one-to-one with the multiple attention units to be verified.

[0239] In this embodiment, the method can verify the degree of difference in attention unit type distribution, comparison perspective, and topic direction, ensuring coverage of at least two or more attention unit types.

[0240] For example, this method can verify that the attention unit covers multiple types, such as image similarity, multi-dimensional comparison, and emotional narrative, to ensure diversity.

[0241] S365. Based on multiple diversity verification results, select the attention units that pass the verification from multiple attention units to be verified, and obtain multiple attention units with high discussion.

[0242] In this embodiment, the method can retain attention units with balanced type distribution and satisfactory perspective differentiation, forming a high-quality attention unit set with high discussion volume.

[0243] S366. Filter the attention units that have been used for material generation from multiple high-discussion attention units to obtain multiple target attention units.

[0244] In this embodiment, the method can compare historical generation records, filter out used attention units, avoid creative duplication, and output new and usable target attention units.

[0245] For example, this method can tag attention units of already generated material and filter to avoid duplicate generation.

[0246] Please refer to Figure 3 , Figure 3 A schematic diagram of an attention unit for mining cue words is shown.

[0247] Please refer to Figure 4 , Figure 4 A schematic diagram of an attention unit mining process is shown.

[0248] In the above embodiments, the method can automatically generate diverse and highly communicable content attention units based on hot topic core information, ensure the standardization of attention unit generation through standardized prompt word templates, and select high-quality attention units by combining multi-dimensional quality assessment to enhance the attractiveness, controversy and resonance of attention units.

[0249] In some embodiments, step S400 may include: S410. Obtain information on multiple attention units that correspond one-to-one with multiple target attention units.

[0250] In this embodiment, the method can read the complete structured fields of each target attention unit and extract the attention unit type, core comparison points, dual-person information, and related controversial topics as the basic input parameters for material generation.

[0251] For example, the method can receive 10-15 high-discussion-level attention units from the attention unit mining output, filter the attention units that have generated material, and determine that each attention unit generates 4 material variants.

[0252] S420: Based on the summary of hot events, preset business scenarios, preset dimension templates, and information on multiple attention units, generate multiple material generation prompts that correspond one-to-one with multiple target attention units.

[0253] In this embodiment, the method can match templates of three dimensions—person, role, and event—according to business scenarios, and combine hot topic summaries and attention unit information to customize exclusive material generation prompt words that include title specifications, number of dimensions, number of characters per field, and rules for filling missing fields.

[0254] For example, the method can construct structured material generation prompts for each attention unit, including generation instructions, hot event content, attention unit information, format requirements, and quality requirements; and adapt corresponding dimension templates for real people, film and television characters, and major events, and fill in "no public information" for missing information, ensuring that there are at least 10 dimensions on each side.

[0255] S430. Generate prompt words based on multiple materials, and construct multiple material generation tasks that correspond one-to-one with multiple target attention units.

[0256] In this embodiment, the method can independently create a generation task for each prompt word, configure the number of variants generated by a single attention unit, task timeout, and failure retry strategy to form a task queue that can be scheduled in parallel.

[0257] S440: Execute multiple material generation tasks in parallel to obtain multiple initial material data corresponding to multiple target attention units.

[0258] In this embodiment, the method can simultaneously execute the multi-attention unit material generation task in parallel, monitor the streaming output of the large language model, and parse and splice the complete initial material structured data in real time.

[0259] For example, this method can use Promise.all (a front-end asynchronous programming API used to execute multiple asynchronous tasks in parallel, wait for all tasks to complete and return the result in a unified manner to achieve multi-task concurrent processing) to process the material generation tasks of multiple attention units in parallel, and set up progress tracking and error handling for each task.

[0260] S450. Verify multiple initial material data separately to obtain multiple material verification results that correspond one-to-one with the multiple initial material data; among them, the material verification results include format verification results and quality verification results.

[0261] In this embodiment, the method can first perform JSON format and required field validity checks; then perform quality checks on the number of dimensions, single field character count, title keyword matching, and missing field tolerance.

[0262] For example, this method can perform format validation to check the validity of JSON, required fields and data types; and quality validation to check the number of dimensions on each side, the number of words in fields such as experience, and the compliance of the title. If the requirements are not met, the data will be regenerated.

[0263] S460. Based on the verification results of multiple materials, select all the initial material data that have passed the verification from multiple initial material data to obtain a candidate material data set.

