Intelligent recommendation method based on KOL

Abstract

The invention discloses an intelligent recommendation method based on KOL, and relates to the technical field of KOL intelligent recommendation and evaluation. Comprising the following steps: acquiring account multi-dimensional data of a user, and preprocessing the account multi-dimensional data; performing deep analysis and representation learning training on the intelligent recommendation model based on the preprocessed account multi-dimensional data to obtain a trained intelligent recommendation model; acquiring natural language data input by a user, performing field identification on the natural language data by utilizing the intelligent recommendation model, and determining a field flow of a natural language request route of the user; the intelligent recommendation model carries out content recommendation of a corresponding field for the user according to the field process of the natural language request routing of the user, and the risk assessment model carries out risk assessment on the content of the field process of the natural language request routing of the user. Through the method, deep semantics and implicit information can be effectively captured, and the analysis precision and depth of recommendation / evaluation tasks are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of KOL intelligent recommendation and evaluation, and particularly relates to a KOL-based intelligent recommendation method. BACKGROUND

[0002] With the popularity of the Internet and the explosive growth of information, in the vast amount of information, users have difficulty finding content of their interest, and a recommendation agent can provide personalized custom recommendations for users by analyzing their historical behavior and preferences, helping users find interesting content.

[0003] With the growing demand for social media marketing, the existing KOL (Key Opinion Leader) recommendation and evaluation mode in the prior art uses classification, regression, clustering and other algorithms, and adopts causal analysis to analyze KOL. Then, the traditional method often relies on the business capabilities of analysts, and there are many algorithm recognition problems, which require a great training cost to optimize, and it is difficult to capture deep semantic and implicit information, resulting in poor analysis accuracy and depth of the recommendation / evaluation task. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a KOL-based intelligent recommendation method, which can effectively capture deep semantic and implicit information, and significantly improve the analysis accuracy and depth of the recommendation / evaluation task.

[0005] To achieve the purpose of the present application, the following technical solutions are adopted: A KOL-based intelligent recommendation method, comprising the following steps: Obtaining user account multidimensional data and preprocessing the account multidimensional data; Based on the preprocessed account multidimensional data, performing deep analysis and representation learning training on a preset intelligent recommendation model to obtain a trained intelligent recommendation model; Obtaining user input natural language data, using the intelligent recommendation model to perform domain identification on the natural language data, and determining the domain process of the user natural language request routing; The intelligent recommendation model performs content recommendation in the corresponding domain for the user according to the domain process of the user natural language request routing.

[0006] In the technical solution, the acquired account multidimensional data is preprocessed, which can effectively improve the practicability and reliability of the data and reduce the requirement for data computing performance; the intelligent recommendation model is trained based on the preprocessed account multidimensional data, so that the intelligent recommendation model can perform end-to-end deep analysis and representation learning on the original text, thereby further improving the accuracy of the text analysis task (such as topic classification and intention understanding), and effectively performing deep semantic understanding, context association and reasoning ability processing on the account multidimensional data, to significantly improve the accuracy of text analysis and recommendation / evaluation and the depth of semantic understanding.

[0007] Further, the process of preprocessing the account multidimensional data includes: The acquired account multidimensional data includes account basic data, post data and comment data. The account basic data, post data and comment data are subjected to sentiment analysis, text classification, entity recognition, interaction quantity merging, persona identification and fan portrait construction processing. KOL represents a key opinion leader.

[0008] Further, the process of identifying the persona of the account basic data, post data and comment data includes: The account basic data, post data and comment data are subjected to theme extraction, style analysis and value inclination identification based on the intelligent recommendation model, to generate a persona profile containing several core labels.

[0009] Further, the process of identifying the domain of the natural language data by using the intelligent recommendation model includes: If the natural language request input by the user involves single social media account analysis, it is routed to the account analysis domain. If the natural language request input by the user involves content generation or script writing, it is routed to the viral content imitation domain. If the natural language request input by the user involves KOL recommendation and persona analysis, it is routed to the influencer recommendation - persona domain. If the natural language request input by the user involves KOL recommendation without persona analysis, it is routed to the influencer recommendation - no persona domain. If the natural language request input by the user involves negative evaluation or public opinion analysis, it is routed to the risk evaluation domain. The domain process includes account analysis, viral content imitation, influencer recommendation and risk evaluation.

