Social media account screening method and system based on vectorized semantic matching

Abstract

The application provides a social media account screening method and system based on vectorized semantic matching. The method includes: performing intent analysis on natural language screening requirements to generate brand requirement fields, structured screening conditions and negative exclusion conditions; performing vectorization processing on the brand requirement fields to generate a brand requirement vector set, and mapping the structured screening conditions and negative exclusion conditions into screening parameters executable by the account platform; searching for candidate accounts in a local account asset library based on the screening parameters; analyzing content, comments, scenes, interactions and business information in the account data to generate an account portrait vector set and business performance characteristics; calculating multi-dimensional semantic matching results and search word acceptance degrees of the candidate accounts, and generating an account comprehensive recommendation score; outputting an account recommendation result according to the account comprehensive recommendation score, and generating recommendation explanation information. The application can improve demand mapping accuracy, improve semantic screening and search acceptance evaluation accuracy, and enhance recommendation explainability.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method and system for filtering social media accounts based on vectorized semantic matching. Background Technology

[0002] With the development of content-based social media platforms, brands typically need to sift through a large number of influencer accounts, consumer sharing accounts, sales guide accounts, and previously partnered accounts to select suitable accounts for their marketing goals when conducting marketing communications, product seeding, search engine optimization, competitor comparisons, and new product launches. Current account selection methods mainly rely on human experience, basic data filtering, keyword search, tag-based recommendations, or external platform criteria, typically using follower count, engagement, pricing, account category, and content tags as the primary judgment criteria.

[0003] However, in refined advertising scenarios, brands need to assess not only an account's surface-level influence but also whether it consistently addresses specific consumer pain points, possesses search term capacity, is suitable for conveying brand selling points, can reach the target decision-making audience, and whether there are issues such as over-commercialization, competitor conflicts, or brand safety risks. Existing methods rely excessively on human experience, have unstable screening criteria, and struggle to establish reusable and explainable screening mechanisms. Screening methods based on follower count, interaction volume, and tags fail to determine the true semantic fit between account content and brand needs. Keyword matching methods struggle to identify semantically similar but differently expressed content, easily leading to missed or incorrect selections. Furthermore, existing systems lack accurate mappings from natural language requirements to platform screening fields, easily mistaking negative requirements for positive recall conditions. They also struggle to uniformly utilize local account assets, external database expansion results, and historical feedback data for ranking optimization, resulting in a lack of explanatory basis and process auditing capabilities for account recommendation results. Summary of the Invention

[0004] In view of this, embodiments of this application provide a social media account screening method and system based on vectorized semantic matching to solve the problems of inaccurate demand mapping, inaccurate semantic and search connection judgment, and lack of explanation in recommendations in the prior art.

[0005] A first aspect of this application provides a method for screening social media accounts based on vectorized semantic matching, comprising: acquiring product information, user identity information, and natural language screening requirements corresponding to brand marketing tasks; performing intent parsing on the natural language screening requirements to generate brand requirement fields, structured screening conditions, and negative exclusion conditions; vectorizing the brand requirement fields to generate a set of brand requirement vectors, and mapping the structured screening conditions and negative exclusion conditions to screening parameters executable by the account platform; retrieving candidate accounts from a local account asset library based on the screening parameters; when the search results do not meet the preset candidate conditions, calling an external account platform to expand the candidate accounts, and storing and vectorizing the expanded accounts; acquiring account data for each candidate account, parsing the content, comments, scenarios, interactions, and commercial information in the account data to generate a set of account profile vectors and commercial performance features; calculating the multidimensional semantic matching results and search term relevance of the candidate accounts based on the set of brand requirement vectors and the set of account profile vectors, and generating a comprehensive account recommendation score by combining commercial performance features, screening parameter satisfaction status, historical preference information, and negative exclusion matching results; outputting account recommendation results based on the comprehensive account recommendation score, and generating recommendation explanation information.

[0006] A second aspect of this application provides a social media account screening system based on vectorized semantic matching, comprising: an acquisition module, configured to acquire product information, user identity information, and natural language screening requirements corresponding to brand marketing tasks, perform intent parsing on the natural language screening requirements, and generate brand requirement fields, structured screening conditions, and negative exclusion conditions; a mapping module, configured to perform vectorization processing on the brand requirement fields, generate a set of brand requirement vectors, and map the structured screening conditions and negative exclusion conditions to screening parameters executable by the account platform; and a retrieval module, configured to first retrieve candidate accounts in a local account asset library based on the screening parameters, and when the retrieval results do not meet the preset candidate conditions... The system utilizes an external account platform to expand candidate accounts, and then stores and vectorizes the expanded accounts. A parsing module acquires account data from each candidate account, analyzes the content, comments, scenarios, interactions, and commercial information within the data, and generates an account profile vector set and commercial performance features. A calculation module calculates the multi-dimensional semantic matching results and search term relevance of candidate accounts based on the brand demand vector set and the account profile vector set, and generates a comprehensive account recommendation score by combining commercial performance features, the status of screening parameter satisfaction, historical preference information, and negative exclusion matching results. A generation module outputs account recommendation results based on the comprehensive recommendation score and generates recommendation explanation information.

[0007] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: By acquiring product information, user identity information, and natural language filtering requirements corresponding to brand marketing tasks, the intent of the natural language filtering requirements is parsed to generate brand requirement fields, structured filtering conditions, and negative exclusion conditions. The brand requirement fields are vectorized to generate a set of brand requirement vectors, and the structured filtering conditions and negative exclusion conditions are mapped to filtering parameters executable by the account platform. Based on the filtering parameters, candidate accounts are first retrieved from the local account asset library. If the search results do not meet the preset candidate conditions, an external account platform is called to expand the candidate accounts, and the expanded accounts are added to the library and vectorized. Account data of each candidate account is acquired, and the content, comments, scenarios, interactions, and commercial information in the account data are parsed to generate a set of account profile vectors and commercial performance characteristics. Based on the set of brand requirement vectors and the set of account profile vectors, the multidimensional semantic matching results and search term relevance of the candidate accounts are calculated, and a comprehensive account recommendation score is generated by combining commercial performance characteristics, filtering parameter satisfaction status, historical preference information, and negative exclusion matching results. The account recommendation results are output based on the comprehensive account recommendation score, and recommendation explanation information is generated. This application can improve the accuracy of demand mapping, enhance the precision of semantic filtering and search acceptance evaluation, and improve the interpretability of recommendations. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating the social media account filtering method based on vectorized semantic matching provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the social media account filtering system based on vectorized semantic matching provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0010] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0011] Current methods for selecting social media accounts typically rely on human experience, basic data filtering, keyword retrieval, tag-based recommendations, or external account platform criteria. These methods often use factors such as follower count, engagement, pricing, account category, and content tags as selection criteria, which can achieve coarse-grained account engagement, but struggle to accurately identify the deep semantic fit between account content and brand selling points, target audience, consumption scenarios, and search terms.

[0012] In refined marketing campaigns, brands typically input filtering requirements that include natural language expressions of campaign objectives, budget constraints, content direction, search term relevance requirements, and negative exclusion criteria. Existing technologies struggle to accurately map these natural language requirements into platform-executable filtering parameters, leading to issues such as missing filtering fields, misrepresentation of negative requirements, and under-selection of semantically similar content. Furthermore, current account recommendation results often lack assessments of search term relevance, comprehensive evaluation of commercial suitability, and explanatory information, resulting in insufficient accuracy in account filtering and reliability in campaign decisions.

[0013] To address the aforementioned issues, this application provides a social media account screening method based on vectorized semantic matching. By acquiring product information, user identity information, and natural language screening requirements corresponding to brand marketing tasks, the method performs intent parsing on the natural language screening requirements to generate brand requirement fields, structured screening conditions, and negative exclusion conditions. The brand requirement fields are then vectorized to generate a set of brand requirement vectors, and the structured screening conditions and negative exclusion conditions are mapped to screening parameters executable by the account platform. Based on these screening parameters, candidate accounts are first retrieved from the local account asset library. If the local retrieval is insufficient, an external account platform is invoked to expand the candidate accounts, and the expanded accounts are then added to the library and vectorized.

[0014] Furthermore, this application analyzes the account data of candidate accounts to generate an account profile vector set and business performance features; calculates multi-dimensional semantic matching results and search term relevance based on the brand demand vector set and the account profile vector set, and generates a comprehensive account recommendation score by combining business performance features, screening parameter satisfaction status, historical preference information, and negative exclusion matching results; finally, outputs the account recommendation results based on the comprehensive account recommendation score and generates recommendation explanation information.

[0015] Through the above technical solutions, this application can improve the mapping accuracy of natural language filtering requirements to platform filtering parameters and reduce the probability of false recall of negative requirements; it can improve the accuracy of account semantic filtering and search term acceptance evaluation based on the vectorized representation of account content, comment requirements and consumption scenarios; it can also generate interpretable account recommendation results by combining commercial suitability and risk status, thereby improving the efficiency of social media account filtering, recommendation accuracy and the reliability of ad placement decisions.