[0264] In this embodiment, the method can filter initial materials that are legally formatted, meet the content dimension requirements, and comply with business specifications, while eliminating invalid materials that are incomplete, violate regulations, or fail to meet standards.

[0265] S470. Perform deduplication on the candidate material data set to obtain hot material data.

[0266] In this embodiment, the method can perform global deduplication based on the similarity between the title and the core dimension content, retaining differentiated material variations, and finally outputting standardized hot material data that can be directly used for rendering and editing. Although deduplication has been performed in the S363 attention unit, due to the uncertainty in the large language model generation process, highly similar materials may still appear, so secondary deduplication is required.

[0267] Please refer to Figure 5 , Figure 5 An example diagram illustrating a material generation process is shown.

[0268] In the above embodiments, the method can combine hot topic cores, business scenarios and high-quality target attention units, and efficiently generate initial materials through standardized prompt words and parallel tasks. After format and quality verification and deduplication, it ensures that the hot topic material data meets business needs, is formatted correctly, meets quality standards and is free of duplication, and achieves batch and efficient production of high-quality hot topic materials, which is suitable for large-scale content creation scenarios.

[0269] Please refer to Figure 6 , Figure 6 A schematic diagram of the architecture of a hotspot material generation system that can execute the above-described AI-based hotspot material generation method is shown. Figure 6 The system architecture shown is Figure 1 The methods and processes shown correspond one-to-one.

[0270] The hot topic material generation system includes: 1. Hot Data Source Access Module: Responsible for accessing multiple data sources such as Weibo, Douyin, and Maoyan's trending dramas, providing a unified data interface.

[0271] 2. Multi-channel parallel search module: responsible for simultaneously collecting hot information from multiple data sources, realizing real-time aggregation and independent display of search results.

[0272] 3. Hot Topic Mining Engine: Responsible for semantic analysis and topic extraction of collected hot topic information, and identifying hot events and key figures with high discussion volume.

[0273] 4. Attention Unit Mining and Cue Word Engineering Module: Responsible for building structured cue word templates to guide the large language model in generating high-discussion-level material creation directions (attention units), ensuring that only attention units with high discussion levels are output.

[0274] 5. Material Batch Generation Engine: Responsible for generating multiple material variations in parallel using a large language model based on high-discussion attention units.

[0275] 6. Structured Material Parser: Responsible for parsing structured material data from the streaming output of the large language model in real time.

[0276] 7. Real-time preview rendering engine: Responsible for rendering structured material data into visual comparison charts, supporting real-time preview and visual editing.

[0277] 8. Cache optimization module: Responsible for caching intermediate data to avoid duplicate requests and improve system performance.

[0278] like Figure 7 As shown, some embodiments of this application provide a structural schematic diagram of an AI-based hotspot material generation device. It should be understood that this device is related to... Figure 1The method executed in the middle corresponds to the steps involved in the aforementioned method. The specific functions and effects of the device can be found in the description above. To avoid repetition, detailed descriptions are omitted here.

[0279] The AI-based hotspot material generation device includes: The retrieval unit 1010 is used to perform parallel retrievals from multiple data sources based on hot keywords to obtain retrieval data. Hotspot mining unit 1020 is used to mine hotspots in the retrieved data using a large language model to obtain summaries of hot events. Parsing unit 1030 is used to parse the hot event summary and retrieved data through a large language model to obtain multiple target attention units for generating materials; The generation unit 1040 is used to generate hot topic material data based on hot topic event summaries, preset business scenarios, and multiple target attention units.

[0280] In some embodiments, the retrieval unit 1010 includes: The first generation subunit 1011 is used to generate cache keys based on hot keywords; The inspection subunit 1012 is used to check whether there is matching non-expired cached data in the memory cache and local storage based on the cache key; The first acquisition subunit 1013 is used to acquire the non-expired cache data that matches the cache key as the retrieval data when checking whether there is matching non-expired cache data in the memory cache and local storage according to the cache key. The first construction subunit 1014 is used to construct multiple data collection tasks corresponding one-to-one with multiple data sources based on hot keywords when there is no matching non-expired cached data. The acquisition subunit 1015 is used to acquire data from multiple data sources in parallel according to multiple data acquisition tasks, and obtain multiple data acquisition results corresponding one-to-one with the multiple data sources; wherein, the multiple data acquisition results constitute the retrieval data.