[0010] Further, the process of the intelligent recommendation model performing content recommendation in the corresponding domain for the user according to the domain process routed by the user's natural language request includes: Expert recommendation - person setting field: Obtaining expert preset person setting data by using a preset person setting query tool; Using the intelligent recommendation model to identify the person setting demand in the natural language data input by the user, combining the preset person setting data, calculating the semantic similarity, and setting a similarity threshold according to the semantic similarity, when the value of the semantic similarity is greater than or equal to the similarity threshold, the corresponding word group is taken as a recommendation condition; Querying the preset expert database based on the recommendation condition, and outputting an expert list containing a recommendation reason by the intelligent recommendation model.

[0011] Further, the process of calculating the semantic similarity further includes: Synonym expansion is performed on the person setting demand in the natural language data input by the user, and the similarity is calculated after supplementing related words by using a WordNet library.

[0012] Further, the method further includes the following steps: Based on the preprocessed account multi-dimensional data, the risk assessment model is trained to obtain a trained risk assessment model; The risk assessment model performs risk assessment on the content of the domain process routed by the user's natural language request, and the intelligent recommendation model performs marking processing on the recommended content according to the risk assessment result, and outputs recommended content containing risk identification.

[0013] Further, the process of the risk assessment model performing risk assessment on the content of the domain process routed by the user's natural language request includes: Using a preset encyclopedia information query tool to perform risk assessment on the to-be-processed information identified by the user's input natural language; If the identified to-be-processed information is an expert post, setting a risk label and a sensitive word library according to the content published by the expert, and setting a first risk threshold, a second risk threshold and a third risk threshold according to the risk label and the sensitive word library; If the content published by the expert is within the first risk threshold, it is marked as a red light risk; if the content published by the expert is within the second risk threshold, it is marked as a yellow light risk; if the content published by the expert is within the third risk threshold, it is marked as a green light risk.

[0014] Further, the method further includes the following steps: Based on the risk assessment model performing risk assessment on the content of the domain process routed by the user's natural language request, the risk assessment result is evaluated to calculate an evaluation score: The evaluation score = (CPE / 5 + CPM / 5 + the completion rate / 5 + the interaction amount / 5 + the browsing amount / 5) / 4*25 + the cooperation competitor matching degree / 4*25 + the fan portrait matching degree / 4*25 + the person set consistency / 4*25; Wherein, CPE represents pay-per-participation, CPM represents cost per thousand impressions, and if some dimension data is missing, the remaining dimension weights are distributed proportionally.

[0015] In the above technical solution, based on the risk label and the sensitive word library, the risks in metaphors, irony and network language can be identified, and the false negative rate is reduced.

[0016] Further, the Moe method is used to process different field processes by using model nodes corresponding to the specialized field, including: Account analysis field: content element extraction is performed by using an intelligent recommendation model; Risk assessment field: risk information of text and images is processed by using a risk assessment model; Wherein, the Moe represents a hybrid expert mode.

[0017] In the above technical solution, the Moe (hybrid expert mode) concept is used, and a dynamic routing mechanism based on field recognition + modular intelligent recommendation model node arrangement (workflow engine coordination) is used, different tasks (account analysis, explosive article imitation, talent recommendation, risk assessment) are processed by intelligent recommendation model nodes specialized in the field, and the illusion rate is reduced.

[0018] Compared with the prior art, the present application has the following advantages: The present application provides an intelligent recommendation method based on KOL, which can effectively improve the practicability and reliability of data and effectively reduce the requirement for data calculation performance by preprocessing the obtained account multidimensional data; the intelligent recommendation model is trained based on the preprocessed account multidimensional data, so that the intelligent recommendation model can perform end-to-end deep analysis and representation learning on the original text, thereby further improving the accuracy of text analysis tasks (such as topic classification and intent understanding), and effectively performing deep semantic understanding, context association and reasoning ability processing on the account multidimensional data, so as to significantly improve the accuracy of text analysis and recommendation / evaluation and the depth of semantic understanding. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A step flowchart of an intelligent recommendation method based on KOL provided by the embodiment of the present application; Figure 2 A field recognition flowchart provided by the embodiment of the present application; Figure 3A process schematic diagram of account analysis / long post analysis field and expert recommendation provided by the embodiment of the present application is provided. Figure 4 A risk assessment identification flowchart provided by the embodiment of the present application is provided. Figure 5 A process schematic diagram of evaluation summary provided by the embodiment of the present application is provided. DETAILED DESCRIPTION

[0020] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the accompanying drawings. The preferred embodiments of the present application are shown in the drawings. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used in the specification of the present application herein are only for the purpose of describing the specific embodiments and are not intended to limit the present application. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0022] Embodiment: The present embodiment provides a KOL-based intelligent recommendation method, referring to Figure 1 , the method comprises the following steps: Step S1: obtaining user account multidimensional data and pre-processing the account multidimensional data; Step S2: based on the pre-processed account multidimensional data, performing deep analysis and representation learning training on the pre-set intelligent recommendation model to obtain a trained intelligent recommendation model; Step S3: obtaining user input natural language data, using the intelligent recommendation model to perform field identification on the natural language data, and determining the field process of user natural language request routing; Step S4: the intelligent recommendation model performs content recommendation in the corresponding field for the user according to the field process of user natural language request routing.