[0016] The technical solution of this application will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0017] Figure 1 This is a flowchart illustrating the social media account filtering method based on vectorized semantic matching provided in an embodiment of this application. Figure 1 As shown, the method may specifically include: S101: Obtain product information, user identity information, and natural language filtering requirements corresponding to the brand marketing task; perform intent parsing on the natural language filtering requirements; and generate brand requirement fields, structured filtering conditions, and negative exclusion conditions. S102, the brand requirement field is vectorized to generate a set of brand requirement vectors, and the structured filtering conditions and negative exclusion conditions are mapped to filtering parameters that can be executed by the account platform. S103: Based on the filtering parameters, first search for candidate accounts in the local account asset library. If the search results do not meet the preset candidate conditions, call the external account platform to expand the candidate accounts, and then put the expanded accounts into the library and vectorize them. S104: Obtain account data for each candidate account, analyze the content, comments, scenarios, interactions, and business information in the account data, and generate a set of account profile vectors and business performance characteristics; S105: Based on the brand demand vector set and the account profile vector set, calculate the multidimensional semantic matching results and search term relevance of candidate accounts, and generate a comprehensive account recommendation score by combining business performance characteristics, screening parameter satisfaction status, historical preference information and negative exclusion matching results. S106: Output the account recommendation results based on the account's overall recommendation score, and generate recommendation explanation information.

[0018] In some embodiments, the intent of the natural language filtering request is parsed to generate a brand requirement field, structured filtering conditions, and negative exclusion conditions, including: Based on the task context of product information, user identity information, and brand marketing tasks, semantic segmentation and intent recognition are performed on natural language filtering requirements to determine semantic requirement segments, numerical constraint segments, and negative expression segments related to the target audience. Classify semantic requirement fragments into brand communication elements to generate brand requirement fields; Perform unit conversion, interval completion, and field normalization on numerical constraint fragments to generate structured filtering conditions; Based on the negative trigger words and their corresponding scope of application, negative expression fragments are analyzed to generate negative exclusion conditions for account filtering or point deduction.

[0019] Specifically, in the infant formula brand account selection scenario, the system receives product information, user identity information, and natural language filtering requirements input by brand operators through user terminals. Product information includes the infant formula product's category attributes, core selling points, target audience, and compliant expression boundaries; user identity information includes the current user's brand affiliation, historical advertising preferences, previously selected account types, and account exclusion records; the brand marketing task context includes advertising objectives such as new product launch promotion, search term acquisition, competitor interception, and scenario-based education. User-inputted requirements could include promoting fresh, nutritious, easily absorbed, and digestively stable infant formula, targeting mothers of babies aged 0-3 years, focusing on accounts related to formula transitions, seasonal feeding changes, infant gastrointestinal sensitivity, and abnormal bowel movements, with 10,000-50,000 followers, a price not exceeding 3500 RMB, and both video and text content are acceptable, but food review accounts are excluded.

[0020] The system first inputs natural language filtering requirements into the intent parsing module, and then performs semantic segmentation by combining product information, user identity information, and task context. During semantic segmentation, the system identifies semantic requirement fragments related to brand communication based on the campaign objective. For example, "fresh and nutritious," "easily absorbed," and "more stable digestive system" are identified as product selling point requirements; "mothers of babies aged 0-3 years," "new mothers," and "mothers concerned about digestion and absorption" are identified as target audience requirements; and "switching formulas," "seasonal feeding," "baby's sensitive stomach," and "abnormal bowel movements" are identified as consumption scenario requirements. If the task context indicates that the current task is related to search terms, the system will also identify expressions such as "recommendation of formula for switching formulas," "formula for babies with sensitive stomachs," and "formula is easily absorbed" as search requirement fragments. If the task context indicates that the current task is competitor interception, the system will identify expressions such as competitor names, competitor selling points, competitor alternatives, and competitor comparisons as competitor requirement fragments.

[0021] During intent recognition, the system simultaneously identifies numerical constraint segments. For follower counts of 10,000-50,000, the system converts Chinese quantity units into unified numerical units, generating a lower limit of 10,000 and an upper limit of 50,000 followers. For prices not exceeding 3,500, the system, based on the content format constraint of both video and image / text, synchronizes the price upper limit to both image / text and video price upper limits, avoiding missing search conditions on external account platforms due to configuring only a single price field. If the user only enters a follower count of less than 50,000, the system completes the lower limit by leaving it empty or using the platform's default lower limit according to field rules. If the user enters 2,000 to 10,000, the system uniformly converts mixed units such as k and w into a platform-executable numerical range. For information such as region, gender, account category, content format, and desired number, the system converts it into unified field codes according to a field mapping table, forming structured filtering conditions.

[0022] For semantic demand fragments, the system categorizes them according to brand communication elements, generating brand demand fields. These fields include at least selling point fields, target audience fields, scenario fields, search fields, competitor fields, and risk fields. Taking a milk powder brand task as an example, the selling point field records communication directions such as fresh nutrition, easy absorption, and stable digestion; the target audience field records mothers of babies aged 0-3 years and related decision-making groups; the scenario field records consumption scenarios such as switching formulas, seasonal changes, digestive sensitivities, and abnormal bowel movements; the search field records high-intent search terms; and the risk field records compliance risk expressions such as exaggerated efficacy claims, medical-like expressions, and absolute promises. Through this categorization method, the system converts mixed expressions in natural language into brand demand fields that can be directly used in subsequent vectorized matching.

[0023] For negative expression fragments, the system analyzes them based on negative trigger words and their scope. The system identifies negative trigger words such as "don't," "exclude," "don't want," "avoid," and "don't recommend again," and determines fragments following these trigger words that are related to account type, content tags, platform category, reported brand, or content direction as negative exclusion targets. Taking the above input as an example, if "don't recommend food review accounts" is analyzed as a negative expression fragment, the system determines "food review" as a negative tag and expands it into negative exclusion conditions by combining category synonyms and historical exclusion records. If a candidate account's account tags, recent content tags, content summary, or comment keywords strongly match the negative exclusion condition, the system directly filters it in subsequent screenings; if it only weakly matches in a small amount of historical content, the system retains the account but configures a negative deduction value for comprehensive recommendation scoring.

[0024] Through the above embodiments, the system can decompose the natural language requirements of the brand into semantic requirement fragments, numerical constraint fragments, and negative expression fragments, and generate brand requirement fields, structured filtering conditions, and negative exclusion conditions respectively. This not only improves the accuracy of the mapping from natural language to platform filtering parameters, but also reduces the probability of negative requirements being mistakenly used as positive recall conditions, providing stable input for subsequent account vector matching, dynamic database expansion, and comprehensive recommendation ranking.

[0025] In some embodiments, the brand requirement field is vectorized to generate a brand requirement vector set, and the structured filtering conditions and negative exclusion conditions are mapped to filtering parameters that can be executed by the account platform, including: The brand demand fields are semantically categorized and standardized according to the brand communication dimensions. A vector generation model is called to generate demand sub-vectors corresponding to each brand communication dimension. Matching weights are configured for the demand sub-vectors according to the brand marketing task's delivery objectives to obtain a set of brand demand vectors. Based on the account platform's field mapping rules, the structured filtering conditions are transformed, their values ​​are normalized, and constraints are merged to generate positive filtering parameters. Exclusion filtering parameters are generated based on negative exclusion conditions, and the positive filtering parameters and exclusion filtering parameters are subjected to consistency verification and conflict resolution to obtain the filtering parameters that can be executed by the account platform.

[0026] Specifically, in the scenario of new product launch and search for infant formula brands, after completing natural language filtering and demand analysis, the system obtains brand demand fields, structured filtering conditions, and negative exclusion conditions. Brand demand fields include selling point fields, target audience fields, scenario fields, search fields, competitor fields, and risk fields. The system first semantically categorizes brand demand fields according to brand communication dimensions. Fresh nutrition, easy absorption, and more stable digestive system are categorized into the selling point dimension; mothers of babies aged 0-3 years, new mothers, and mothers concerned about their babies' digestion and absorption are categorized into the target audience dimension; switching formulas, seasonal feeding, babies with sensitive stomachs, and abnormal bowel movements are categorized into the scenario dimension; formula recommendations for switching formulas, formulas for babies with sensitive stomachs, formulas that are easily absorbed, and formula reviews are categorized into the search dimension; competitor brand names, competitor alternatives, and competitor comparisons are categorized into the competitor dimension; and exaggerated efficacy claims, medical-like expressions, and absolute promises are categorized into the risk dimension.