[0281] In some embodiments, the hotspot detection unit 1020 includes: The first extraction subunit 1021 is used to extract the first key field data from the retrieved data; wherein, the first key field data includes the search results list, the list of popular comments from netizens, and the data source identifier; Subunit 1022 is invoked to extract entity recognition results from the first key field data by calling the large language model based on preset semantic analysis prompts; wherein, the entity recognition results include at least key events, key figures, time and place, sentiment tendencies and points of contention; The statistical subunit 1023 is used to perform word frequency statistics on the list of popular comments from netizens and obtain the word frequency statistics results. The first extraction subunit 1021 is also used to extract multiple hot topics with high discussion from the list of netizens' hot comments based on word frequency statistics results. The first extraction subunit 1021 is also used to extract core figure information based on entity recognition results and multiple highly discussed hot topics; wherein, the core figure information includes core figure pairs, the relationship type between core figure pairs, the relationship features between core figure pairs, and the feature information of each core figure in the core figure pair. The second generation subunit 1024 is used to generate a summary of hot events based on multiple highly discussed hot topics, entity recognition results, and key figure information.

[0282] In some embodiments, the first extraction subunit 1021 is specifically used to extract discussion topics from the list of popular comments by netizens based on word frequency statistics results, and obtain multiple initial candidate topics; Based on the data from the first key field, multiple interaction metrics are calculated that correspond one-to-one with multiple initial candidate topics. Among them, the interaction metrics include at least one of the following: overall popularity score, cross-platform coverage, topic growth rate, and sentiment polarization index. Multiple initial candidate topics were filtered based on several interaction metrics, resulting in several trending topics with high discussion levels.

[0283] In some embodiments, the first extraction subunit 1021 is further configured to obtain multiple interaction data corresponding one-to-one with multiple initial candidate topics based on the list of popular online comments; wherein the interaction data is used to reflect the degree of user participation in the multiple initial candidate topics; and / or Calculate multiple comprehensive popularity scores based on multiple interaction data points, each corresponding to one of the multiple initial candidate topics; and / or Each popular comment in the list of popular comments is categorized by sentiment, resulting in a list of sentiment types corresponding to the list of popular comments. The sentiment type list includes the sentiment type corresponding to each popular comment in the list, which can be one of the following: positive, neutral, or negative; and / or Based on the list of popular online comments and the list of sentiment types, calculate multiple sentiment polarization indices that correspond one-to-one with the multiple initial candidate topics; and / or Calculate cross-platform coverage based on data source identifiers, corresponding one-to-one with multiple initial candidate topics; and / or Calculate the growth rate of multiple topics that correspond one-to-one with multiple initial candidate topics based on the search results list; The interaction metrics include at least one of the following: overall popularity score, cross-platform coverage, topic growth rate, and emotional polarization index.

[0284] In some embodiments, the parsing unit 1030 includes: The second extraction subunit 1031 is used to extract key information from the summary of hot events; among which, the key information includes the event name, key figures, controversial topics, emotional resonance points, and high-frequency topics in netizens' comments; The second construction subunit 1032 is used to construct attention units to mine prompt word templates based on key information in the summary; The third generation subunit 1033 is used to generate attention unit mining prompt text based on the key information of the summary and the attention unit mining prompt word template; The third generation subunit 1033 is also used to mine cue word text through a large language model and attention units to generate multiple initial attention units; among them, the fields in the initial attention units include attention unit type, core comparison points, discussion degree, list of controversial topics, brief description, key figures, and comparison figures; Evaluation subunit 1034 is used to perform quality evaluation on multiple initial attention units respectively, and obtain multiple evaluation results corresponding to the multiple initial attention units one by one; among them, the evaluation results include discussion level, completeness evaluation result, and comparison point clarity evaluation result; The preprocessing subunit 1035 is used to preprocess multiple initial attention units based on multiple evaluation results to obtain multiple target attention units.

[0285] In some embodiments, the second construction subunit 1032 is specifically used to determine the hot event type and the priority attention unit type based on the summary key information; Based on the priority attention unit type and hot event type, construct a structured attention unit mining prompt word template; Among them, trending events include either popular TV dramas or real-life events. Priority attention unit types include one or more of the following: image similarity type, multi-dimensional contrast type, and emotional narrative type; The attention unit mining prompt template includes instructions, a summary of hot events, an attention unit type definition, an attention unit type selection strategy, attention unit quality requirements, special scenario handling rules, output format requirements, and requirements for enhancing online appeal and attractiveness.