[0027] Furthermore, after semantic classification, the system standardizes the fields for each brand communication dimension. For example, expressions like "baby's stomach upset," "stomach sensitivity," and "poor bowel movements after drinking milk" are grouped under the "stomach sensitivity" scenario; "changing formula," "changing stage," and "incompatibility with switching formulas" are grouped under the "switching formula" scenario; and "easy to digest," "good absorption," and "no bloating" are grouped under the "easy absorption" selling point. During field standardization, the system combines the brand's SPU knowledge base, historical ad terminology database, and platform search terminology database to generate standard field identifiers, preventing the same requirement from being split into multiple unrelated fields due to differences in expression. Subsequently, the system calls the vector generation model to encode each standard field and its contextual description, generating sub-vectors for selling point requirements, audience requirements, scenario requirements, search requirements, competitor requirements, and risk requirements. These sub-vectors are then stored in association according to field identifiers, source text, and task identifiers, forming a brand requirement vector set.

[0028] When configuring matching weights, the system dynamically determines the importance of each demand sub-vector based on the brand's marketing campaign objectives. If the current task is new product launch awareness, the system increases the matching weight of the selling point demand sub-vector and the scenario demand sub-vector, prioritizing accounts that consistently cover topics such as formula transition, gastrointestinal sensitivity, and formula reviews. If the current task is search engagement, the system increases the matching weight of the search demand sub-vector and the audience demand sub-vector, giving higher matching priority to accounts with comments addressing purchase inquiries, brand comparisons, and search decision-making issues. If the current task is competitor interception, the system increases the weight of the competitor demand sub-vector and uses the competitor's historical content coverage as a crucial basis for subsequent matching calculations. If the product category belongs to high-risk industries such as mother and baby products, the system increases the constraint weight of the risk demand sub-vector in the risk identification phase.

[0029] For structured filtering criteria, the system performs field transformation, value normalization, and constraint merging according to the account platform's field mapping rules. For example, a follower count of 10,000-50,000 is converted into a lower limit of 10,000 followers and a higher limit of 50,000 followers; when the price does not exceed 3,500 and both video and image / text are acceptable, the system synchronously writes the price constraint into the upper limit of the image / text price and the upper limit of the video price, avoiding generating only a single price field; conditions such as region, gender, account category, content format, and expected quantity are converted into standard enumerated values ​​based on the platform's field dictionary. If the user modifies the follower count to 20,000-100,000 in multiple rounds of dialogue, the system updates the follower count range based on the fields in this round, while retaining the unmodified price, content direction, and risk constraints, generating positive filtering parameters.

[0030] For negative exclusion conditions, the system generates exclusion filtering parameters based on the negative expression fragment. For example, when a user inputs "don't want food review accounts," the system categorizes food reviews as an exclusion tag and further associates it with account category, recent note tags, content summary, and comment keywords. If the excluded object can be directly identified by the account platform, the system writes it into the platform's exclusion field; if the platform does not support direct exclusion, the system writes it into local filtering rules and rating deduction rules. The system then performs consistency checks on the positive filtering parameters and exclusion filtering parameters to determine if there is a conflict where the same tag is used as both a positive recall condition and an exclusion condition. For example, if "Sam's Club food" is strongly related to positive semantic needs, and "don't want food reviews" is a negative exclusion condition, the system distinguishes between retail scenario-related and review content type exclusions. Instead of directly deleting the Sam's Club food requirement, it excludes accounts primarily focused on reviews during the account filtering stage, while retaining accounts that can handle Sam's Club product reviews, family stockpiling, and food selection content.

[0031] Through the above embodiments, the system can transform brand demand fields into a set of brand demand vectors organized by communication dimensions, and convert structured filtering conditions and negative exclusion conditions into filtering parameters that can be executed by the account platform. At the same time, consistency verification and conflict resolution avoid mutual interference between positive recall and negative exclusion, thereby improving the accuracy of platform retrieval parameter generation and the stability of subsequent account semantic matching.

[0032] In some embodiments, candidate accounts are first retrieved from the local account asset library based on filtering parameters. If the retrieval results do not meet the preset candidate conditions, an external account platform is invoked to expand the candidate accounts. The expanded accounts are then stored in the library and vectorized, including: Based on the screening parameters, structured conditional matching and vector similarity recall are performed in the local account asset database to obtain a local candidate account set; Based on the number of accounts, matching quality, and coverage of the local candidate account set, determine whether the search results meet the preset candidate conditions; When the search results do not meet the preset candidate conditions, a pagination expansion request is generated based on the filtering parameters, and an external account platform is called to obtain extended account data. The extended account data is deduplicated, standardized, and validated. The validated extended accounts are written to the local account asset library, and corresponding account vectors are generated to update the candidate accounts.

[0033] Specifically, when milk powder brands are selecting accounts to promote new product launches and screen potential clients, the system first retrieves data from its local account asset database after generating executable filtering parameters for the account platform. This local database stores historically partnered accounts, imported accounts, accounts retrieved from past tasks, and account profiles that have already undergone content parsing and vectorization. Filtering parameters include a fan count of 10,000 to 50,000, a maximum image / text price of 3,500, a maximum video price of 3,500, a focus on mother-and-baby related content, semantic requirements such as switching formulas, addressing infant gastrointestinal sensitivity, and ensuring good formula absorption, as well as negative exclusion criteria such as restaurant reviews.

[0034] The system first performs structured conditional matching on the local account asset library, filtering accounts based on follower count range, bid ceiling, account status, content format, platform category, and availability. Accounts with bids exceeding limits, inactive accounts, historically abnormal campaigns, or those matching strong exclusion tags are removed. For accounts that pass structured conditional matching, the system continues to use vector retrieval services to perform similarity retrieval between the brand demand vector set and the account content semantic vector, the account audience demand vector, and the account scenario vector. For example, if a historically partnered account has consistently covered baby feeding status, formula reviews, formula switching experiences, and abnormal bowel movements for the past 90 days, and the comment section contains inquiries about whether it's suitable for sensitive babies, whether switching formulas will cause diarrhea, and how to buy it, the system will retrieve this account as a local candidate account. After structured conditional matching and vector similarity retrieval, the system forms a local candidate account set and records the source, matched fields, similarity score, and exclusion verification results for each account.

[0035] The system then determines the sufficiency of local search results based on preset candidate criteria. These criteria include not only a threshold for the number of accounts, but also thresholds for matching quality and coverage. For example, if a brand wants to generate 30 recommended accounts, the system can require at least 80 local candidate accounts, with at least 40 highly matching accounts, covering candidate areas such as core product seeding, search engagement, competitor interception, and cost-effective product placement. If the number of local candidate accounts meets the requirements, but most accounts only cover general maternal and infant life and lack specific scenarios such as formula transition, gastrointestinal sensitivity, and formula reviews, the system will still determine that the coverage is insufficient. If highly matching accounts are concentrated in a single account type and cannot meet the needs of different campaign stages, the system will also determine that the search results do not meet the preset candidate criteria.

[0036] When local search results are insufficient, the system generates a paginated expansion request based on filtering parameters and calls the external account platform interface to obtain expanded account data. The paginated expansion request carries structured positive filtering parameters, exclusion filtering parameters, pagination page number, number of pages, sorting method, and task identifier. For semantic requirements not supported by the external platform, the system retains them as local secondary filtering conditions and does not force them into the external platform fields. For example, the external platform can identify follower count, price, platform category, and geographic conditions, but cannot directly identify the semantic requirement of baby formula for sensitive stomachs. The system expands the database on the external platform by mother and baby category and price conditions, and then performs local semantic matching on the newly added account content. If the number of valid new accounts after expanding the first page is still insufficient, the system continues to request the next page according to resource control rules, recording the expanded page number, total platform hits, actual number of new accounts, and interface call status.

[0037] Furthermore, after obtaining the extended account data, the system performs deduplication, field standardization, and validity checks on the externally returned accounts. During deduplication, the system identifies duplicate accounts based on platform account identifiers, account homepage links, account name similarity, and historical data mapping relationships, preventing the same account from being repeatedly added to the database due to historical imports or nickname changes. During field standardization, the system converts the number of followers, interactions, pricing, content format, account category, and cooperation status returned by external platforms into locally unified fields. During validity checks, the system removes accounts with inaccessible homepages, missing key fields, no recent valid content, abnormal pricing, suspected fake interactions, or those that meet strong exclusion criteria. Extended accounts that pass the validation are written to the local account asset database and associated with the current brand marketing task, database source, and filtering parameter version.

[0038] For newly added accounts, the system continues to capture recent notes, titles, body text, tags, comment text, image recognition text, and video recognition text within a preset time range. This data is then cleaned, subject-specific, scene-specific, and vectorized to generate corresponding semantic vectors for the account content, audience needs, and scene-specific data. If a newly added account consistently contains content related to switching formulas, feeding babies with sensitive stomachs, or formula reviews, and has purchase inquiries in the comments section, the system adds that account to the candidate list. If only the number of followers and the price are met, but the content leans towards parenting or children's education, the system retains the account's entry record but lowers its candidate priority.