[0286] In some embodiments, the preprocessing subunit 1035 is specifically used to filter multiple initial attention units based on multiple evaluation results to obtain multiple candidate attention units; Determine whether the total number of multiple candidate attention units is less than a preset attention unit number threshold; When the total number of multiple candidate attention units is not less than the preset attention unit number threshold, the multiple candidate attention units are deduplicated to obtain multiple attention units to be verified. Perform diversity verification on multiple attention units to be verified, and obtain multiple diversity verification results that correspond one-to-one with the multiple attention units to be verified. Based on multiple diversity verification results, attention units that pass the verification are selected from multiple attention units to be verified, resulting in multiple attention units with high discussion. The attention units that have already been used for material generation from multiple high-discussion attention units are filtered to obtain multiple target attention units.

[0287] In some embodiments, the generation unit 1040 includes: The second acquisition subunit 1041 is used to acquire information of multiple attention units that correspond one-to-one with multiple target attention units; The fourth generation subunit 1042 is used to generate multiple material generation prompts corresponding to multiple target attention units based on the hot event summary, preset business scenarios, preset dimension templates and multiple attention unit information; The third construction subunit 1043 is used to generate prompt words based on multiple materials and construct multiple material generation tasks that correspond one-to-one with multiple target attention units; Execution subunit 1044 is used to execute multiple material generation tasks in parallel to obtain multiple initial material data corresponding to multiple target attention units; The verification subunit 1045 is used to verify multiple initial material data respectively, and obtain multiple material verification results corresponding one-to-one with the multiple initial material data; among them, the material verification results include format verification results and quality verification results; Select subunit 1046 is used to select all the initial material data that have passed the verification from multiple initial material data based on the verification results of multiple material data, so as to obtain a candidate material data set; The deduplication subunit 1047 is used to perform deduplication on the candidate material data set to obtain hot material data.

[0288] In some embodiments, the AI-based hotspot material generation device further includes: Extraction unit 1050 is used to extract the second key field from the hot material data; wherein, the second key field includes at least one or more of the following: title, left-side character data object, right-side character data object, and brand embedded content; both the left-side character data object and the right-side character data object contain a field order array and a red-marked array; Building unit 1060 is used to construct a DOM structure for rendering a visual comparison chart based on the data object in the second key field; The receiving unit 1070 is used to receive editing instructions input by the user for the DOM structure; The editing unit 1080 is used to edit the DOM structure according to editing instructions to obtain the target DOM structure; wherein, the editing instructions include at least one of text style editing instructions, image replacement instructions, and layout adjustment instructions; Rendering unit 1090 is used to render based on the target DOM structure and output a visual comparison chart.

[0289] like Figure 8 As shown, this application provides an electronic device 1100, which includes a processor 1101 and a memory 1102. The processor 1101 and the memory 1102 are interconnected and communicate with each other through a communication bus 1103 and / or other forms of connection mechanism (not shown). The memory 1102 stores a computer program that can be executed by the processor 1101. When the computing device is running, the processor 1101 executes the computer program to perform the method in any of the aforementioned optional implementations.

[0290] This application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the method in any of the aforementioned optional implementations.

[0291] The computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0292] This application provides a computer program product, which includes a computer program that, when run by a processor, executes the method in any of the aforementioned optional implementations.

[0293] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and not to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and they should all be covered within the scope of the claims and specification of this application. In particular, as long as there is no conflict, the various technical features mentioned in the embodiments can be combined in any way. This application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for generating trending topics based on AI, characterized in that, include: The search data is obtained by performing parallel searches on multiple data sources based on trending keywords. Hotspot mining is performed on the retrieved data using a large language model to obtain a summary of hot events; The hot topic event summary and the retrieved data are parsed using a large language model to obtain multiple target attention units for generating materials; Based on the hot topic event summary, the preset business scenario, and the multiple target attention units, hot topic material data is generated.