[0039] Through the above embodiments, the system can prioritize the reuse of local account assets, and then dynamically expand the database from external platforms when local candidates are insufficient. The expanded accounts are deduplicated, standardized, verified, stored, and vectorized, which reduces redundant crawling and waste of interface resources, and ensures that the number of candidate accounts, matching quality, and scenario coverage meet the brand's advertising needs. This provides a sufficient, stable, and traceable candidate account foundation for subsequent comprehensive recommendation and ranking.

[0040] In some embodiments, account data for each candidate account is obtained, and the content, comments, scenarios, interactions, and commercial information in the account data are analyzed to generate an account profile vector set and commercial performance characteristics, including: Acquire account data for each candidate account within a preset time range, clean, deduplicate, and normalize the format of the account data to generate standardized account data; The content published in standardized account data is subject to topic identification and semantic vectorization. The content vectors are then weighted and aggregated based on publication time, interaction performance, and content quality to generate semantic vectors for account content. The system identifies the demand intent of comment text in standardized account data, extracts semantic information of comments related to user inquiries, decision comparisons, and consumption feedback, and generates an account audience demand vector. Based on the semantic information of published content and comments, identify the consumption scenarios covered by the account and generate an account scenario vector; An account profile vector set is generated based on the semantic vector of account content, the demand vector of account audience, and the scenario vector of account. Business performance features are then generated based on interaction quality, pricing status, and cooperation stability.

[0041] Specifically, in the infant formula brand account screening scenario, after completing local searches and external database expansion, the system retrieves account data for each candidate account within a preset time range. The preset time range can be set to the last 90 days or the last 180 days, depending on the task period. Account data includes account homepage information, historical notes, recent notes, titles, body text, platform tags, cover image recognized text, video recognized text, comment section text, posting time, readership, likes, favorites, comments, reposts, image / text pricing, video pricing, historical collaboration records, and campaign feedback records. For infant formula brand tasks, the system focuses on collecting content snippets related to maternal and infant feeding, infant formula reviews, formula switching experiences, baby's gastrointestinal status, abnormal bowel movements, and competitor comparisons, while retaining the account's recent content update frequency and collaboration density.

[0042] The system first cleanses, deduplicates, and standardizes the format of account data. For instances where the same note appears repeatedly on external platforms, in the local asset library, and in historical tasks, the system deduplicates based on note identifier, publication time, text summary, and image fingerprint. For comment text, the system removes semantically meaningless emojis, spam content, advertising content, and obviously abnormal comments. For interaction data, the system converts metrics such as likes, favorites, comments, and shares into a unified statistical standard. For pricing data, the system converts text / image pricing, video pricing, and combined pricing into a unified price field. After processing, the system generates standardized account data and retains the source, time, interaction status, and quality status for each piece of content.

[0043] Subsequently, the system performs topic identification and semantic vectorization on the content published in the standardized account data. Taking candidate account A as an example, the system identifies multiple posts published by the account in the past 90 days that involve topics such as baby's feeding status, precautions for switching formulas, formula reviews, feeding of babies with sensitive stomachs, and observation of baby's stools. The system merges the title, body text, tags, and image recognition text of each post into content semantic text, which is then input into the vector generation model to obtain content vectors. The system does not directly take a single highly interactive post as the account profile, but instead assigns aggregation weights to each content vector based on the posting time, interaction performance, and content quality. The more recent the posting time, the higher the time weight; the higher the collection rate, comment rate, and effective interaction ratio, the higher the interaction weight; and the higher the content quality weight is for posts with clear themes, sufficient genuine experiences, and clear demands in the comment section. Based on the above weights, the system performs weighted aggregation on multiple content vectors to generate an account content semantic vector, thereby reducing the interference of a single viral post or short-term hot topic on the judgment of account suitability.

[0044] For comment text, the system identifies the user's intent and extracts semantic information related to user inquiries, decision comparisons, and consumption feedback. For example, comments such as "Is this suitable for sensitive babies?", "Will switching formulas cause diarrhea?", "Can I switch to this brand if I have constipation?", "Which is better compared to this brand?", and "How to buy?" are identified as product inquiries, purchase decisions, competitor comparisons, and usage feedback intents. The system removes deduplication and merges this semantic information, and combines it with comment frequency, question completeness, and account response quality to generate comment demand text. This text is then transformed into an account audience demand vector using a vector generation model. This vector allows the system to identify whether the account's target audience has search, comparison, inquiry, and conversion intents, rather than simply identifying the content the account has published.

[0045] In the scenario identification stage, the system utilizes both published content and semantic information from comments to determine the consumption scenarios covered by the account. If the account's content consistently features topics such as switching formulas, seasonal feeding changes, gastrointestinal sensitivity, abnormal bowel movements, night feedings, weaning, and formula selection for new mothers, and related inquiries appear continuously in the comments section, the system configures these scenarios as stable coverage scenarios for the account. If only a small amount of historical content mentions these topics occasionally, they are configured as low-confidence scenarios. The system generates account scenario vectors based on the content proportion of each scenario, its recent persistence, and the intensity of comment demand, enabling the account profile to reflect specific consumption decision-making scenarios, rather than remaining within the broad category of "mother and baby account."

[0046] Finally, the system associates the semantic vector of account content, the demand vector of the account audience, and the scenario vector of the account to form an account profile vector set, and generates commercial performance characteristics based on interaction quality, pricing status, and cooperation stability. Interaction quality is determined based on the percentage of effective interactions, the structure of collections and comments, and the proportion of abnormal interactions; pricing status is determined based on the account's pricing compared to the price range of accounts at the same level, budget constraints, and historical transaction data; cooperation stability is determined based on the account's update frequency, historical cooperation density, performance record, and campaign feedback.

[0047] Through the above processing, the system can form a basic account profile that simultaneously represents content adaptation, audience intent, scenario coverage, and commercial value, providing accurate data support for subsequent multi-dimensional semantic matching, search term relevance calculation, and comprehensive recommendation ranking.

[0048] In some embodiments, based on the brand demand vector set and the account profile vector set, the multidimensional semantic matching results and search term relevance of candidate accounts are calculated, including: Based on the brand marketing campaign objectives, determine the sub-vectors of needs to be matched and their corresponding dimension weights from the brand demand vector set; The similarity between the demand sub-vector and the corresponding profile vector in the account profile vector set is calculated, and the similarity of each dimension is fused according to the corresponding dimension weight to obtain the multi-dimensional semantic matching result. The search demand vector is extracted from the brand demand vector set, and then jointly matched with the account content semantic vector, the account audience demand vector, and the account scenario vector to generate search term relevance.

[0049] Specifically, when milk powder brands conduct new product launch awareness campaigns and search engine optimization (SEO) campaigns, the system generates a brand demand vector set and multiple candidate account profile vector sets. The brand demand vector set includes sub-vectors for selling points, target audience needs, scenario needs, search needs, competitor needs, and risk needs. The account profile vector set includes semantic vectors for account content, target audience needs, and scenario needs, and is linked to the account's recent content themes, comment intent, and commercial performance characteristics.

[0050] The system first determines the matching focus for this round based on the brand's marketing campaign objectives. If the current campaign is about new product launch and brand building, the system prioritizes sub-vectors representing selling points such as freshness, nutrition, easy absorption, and improved gut health, as well as scenario-based needs such as transitioning to new formula, seasonal feeding changes, sensitive baby digestion, and abnormal bowel movements. It also increases the weight of selling point and scenario dimensions. If the current campaign is about search engagement, the system prioritizes sub-vectors representing search needs such as recommendations for new formula, formula for babies with sensitive digestion, easy formula absorption, and formula reviews. It also increases the weight of search, audience needs, and comment intent dimensions. If the campaign is focused on competitor interception, the system includes competitor needs such as competitor names, alternatives, and comparisons in the key matching scope, and increases the weight of competitor-related content coverage. This process allows the matching focus to dynamically change with different campaign objectives, rather than using fixed keywords or uniform ranking rules.

[0051] For candidate account A, the system retrieves the account content semantic vector, the account audience demand vector, and the account scenario vector from the account profile vector set. Account A's content over the past 90 days has consistently focused on baby's feeding status, formula switching experiences, formula reviews, feeding of babies with sensitive stomachs, and observations of baby's bowel movements. Comments in the comment section include questions such as "Is it suitable for sensitive babies?", "Will switching formulas cause diarrhea?", "Which is better compared to another brand?", and "How to buy?". The system calculates the similarity between the selling point demand sub-vector and the account content semantic vector to determine if the account content effectively conveys brand selling points such as good absorption and stable digestion. It also calculates the similarity between the audience demand sub-vector and the account audience demand vector to determine if the comment section users closely resemble the target decision-making group, such as mothers of babies aged 0-3 years and new mothers. Finally, it calculates the similarity between the scenario demand sub-vector and the account scenario vector to determine if the account consistently covers consumption scenarios such as formula switching, sensitive stomachs, and abnormal bowel movements. After calculating the similarity across all dimensions, the system integrates them according to the dimensional weights corresponding to the advertising objectives, generating a multi-dimensional semantic matching result.