2. The AI-based hotspot material generation method according to claim 1, characterized in that, The process of performing parallel searches across multiple data sources based on trending keywords to obtain search data includes: Generate cache keys based on trending keywords; Check whether there is matching non-expired cached data in the memory cache and local storage based on the cache key; When a matching non-expired cached data is found in the memory cache and local storage based on the cache key, the non-expired cached data matching the cache key is retrieved as the search data; When no matching non-expired cached data exists, multiple data collection tasks are constructed based on the hot keywords, each corresponding to one of the multiple data sources. Based on the multiple data acquisition tasks, data is acquired from the multiple data sources in parallel to obtain multiple data acquisition results that correspond one-to-one with the multiple data sources; wherein, the multiple data acquisition results constitute the retrieval data.

3. The AI-based hotspot material generation method according to claim 1, characterized in that, The process of performing hotspot mining on the retrieved data using a large language model to obtain hotspot event summaries includes: Extract the first key field data from the search data; wherein, the first key field data includes the search results list, the list of popular comments from netizens, and the data source identifier; Based on preset semantic analysis prompts, a large language model is invoked to extract entity recognition results from the first key field data; wherein, the entity recognition results include at least key events, key figures, time and place, sentiment tendencies, and points of contention; The word frequency statistics were performed on the list of popular comments from netizens, and the word frequency statistics results were obtained. Based on the word frequency statistics, several highly discussed and trending topics were extracted from the list of popular online comments. Based on the entity recognition results and the multiple highly discussed hot topics, core figure information is extracted; wherein, the core figure information includes core figure pairs, the relationship type between the core figure pairs, the relationship features between the core figure pairs, and the feature information of each core figure in the core figure pair; Based on the multiple highly discussed trending topics, the entity recognition results, and the key figures' information, a summary of trending events is generated.

4. The AI-based hotspot material generation method according to claim 3, characterized in that, The process involves extracting multiple highly discussed trending topics from the list of popular online comments based on the word frequency statistics results, including: Based on the word frequency statistics, discussion topics in the list of popular online comments are extracted to obtain multiple initial candidate topics; Based on the first key field data, calculate multiple interaction indicators that correspond one-to-one with the multiple initial candidate topics; wherein, the interaction indicators include at least one of the following: comprehensive popularity score, cross-platform coverage, topic growth rate, and emotional polarization index; Based on the multiple interaction metrics, the multiple initial candidate topics were filtered to obtain multiple hot topics with high discussion levels.

5. The AI-based hotspot material generation method according to claim 4, characterized in that, The calculation of multiple interaction metrics corresponding one-to-one with the multiple initial candidate topics based on the first key field data includes: Based on the list of popular user comments, obtain multiple interaction data points corresponding one-to-one with the multiple initial candidate topics; wherein, the interaction data is used to reflect the degree of user participation in the multiple initial candidate topics; and / or Calculate multiple comprehensive popularity scores corresponding one-to-one with the multiple initial candidate topics based on the multiple interaction data; and / or Each popular comment in the list of popular comments is categorized by sentiment to obtain a list of sentiment types corresponding to the list of popular comments; wherein, the list of sentiment types includes the sentiment type corresponding to each popular comment in the list of popular comments, and the sentiment type is one of positive, neutral, and negative; and / or Based on the list of popular online comments and the list of sentiment types, calculate multiple sentiment polarization indices that correspond one-to-one with the multiple initial candidate topics; and / or Calculate multiple cross-platform coverages corresponding one-to-one with the multiple initial candidate topics based on the data source identifier; and / or The growth rate of multiple topics corresponding one-to-one with the multiple initial candidate topics is calculated based on the search results list.

6. The AI-based hotspot material generation method according to claim 1, characterized in that, The process involves parsing the hot topic event summary and the retrieved data using a large language model to obtain multiple target attention units for generating content, including: Extract key information from the summary of the hot events; wherein, the key information includes the event name, key figures, controversial topics, points of emotional resonance, and frequently mentioned topics in netizens' comments; Based on the key information in the summary, an attention unit is constructed to mine cue word templates; Based on the key information in the summary and the attention unit mining prompt word template, generate attention unit mining prompt word text; By mining cue word text using a large language model and the attention units, multiple initial attention units are generated; wherein, the fields in the initial attention units include attention unit type, core comparison points, discussion level, list of controversial topics, brief description, key figures, and comparison figures; Each of the multiple initial attention units is subjected to a quality assessment, resulting in multiple assessment results corresponding to each of the multiple initial attention units; wherein, the assessment results include discussion level, completeness assessment result, and comparison point clarity assessment result; Based on the multiple evaluation results, the multiple initial attention units are preprocessed to obtain multiple target attention units.