[0052] During similarity fusion, the system also verifies the effectiveness of the matching dimensions. If an account's content has high semantic similarity, but the comment section lacks genuine consultation and decision-making comparison questions, the audience demand dimension score will be low, and the overall matching result will not be excessively improved solely due to content similarity. If an account meets the requirements for follower count and pricing, but its content mainly focuses on parenting, children's education, or family daily life, with only a few notes mentioning baby formula, the system will reduce the contribution of the scenario and selling point dimensions. If an account's recent content themes differ significantly from its historical themes, the system will adjust the account's content semantic vector based on the weight of recent content, making the multi-dimensional semantic matching results more reflective of the current ad availability.

[0053] For search term relevance, the system extracts search demand vectors from the brand demand vector set and performs joint matching with account content semantic vectors, account audience demand vectors, and account scenario vectors. Taking baby formula for sensitive stomachs as an example, even if candidate account B's historical content does not directly contain the complete search term, multiple posts contain semantic expressions such as baby's sensitive stomach, discomfort after switching formulas, irregular bowel movements, and the need to consider absorption when choosing formula. Furthermore, the comments section contains questions about whether the formula is suitable for a sensitive baby, whether it will cause bloating, and how to switch formulas. Therefore, the system can still determine that account B is suitable for relevance based on the similarity of content semantics, comment demand, and scenario coverage. For recommendations on formula for switching baby formulas, the system focuses on the semantic content of the account's content, such as switching experiences, stage change records, and feeding feedback, as well as the intensity of questions in the comments section regarding switching methods, diarrhea risks, and brand comparisons, generating corresponding search term relevance.

[0054] After calculating the relevance of each target search term, the system establishes a connection between the account and the applicable search terms, retaining the basis for this connection. For example, account A can be tagged as suitable for recommending formula for babies with sensitive stomachs, formula reviews, and other related topics; account B can be tagged as suitable for recommending easily absorbed formula and alternative formulas to competitors. This connection is subsequently used to generate recommendation explanations, enabling advertisers to clearly define the appropriate search directions and content themes for each account.

[0055] Through the above embodiments, the system can dynamically select demand sub-vectors and dimension weights according to different delivery targets, and generate multi-dimensional semantic matching results and search term relevance by calculating the joint similarity of account content, comment demand and consumption scenario. This reduces omissions and misselections caused by relying solely on keywords or account categories, and improves the accuracy of account semantic filtering and the reliability of search relevance evaluation.

[0056] In some embodiments, the similarity between the demand sub-vector and the corresponding profile vector in the account profile vector set is calculated, and the similarity of each dimension is fused according to the corresponding dimension weights to obtain a multi-dimensional semantic matching result, including: Based on the dimension identifier of the demand sub-vector, the corresponding portrait vector with the same or related semantic dimensions is determined in the account portrait vector set; Vector normalization is performed on the demand sub-vector and the corresponding profile vector, and the dimensional similarity of each matching dimension is calculated according to the preset similarity function; Based on the brand marketing campaign objectives and the effectiveness status of each matching dimension, the corresponding dimension weights are calibrated to obtain the calibrated dimension weights. The similarity of each dimension is fused by calibrating the dimension weights, and multidimensional semantic matching results are generated by combining the missing dimension compensation rules.

[0057] Specifically, in the scenario of selecting accounts for infant formula brand search, the system has generated a set of brand demand vectors and a set of candidate account profile vectors. In the brand demand vector set, each demand sub-vector carries a dimension identifier, a task identifier, and a source field identifier. For example, "easy to absorb" and "stable digestive system" correspond to the selling point dimension; "0-3 year old babies and their mothers" corresponds to the target audience dimension; "switching formula," "baby's sensitive digestive system," and "abnormal bowel movements" correspond to the scenario dimension; and "baby's sensitive digestive system formula" and "switching formula recommendations" correspond to the search dimension. In the account profile vector set, the account content semantic vector, the account audience demand vector, and the account scenario vector correspond to the account's published content, comment intent, and consumption scenario coverage, respectively.

[0058] Based on the dimensional identifiers of the demand sub-vectors, the system searches for profile vectors with the same or related semantic dimensions in the account profile vector set. For selling point demand sub-vectors, the system prioritizes matching the account's content semantic vector to determine whether the account's notes effectively convey selling points such as good absorption, stable digestion, and fresh nutrition. For audience demand sub-vectors, the system prioritizes matching the account's audience demand vector to determine whether commenters are similar to new mothers, mothers of babies aged 0-3, and those concerned with digestion and absorption. For scenario demand sub-vectors, the system prioritizes matching the account's scenario vector to determine whether the account consistently covers scenarios such as transitioning to new formula, seasonal feeding changes, digestive sensitivities, and abnormal bowel movements. If no two demand sub-vectors have completely identical dimensions, the system determines the related semantic dimensions based on the semantic dimension mapping table. For example, a baby's discomfort after drinking milk can be associated with both digestive sensitivities and good absorption selling points; comparing a product to another brand can be associated with both competitor comparisons and the target audience for purchasing decisions.

[0059] After determining the corresponding profile vector, the system performs vector normalization on the demand sub-vector and the corresponding profile vector to ensure that vectors generated from different model versions, field lengths, and text sources are on a uniform scale. After normalization, the system calculates the dimensional similarity of each matching dimension according to a preset similarity function. This similarity function can be cosine similarity or a similarity function after distance transformation. Taking candidate account A as an example, its many posts in the past 90 days involve experiences with switching formulas, formula reviews, feeding babies with sensitive stomachs, and observations of baby stools. The comments section frequently raises questions such as whether it is suitable for sensitive babies, whether switching formulas will cause diarrhea, and which brand is better compared to another. The system calculates that account A has high similarity in the selling point dimension, scenario dimension, and target audience demand dimension, but only a small amount of content matches in the competitor interception dimension, thus the similarity in this dimension is low.

[0060] The system further calibrates the weights of corresponding dimensions based on the brand marketing task's objectives and the effectiveness status of each matching dimension. For example, if the current task is new product launch promotion, the basic weights of the selling point and scenario dimensions are higher; if the current task is search engagement, the basic weights of the search dimension, audience needs dimension, and scenario dimension are higher; if the current task is competitor interception, the basic weights of the competitor dimension and comparison intent dimension are higher. The effectiveness status is used to determine whether the data for a certain dimension is sufficient to support matching. For example, if the number of account comments is too low, the comment text effectiveness rate is low, or the recent content sample is insufficient, the system reduces the weight of the corresponding dimension; if the account has recently been continuously publishing milk powder reviews and switching experiences with sufficient inquiries in the comment section, the system increases the weight of the relevant dimension. Through this calibration method, the system avoids the abnormal amplification of a single dimension due to missing or low-quality data.

[0061] During the weighted fusion phase, the system uses calibration dimension weights to weightedly fuse the similarities of each dimension, generating initial multidimensional semantic matching results. If a candidate account has missing dimensions, the system processes it according to the missing dimension compensation rules. For accounts that have not received comment data from the platform but whose published content is highly consistent with the target scenario, the system does not directly determine them as low-matching accounts, but instead lowers the confidence level of the audience demand dimension and provides limited compensation using the content semantic dimension and scenario dimension. For accounts that only have structured data such as follower count and pricing but lack recent content, the system does not provide high-score compensation, but instead marks them as accounts awaiting supplementation of their profile. If an account only hits the milk powder keyword in a single viral post, but its long-term content leans towards parenting or children's education, the system lowers the fusion result due to insufficient scenario continuity.

[0062] Through the above embodiments, the system can accurately align the demand sub-vector and the account profile vector based on the dimension identifier, and generate stable multi-dimensional semantic matching results through normalization, similarity calculation, weight calibration and missing dimension compensation. This reduces misselection caused by mismatch of different semantic dimensions and avoids distortion of account suitability judgment due to missing local data or abnormal single content.

[0063] In some embodiments, a comprehensive account recommendation score is generated by combining business performance characteristics, the status of screening parameter satisfaction, historical preference information, and negative exclusion matching results, including: The commercial suitability of candidate accounts is determined based on their commercial performance characteristics, and the satisfaction of structured conditions is determined based on the status of the screening parameters. Based on user account operation behavior and campaign performance data in historical tasks, generate preference fit corresponding to the current brand marketing task; The risk filtering status and negative deduction value are determined based on the negative exclusion matching results; The scoring weights are configured according to the brand marketing task's delivery objectives. The multi-dimensional semantic matching results, search term relevance, commercial suitability, structured condition satisfaction, preference suitability, and negative deduction values ​​are integrated to generate a comprehensive account recommendation score.