7. The AI-based hotspot material generation method according to claim 6, characterized in that, The step of constructing an attention unit to mine prompt word templates based on the key information in the summary includes: Based on the key information in the summary, determine the types of hot events and the types of priority attention units; Based on the priority attention unit type and the hot event type, construct a structured attention unit mining prompt word template; The types of trending events include popular TV dramas or real-life events involving people. The priority attention unit types include one or more of the following: image similarity type, multi-dimensional comparison type, and emotional narrative type; The attention unit mining prompt template includes instructions, a summary of hot events, an attention unit type definition, an attention unit type selection strategy, attention unit quality requirements, special scenario processing rules, output format requirements, and requirements for enhancing online appeal and attractiveness.

8. The AI-based hotspot material generation method according to claim 6 or 7, characterized in that, The step of preprocessing the multiple initial attention units based on the multiple evaluation results to obtain multiple target attention units includes: Based on the multiple evaluation results, the multiple initial attention units are screened to obtain multiple candidate attention units; Determine whether the total number of the plurality of candidate attention units is less than a preset attention unit number threshold; When the total number of the multiple candidate attention units is not less than the preset attention unit number threshold, the multiple candidate attention units are deduplicated to obtain multiple attention units to be verified. Perform diversity verification on the plurality of attention units to be verified respectively to obtain a plurality of diversity verification results corresponding one-to-one with the plurality of attention units to be verified; Based on the multiple diversity verification results, the attention units that pass the verification are selected from the multiple attention units to be verified, resulting in multiple attention units with high discussion. The attention units that have already been used for material generation among the multiple high-discussion attention units are filtered to obtain multiple target attention units.

9. The AI-based hotspot material generation method according to claim 1, characterized in that, The step of generating hot topic material data based on the hot topic event summary, the preset business scenario, and the multiple target attention units includes: Obtain information about multiple attention units that correspond one-to-one with the multiple target attention units; Based on the hot topic event summary, preset business scenarios, preset dimension templates, and the information of the multiple attention units, generate multiple material generation prompts that correspond one-to-one with the multiple target attention units; Based on the multiple materials, prompt words are generated, and multiple material generation tasks are constructed that correspond one-to-one with the multiple target attention units; The multiple material generation tasks are executed in parallel to obtain multiple initial material data corresponding one-to-one with the multiple target attention units; The multiple initial material data are verified one by one to obtain multiple material verification results corresponding to the multiple initial material data; wherein, the material verification results include format verification results and quality verification results; Based on the multiple material verification results, select all the initial material data that passed the verification from the multiple initial material data to obtain a candidate material data set; The candidate material data set is deduplicated to obtain hotspot material data.

10. The AI-based hotspot material generation method according to claim 1, characterized in that, The method further includes: Extract the second key field from the hot material data; wherein, the second key field includes at least one or more of the following: title, left-side character data object, right-side character data object, and brand embedded content; both the left-side character data object and the right-side character data object contain a field order array and a red-marked array; Based on the data object in the second key field, construct a DOM structure for rendering the visual comparison chart; Receive editing instructions from the user regarding the DOM structure; The DOM structure is edited according to the editing instructions to obtain the target DOM structure; wherein, the editing instructions include at least one of text style editing instructions, image replacement instructions, and layout adjustment instructions; Render the target DOM structure and output a visual comparison chart.

11. A device for generating trending topics based on AI, characterized in that, The AI-based hotspot material generation device includes: The retrieval unit is used to perform parallel searches across multiple data sources based on hot keywords to obtain retrieval data. The hotspot mining unit is used to perform hotspot mining on the retrieved data using a large language model to obtain a summary of hot events. The parsing unit is used to parse the hot event summary and the retrieved data through a large language model to obtain multiple target attention units for generating materials; The generation unit is used to generate hot topic material data based on the hot topic event summary, the preset business scenario, and the multiple target attention units.

12. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the AI-based hotspot material generation method according to any one of claims 1 to 10.

13. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which, when executed by a processor, performs the AI-based hotspot material generation method according to any one of claims 1 to 10.

14. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, performs the AI-based hotspot material generation method according to any one of claims 1 to 10.

Images