[0064] Specifically, in the scenario of recommending and ranking accounts for infant formula brands, after obtaining the multi-dimensional semantic matching results and search term relevance of candidate accounts, the system further combines the account's commercial performance, the fulfillment of platform screening parameters, user historical preferences, and negative exclusion matching to generate a comprehensive account recommendation score. Taking the current brand marketing task of generating interest in new product launches while also considering search relevance as an example, the brand requires accounts to express selling points such as fresh nutrition, easy absorption, and more stable digestive system, covering scenarios such as switching formulas, seasonal feeding changes, infant gastrointestinal sensitivity, and abnormal bowel movements, and meeting screening conditions such as 10,000-50,000 followers, a price not exceeding 3,500, the ability to advertise both videos and images, and the exclusion of food review accounts.

[0065] The system first determines the commercial suitability of candidate accounts based on their business performance characteristics. For candidate account A, the system reads its effective interaction ratio, collection and comment structure, comment demand intensity, pricing level, content update stability, historical cooperation density, and historical campaign feedback. If account A has stable content interaction over the past 90 days, a collection rate higher than the average of similar maternal and infant accounts, a large number of genuine product inquiries and brand comparison questions in the comments section, and both text / image and video pricing are below the budget limit, and its historical cooperation performance record is normal, the system assigns account A a high commercial suitability score. For candidate account B, although the content semantic match is high, the pricing is significantly higher than the range of similar accounts, and the proportion of historical commercial notes is too high, with few purchase inquiries in the comments section. The system lowers its pricing rationality and interaction quality scores to prevent the commercial suitability score from being excessively increased due to semantic relevance.

[0066] Subsequently, the system determines the satisfaction level of structured criteria based on the status of the filtering parameters. The system matches fields such as account follower count, price, content format, platform category, region, and account status against the filtering parameters for this round. If a candidate account has 32,000 followers, a text / image price of 2,800, a video price of 3,300, supports both text / image and video content, and is in a "cooperation-ready" status, then the structured criteria satisfaction level is relatively high. If the account's follower count is within the required range but the video price is missing, the system lowers the satisfaction level according to the field missing rules and marks it as needing a supplementary price; if the account's follower count or price exceeds the hard limits, the system can directly downgrade or remove the account to prevent it from entering the high-priority recommendation pool.

[0067] When generating preference fit, the system retrieves user account operation behavior and campaign performance data from historical tasks. Historical task data includes records of accounts the user has previously selected, pending, eliminated, favorited, and actually collaborated with, as well as post-campaign data such as readership, interaction rate, favorite rate, comment quality, search traffic, store visits, conversion rates, and cost performance. If the current user has repeatedly selected maternal and infant KOC accounts with authentic content, comprehensive comments and consultations, and 10,000-50,000 followers in past milk powder tasks, and has eliminated accounts that are overly commercialized or overly focused on parenting, the system maps this preference to the preference fit for the current task. If an account is similar to a user's historically high-performing accounts in terms of content scenario, account level, pricing range, and comment intent, the preference fit increases; if the account is similar to accounts the user has repeatedly eliminated, the preference fit decreases.

[0068] For negative exclusion matching results, the system determines the risk filtering status and negative deduction value. The system matches the user-recommended food review accounts with the static tags, recent note tags, content summaries, platform categories, and comment keywords of the candidate accounts. If an account's recent content mainly consists of restaurant reviews, store visits, and food evaluations, even if it matches Sam's Club or family consumption related semantics, the system still judges it as a strong match for negative exclusion conditions and configures the filtering status. If an account only occasionally posts content about family food selection, but the main content is about maternal and infant feeding, formula reviews, and baby gastrointestinal care, the system judges it as a weak match and only applies a low negative deduction. For accounts in the maternal and infant category that involve risky content such as exaggerated efficacy claims, medical expressions, or absolute promises, the system will also add brand safety deductions.

[0069] The system assigns scoring weights based on the brand marketing campaign's objectives. In new product launch and seeding campaigns, multi-dimensional semantic matching results and commercial fit are given higher weights to ensure recommended accounts both convey the brand's selling points and possess advertising value. In search campaigns, search term fit and comment demand relevance are weighted more heavily, prioritizing accounts suitable for handling searches related to baby formula for sensitive stomachs or recommendations for transitioning to other formulas. In low-cost volume expansion campaigns, structured condition fulfillment and pricing reasonableness are weighted more heavily. In high-risk category campaigns, negative deductions and brand safety risk are weighted more heavily. The system integrates multi-dimensional semantic matching results, search term fit, commercial fit, structured condition fulfillment, preference fit, and negative deductions according to the aforementioned weights to generate a comprehensive account recommendation score, retaining the contribution of each component as a basis for subsequent recommendation interpretation.

[0070] Through the above embodiments, the system can avoid relying solely on content similarity, number of followers, or price for ranking. Instead, it incorporates account semantic adaptation, search acceptance, commercial value, hard conditions, historical preferences, and negative risks into a comprehensive scoring process. This improves the consistency between account recommendation results and brand advertising goals, and reduces the probability of recommending accounts that do not meet the user's explicit exclusion criteria or lack commercial value.

[0071] In some embodiments, account recommendation results are output based on the comprehensive account recommendation score, and recommendation explanation information is generated, including: Candidate accounts are sorted according to their comprehensive recommendation scores, and then divided into sets of recommended accounts for different purposes based on preset recommendation thresholds and ad placement rules. For target accounts in the recommended account set, extract the matching dimensions that contribute to the account's overall recommendation score, the screening parameters that are met, and the risk factors that are hit; Based on the extracted results, recommendation explanation information is generated, and the target account, the account's overall recommendation score, the corresponding purpose of the campaign, and the recommendation explanation information are linked and output as the account recommendation results.

[0072] Specifically, after the milk powder brand completes the comprehensive scoring of candidate accounts, the system enters the process of outputting recommendation results and generating explanations. The current brand marketing task is to generate buzz for the new product launch while also ensuring search traffic. The product emphasizes fresh nutrition, easy absorption, and a more stable digestive system for babies. The target audience is mothers of babies aged 0-3 years, with key scenarios including switching formulas, seasonal feeding changes, and addressing infant gastrointestinal sensitivity and bowel abnormalities. Target search terms include milk powder recommendations for switching formulas, formula for babies with sensitive stomachs, easily absorbed formula, and formula reviews. The system has generated a comprehensive recommendation score for each candidate account and retains sub-results such as multi-dimensional semantic matching results, search term relevance, commercial suitability, structured condition satisfaction, preference suitability, and negative deductions.

[0073] The system first sorts candidate accounts in descending order based on their overall recommendation score, and then generates a set of recommended accounts based on preset recommendation thresholds and campaign rules. The preset recommendation thresholds can be dynamically determined based on the brand's budget, the desired number of accounts, and the quality of the candidate accounts. For example, accounts with a total score above 85 and no strong risk factors are included in the core recommendation range; accounts with a total score between 70 and 85 and a single advantageous scenario are included in the special recommendation range; and accounts with a total score below 70 or that meet strong exclusion criteria are included in the low-priority or non-recommended range.

[0074] In some examples, the ad placement rules are used to further categorize account purposes. If an account has high selling point matching, scenario coverage, and content quality, the system will classify it into the core seeding account set. If an account has a high degree of relevance to target search terms and the intention to inquire in the comments section is obvious, the system will classify it into the search relevance account set. If an account's historical content has competitor comparison semantics and the risk is controllable, the system will classify it into the competitor interception account set. If an account has a low price, stable interaction, and a high degree of fulfillment of structured conditions, the system will classify it into the high cost-effectiveness volume expansion account set. If an account hits the tags of exaggerated efficacy, medical expression, excessive commercialization, or user explicit exclusion, the system will classify it into the brand safety risk account set or the not recommended account set.

[0075] For target accounts within the recommended account set, the system extracts the matching dimensions that primarily contribute to the account's overall recommendation score. Taking target account A as an example, the system identifies a high proportion of its content from the past 90 days related to switching formulas, formula reviews, feeding advice for babies with sensitive digestive systems, and observations of bowel movements. These selling point and scenario dimensions significantly contribute to the overall score. In the comments section, questions such as "Is this suitable for sensitive babies?", "Will switching formulas cause diarrhea?", "Which is better compared to another brand?", and "How to buy?" appear frequently, indicating significant contributions from audience needs and search term relevance. The account's collection rate, effective comment rate, and content update stability are higher than the average for similar mother and baby accounts, suggesting a high degree of commercial suitability. The system also extracts screening parameters that the target account meets, such as a follower count between 10,000 and 50,000, text / image and video pricing not exceeding 3,500, support for both text / image and video content, an account status indicating potential collaboration, and no strong exclusion criteria for food reviews.

[0076] For accounts with risk factors, the system simultaneously extracts the evidence for risk detection. Taking target account B as an example, its content semantics are quite similar to baby formula for sensitive stomachs, but it has recently had a high density of commercial collaborations, and some notes contain risky text with overly strong efficacy claims. The system includes the risks of excessive commercialization and compliant expression as risk factors in the explanation information and lowers its recommended usage level. For accounts with weak negative exclusion criteria, the system does not directly remove them, but explains in the recommendation explanation that they have a small amount of historical content that does not conform to the exclusion direction, and lists them as alternative accounts; for accounts with strong negative exclusion criteria, the system does not output them as recommended accounts, but only saves the filtering reasons in the background audit records.

[0077] Based on the extracted results, the system generates recommendation explanations. These explanations include the account's suitable advertising purpose, core recommendation criteria, applicable search terms, suggested content direction, met hard screening conditions, and potential risk warnings. For example, for account A, the system-generated explanation can be summarized as follows: This account's recent content consistently covers switching formulas, formula reviews, feeding of babies with sensitive stomachs, and observation of bowel movements, which aligns well with the selling points of good absorption and stable digestion; the comment section contains numerous inquiries about switching formulas, sensitive babies, brand comparisons, and purchasing methods, making it suitable for handling search terms such as switching formulas and formulas for babies with sensitive stomachs; the price and number of followers meet the screening conditions for this round, making it suitable for core product seeding and search engine optimization. The system correlates the target account, overall account recommendation score, sub-scores, corresponding advertising purpose, applicable search terms, suggested content direction, and risk warnings, outputting the account recommendation results, which are then grouped and displayed on the front-end page according to the recommended account set.

[0078] Through the above embodiments, the system can convert the comprehensive account recommendation score into account recommendation results for ad placement decisions, and simultaneously output recommendation explanation information corresponding to the scoring process. This allows brands to not only see the ranking of recommended accounts, but also understand what ad placement purpose the account is suitable for, why it was recommended, which search terms it can handle, and what risks exist, thereby improving the interpretability of account screening results and the reliability of ad placement decisions.

[0079] The following are system embodiments of this application, which can be used to execute the method embodiments of this application. For details not disclosed in the system embodiments of this application, please refer to the method embodiments of this application.

[0080] Figure 2 This is a schematic diagram of the structure of a social media account filtering system based on vectorized semantic matching provided in an embodiment of this application. Figure 2 As shown, the system includes: The acquisition module 201 is used to acquire product information, user identity information and natural language filtering requirements corresponding to brand marketing tasks, perform intent parsing on natural language filtering requirements, and generate brand requirement fields, structured filtering conditions and negative exclusion conditions. The mapping module 202 is used to vectorize the brand requirement field, generate a set of brand requirement vectors, and map the structured filtering conditions and negative exclusion conditions into filtering parameters that can be executed by the account platform. The retrieval module 203 is used to first retrieve candidate accounts in the local account asset library based on the filtering parameters, and when the retrieval results do not meet the preset candidate conditions, call the external account platform to expand the candidate accounts, and perform database entry and vectorization processing on the expanded accounts. The parsing module 204 is used to obtain the account data of each candidate account, and to parse the content, comments, scenarios, interactions and business information in the account data to generate an account profile vector set and business performance characteristics; The calculation module 205 is used to calculate the multidimensional semantic matching results and search term relevance of candidate accounts based on the brand demand vector set and the account profile vector set, and generate a comprehensive recommendation score for the account by combining business performance characteristics, the status of screening parameters, historical preference information and negative exclusion matching results. The generation module 206 is used to output account recommendation results based on the comprehensive recommendation score of the account and generate recommendation explanation information.

[0081] In some embodiments, Figure 2 The acquisition module 201, based on product information, user identity information, and the task context of brand marketing tasks, performs semantic segmentation and intent recognition on natural language filtering requirements to determine semantic requirement fragments, numerical constraint fragments, and negative expression fragments related to the target audience; it categorizes the semantic requirement fragments into brand communication elements to generate brand requirement fields; it performs unit conversion, interval completion, and field normalization on the numerical constraint fragments to generate structured filtering conditions; and it parses the negative expression fragments based on the negative trigger words and their corresponding scope of application to generate negative exclusion conditions for account filtering or deduction.

[0082] In some embodiments, Figure 2 The mapping module 202 performs semantic classification and field standardization of brand demand fields according to brand communication dimensions, calls the vector generation model to generate demand sub-vectors corresponding to each brand communication dimension, and configures matching weights for demand sub-vectors according to the delivery goals of brand marketing tasks to obtain a set of brand demand vectors; according to the field mapping rules of the account platform, it performs field transformation, value normalization and constraint merging on the structured filtering conditions to generate positive filtering parameters; it generates exclusion filtering parameters according to negative exclusion conditions, and performs consistency verification and conflict resolution on the positive filtering parameters and exclusion filtering parameters to obtain the filtering parameters that can be executed by the account platform.

[0083] In some embodiments, Figure 2 The retrieval module 203 performs structured conditional matching and vector similarity recall in the local account asset library according to the filtering parameters to obtain a local candidate account set. Based on the number of accounts, matching quality and coverage of the local candidate account set, it determines whether the retrieval results meet the preset candidate conditions. When the retrieval results do not meet the preset candidate conditions, it generates a pagination expansion request according to the filtering parameters and calls an external account platform to obtain extended account data. It performs deduplication, field standardization and validity verification on the extended account data, writes the verified extended accounts into the local account asset library, and generates corresponding account vectors to update the candidate accounts.

[0084] In some embodiments, Figure 2 The parsing module 204 acquires account data for each candidate account within a preset time range, cleans, deduplicates, and normalizes the format of the account data to generate standardized account data; it performs topic identification and semantic vectorization on the published content in the standardized account data, and weights and aggregates the content vectors based on publication time, interaction performance, and content quality to generate account content semantic vectors; it identifies the demand intent of the comment text in the standardized account data, extracts comment semantic information related to user consultation, decision comparison, and consumption feedback, and generates account audience demand vectors; it identifies the consumption scenarios covered by the account based on the published content and comment semantic information, and generates account scenario vectors; it generates an account profile vector set based on the account content semantic vectors, account audience demand vectors, and account scenario vectors, and generates business performance characteristics based on interaction quality, pricing status, and cooperation stability.

[0085] In some embodiments, Figure 2 The calculation module 205 determines the demand sub-vectors to be matched and their corresponding dimension weights from the brand demand vector set according to the brand marketing task's delivery objectives; it calculates the similarity between the demand sub-vectors and the corresponding portrait vectors in the account portrait vector set, and merges the similarity of each dimension according to the corresponding dimension weights to obtain a multi-dimensional semantic matching result; it extracts the search demand vectors from the brand demand vector set, and performs joint matching between the search demand vectors and the account content semantic vectors, account audience demand vectors, and account scenario vectors respectively to generate search term relevance.

[0086] In some embodiments, Figure 2 The calculation module 205, based on the dimension identifier of the demand sub-vector, determines the corresponding portrait vector with the same or related semantic dimensions in the account portrait vector set; performs vector normalization on the demand sub-vector and the corresponding portrait vector, and calculates the dimensional similarity of each matching dimension according to the preset similarity function; calibrates the corresponding dimension weights according to the delivery goals of the brand marketing task and the effectiveness status of each matching dimension, and obtains the calibrated dimension weights; uses the calibrated dimension weights to fuse the similarity of each dimension, and generates a multi-dimensional semantic matching result by combining the missing dimension compensation rule.

[0087] In some embodiments, Figure 2The calculation module 205 determines the commercial suitability of candidate accounts based on their commercial performance characteristics and the structured condition satisfaction based on the status of the screening parameters. Based on the user's account operation behavior and campaign performance data in historical tasks, it generates the preference suitability corresponding to the current brand marketing task. Based on the negative exclusion matching results, it determines the risk filtering status and negative deduction value. According to the campaign goals of the brand marketing task, it configures the scoring weights and integrates the multi-dimensional semantic matching results, search term relevance, commercial suitability, structured condition satisfaction, preference suitability, and negative deduction value to generate a comprehensive account recommendation score.

[0088] In some embodiments, Figure 2 The generation module 206 sorts candidate accounts according to their comprehensive recommendation scores and divides them into sets of recommended accounts for different purposes based on preset recommendation thresholds and delivery scenario rules. For target accounts in the recommended account sets, it extracts the matching dimensions that contribute to the comprehensive recommendation score, the screening parameters that are met, and the risk factors that are hit. Based on the extraction results, it generates recommendation explanation information and outputs the target account, the comprehensive recommendation score, the corresponding delivery purpose, and the recommendation explanation information as the account recommendation result.

[0089] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0090] Figure 3 This is a schematic diagram of the electronic device 3 provided in an embodiment of this application. Figure 3 As shown, the electronic device 3 of this embodiment includes: a processor 301, a memory 302, and a computer program 303 stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program 303, it implements the steps in the various method embodiments described above. Alternatively, when the processor 301 executes the computer program 303, it implements the functions of each module / unit in the various system embodiments described above.

[0091] Electronic device 3 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 3 may include, but is not limited to, processor 301 and memory 302. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or different components.

[0092] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0093] The memory 302 can be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 302 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 3. The memory 302 can also include both internal and external storage units of the electronic device 3. The memory 302 is used to store computer programs and other programs and data required by the electronic device.

[0094] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0095] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a readable storage medium (e.g., a computer-readable storage medium). Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which may be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0096] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for filtering social media accounts based on vectorized semantic matching, characterized in that, include: Obtain product information, user identity information, and natural language filtering requirements corresponding to brand marketing tasks; perform intent parsing on the natural language filtering requirements; and generate brand requirement fields, structured filtering conditions, and negative exclusion conditions. The brand requirement field is vectorized to generate a brand requirement vector set, and the structured filtering conditions and the negative exclusion conditions are mapped to filtering parameters that can be executed by the account platform. Based on the filtering parameters, candidate accounts are first retrieved from the local account asset library. If the retrieval results do not meet the preset candidate conditions, an external account platform is called to expand the candidate accounts, and the expanded accounts are then stored in the library and vectorized. Acquire account data for each candidate account, analyze the content, comments, scenarios, interactions, and business information in the account data, and generate an account profile vector set and business performance characteristics; Based on the brand demand vector set and the account profile vector set, calculate the multidimensional semantic matching results and search term relevance of the candidate accounts, and generate a comprehensive account recommendation score by combining the business performance characteristics, the status of the screening parameters, historical preference information and negative exclusion matching results. Based on the comprehensive recommendation score of the account, the account recommendation results are output, and recommendation explanation information is generated.

2. The method according to claim 1, characterized in that, The process of parsing the natural language filtering request to generate brand requirement fields, structured filtering conditions, and negative exclusion conditions includes: Based on the product information, the user identity information, and the task context of the brand marketing task, the natural language filtering requirements are semantically segmented and intent-recognized to determine semantic requirement segments, numerical constraint segments, and negative expression segments related to the target audience. The semantic requirement fragments are categorized into brand communication elements to generate the brand requirement field; The numerical constraint fragment is subjected to unit conversion, interval completion, and field normalization to generate the structured filtering conditions; The negative expression fragment is parsed based on the negative trigger word and its corresponding scope of application to generate the negative exclusion conditions used for account filtering or point deduction.

3. The method according to claim 1, characterized in that, The process of vectorizing the brand requirement field to generate a brand requirement vector set, and mapping the structured filtering conditions and the negative exclusion conditions to filtering parameters executable by the account platform, includes: The brand demand fields are semantically categorized and standardized according to the brand communication dimension. A vector generation model is called to generate demand sub-vectors corresponding to each brand communication dimension. Matching weights are configured for the demand sub-vectors according to the delivery goals of the brand marketing task to obtain the brand demand vector set. Based on the account platform's field mapping rules, the structured filtering conditions are subjected to field transformation, value normalization, and constraint merging to generate positive filtering parameters; Based on the negative exclusion conditions, exclusion filtering parameters are generated, and the positive filtering parameters and the exclusion filtering parameters are subjected to consistency verification and conflict resolution to obtain the filtering parameters that can be executed by the account platform.

4. The method according to claim 1, characterized in that, The process involves first retrieving candidate accounts from the local account asset database based on the filtering parameters. If the search results do not meet the preset candidate conditions, an external account platform is invoked to expand the candidate accounts. The expanded accounts are then stored in the database and vectorized, including: Based on the filtering parameters, structured conditional matching and vector similarity recall are performed in the local account asset database to obtain a local candidate account set; Based on the number of accounts, matching quality, and coverage of the local candidate account set, determine whether the search results meet the preset candidate conditions; When the search results do not meet the preset candidate conditions, a pagination expansion request is generated according to the filtering parameters, and the external account platform is called to obtain extended account data. The extended account data is deduplicated, standardized, and validated. The validated extended accounts are written into the local account asset library, and corresponding account vectors are generated to update the candidate accounts.

5. The method according to claim 1, characterized in that, The process of acquiring account data for each candidate account, analyzing the content, comments, scenarios, interactions, and commercial information within the account data, and generating an account profile vector set and commercial performance characteristics includes: Obtain account data for each candidate account within a preset time range, clean, deduplicate, and normalize the format of the account data to generate standardized account data; The standardized account data is used to identify topics and semantically vectorize the published content. The content vectors are then weighted and aggregated based on publication time, interaction performance, and content quality to generate account content semantic vectors. The demand intent is identified in the comment text of the standardized account data, and the semantic information of the comments related to user consultation, decision comparison and consumption feedback is extracted to generate an account audience demand vector. Based on the published content and the semantic information of the comments, the consumption scenarios covered by the account are identified, and an account scenario vector is generated; The account profile vector set is generated based on the account content semantic vector, the account audience demand vector, and the account scenario vector, and the business performance characteristics are generated based on interaction quality, pricing status, and cooperation stability.

6. The method according to claim 1, characterized in that, The step of calculating the multidimensional semantic matching results and search term relevance of candidate accounts based on the brand demand vector set and the account profile vector set includes: Based on the target of the brand marketing task, determine the demand sub-vectors to be matched and their corresponding dimension weights from the brand demand vector set; The similarity between the demand sub-vector and the corresponding portrait vector in the account portrait vector set is calculated, and the similarity of each dimension is fused according to the corresponding dimension weight to obtain the multi-dimensional semantic matching result. The search demand vector is extracted from the brand demand vector set, and the search demand vector is jointly matched with the account content semantic vector, the account audience demand vector and the account scenario vector to generate the search term relevance.

7. The method according to claim 6, characterized in that, The step of calculating the similarity between the demand sub-vector and the corresponding profile vector in the account profile vector set, and fusing the similarities of each dimension according to the corresponding dimension weights to obtain the multi-dimensional semantic matching result includes: Based on the dimension identifier of the demand sub-vector, determine the corresponding portrait vector with the same semantic dimension or related semantic dimension in the account portrait vector set; The demand sub-vector and the corresponding profile vector are normalized, and the dimensional similarity of each matching dimension is calculated according to a preset similarity function. Based on the campaign objectives of the brand marketing task and the effectiveness status of each matching dimension, the weights of the corresponding dimensions are calibrated to obtain the calibrated dimension weights. The similarity of each dimension is fused using the calibration dimension weights, and the multidimensional semantic matching result is generated by combining the missing dimension compensation rule.

8. The method according to claim 6, characterized in that, The process of generating a comprehensive account recommendation score by combining the aforementioned business performance characteristics, the status of screening parameters, historical preference information, and negative exclusion matching results includes: The commercial suitability of candidate accounts is determined based on the aforementioned commercial performance characteristics, and the satisfaction of structured conditions is determined based on the satisfaction status of the aforementioned screening parameters. Based on user account operation behavior and campaign performance data in historical tasks, generate preference fit corresponding to the current brand marketing task; The risk filtering status and negative deduction value are determined based on the negative exclusion matching results; The scoring weights are configured according to the brand marketing task's deployment objectives. The multi-dimensional semantic matching results, search term relevance, commercial suitability, structured condition satisfaction, preference suitability, and negative deduction values ​​are then integrated to generate the account's comprehensive recommendation score.

9. The method according to claim 1, characterized in that, The step of outputting account recommendation results based on the comprehensive account recommendation score and generating recommendation explanation information includes: Candidate accounts are sorted according to their comprehensive recommendation scores, and then divided into sets of recommended accounts for different purposes based on preset recommendation thresholds and delivery scenario rules. For the target account in the recommended account set, extract the matching dimensions that contribute to the overall recommendation score of the account, the filtering parameters that are satisfied, and the risk factors that are hit; The recommendation explanation information is generated based on the extraction results, and the target account, the account's comprehensive recommendation score, the corresponding purpose of the campaign, and the recommendation explanation information are associated and output as the account recommendation result.

10. A social media account filtering system based on vectorized semantic matching, characterized in that, include: The acquisition module is used to acquire product information, user identity information and natural language filtering requirements corresponding to brand marketing tasks, perform intent parsing on the natural language filtering requirements, and generate brand requirement fields, structured filtering conditions and negative exclusion conditions. The mapping module is used to vectorize the brand requirement field, generate a set of brand requirement vectors, and map the structured filtering conditions and the negative exclusion conditions into filtering parameters that can be executed by the account platform. The retrieval module is used to first retrieve candidate accounts in the local account asset library based on the filtering parameters, and when the retrieval results do not meet the preset candidate conditions, call the external account platform to expand the candidate accounts, and perform database entry and vectorization processing on the expanded accounts. The parsing module is used to obtain account data for each candidate account, and to parse the content, comments, scenarios, interactions and business information in the account data to generate an account profile vector set and business performance characteristics; The calculation module is used to calculate the multidimensional semantic matching results and search term relevance of candidate accounts based on the brand demand vector set and the account profile vector set, and generate a comprehensive recommendation score for the account by combining the business performance characteristics, the status of the screening parameters, historical preference information and negative exclusion matching results. The generation module is used to output account recommendation results based on the comprehensive recommendation score of the account, and generate recommendation explanation information.

Images