[0023] As a preferred embodiment, in step S1, the pre-processing process of the account multidimensional data comprises: The obtained account multidimensional data includes account basic data, post data and comment data; The account basic data, post data and comment data are subjected to sentiment analysis, text classification, entity recognition, interaction quantity merging, persona identification and fan portrait construction processing; Wherein, KOL represents key opinion leader.

[0024] Further, the process of performing persona recognition on the account basic data, post data and comment data includes: Based on the intelligent recommendation model, the account basic data, post data and comment data are subjected to theme extraction, style analysis and value inclination recognition to generate a persona profile containing several core labels.

[0025] For example, the collected account basic information, posts, comments and other social media data are subjected to pre-processing data processing such as sentiment algorithm, text classification, entity recognition, code table normalization, interaction quantity merging, account type algorithm, business order label processing, persona recognition, and fan portrait construction.

[0026] As a preferred embodiment, in step S2, based on the pre-processed account multi-dimensional data, a prompt (prompt word) is input to the intelligent recommendation model according to the actual business, and an intelligent recommendation model (such as DeepSeek-R1 / DeepSeek-V3-0324) based on the Transformer architecture (i.e., the intelligent recommendation model) is used to replace the discrete data processing step in the traditional algorithm. The intelligent recommendation model has strong deep semantic understanding ability and context modeling ability (mainly through its core self-attention mechanism), which can perform end-to-end deep analysis and representation learning on the original text, thereby further improving the accuracy of text analysis tasks (such as topic classification and intent understanding).

[0027] As a preferred embodiment, in step S3, the process of using the intelligent recommendation model to perform domain recognition on the natural language data includes: If the user input natural language request involves single social media account analysis, it is routed to the account analysis domain; If the user input natural language request involves content generation or script writing, it is routed to the viral content imitation domain; If the user input natural language request involves KOL recommendation and persona analysis, it is routed to the influencer recommendation - persona domain; If the user input natural language request involves KOL recommendation without persona analysis, it is routed to the influencer recommendation - no persona domain; If the user input natural language request involves negative evaluation or public opinion analysis, it is routed to the risk assessment domain; The domain process includes account analysis, viral content imitation, influencer recommendation and risk assessment.

[0028] For example, the embodiment provides a domain recognition method based on an intelligent recommendation model, as shown in Figure 2 The method includes the following steps: Recognition and processing based on natural language input from the user: The analysis involves analyzing a single account on a particular social media platform. Batch KOL recommendation - Persona refers to recommending KOLs or proposing bulk placement conditions and plans, and requiring analysis of persona, content type, account type, influencer type, style, and account attributes. This involves recommending KOLs or proposing bulk placement conditions and plans, and requires analysis of content type, account type, influencer type, style, and account attributes, but does not require bulk recommendations of influencers based on persona - excluding persona; Anything involving negative assessments or public opinion analysis is considered a negative assessment. This involves imitating content, generating content, and writing scripts, all for the purpose of imitating viral articles. If the above-mentioned domain identification conditions are not met, other conditions are considered (such as evaluation and summary), and this invention does not limit them.

[0029] In a preferred embodiment, in step S4, the process by which the intelligent recommendation model recommends content for the user in the corresponding domain based on the domain flow of the user's natural language request routing includes: Expert Recommendation - Persona Category: Use a pre-set persona lookup tool to obtain the influencer's pre-set persona data; The intelligent recommendation model is used to identify the persona requirements in the natural language data input by the user, and combined with the preset persona data, the semantic similarity is calculated. A similarity threshold is set according to the semantic similarity. When the value of the semantic similarity is greater than or equal to the similarity threshold, the corresponding word group is used as the recommendation condition. Based on the recommendation criteria, the system queries a pre-defined database of influencers and outputs a list of influencers with reasons for the recommendations using an intelligent recommendation model.

[0030] Furthermore, the process of calculating semantic similarity also includes: The system expands the user's input natural language data with synonyms for the desired persona, supplements the vocabulary with relevant words from the WordNet lexicon, and then calculates the similarity.

[0031] As a preferred embodiment, this implementation provides a method for influencer recommendation, account analysis, and viral article imitation, such as... Figure 3 As shown, it includes the following steps: Based on the natural language input from the user, the system performs domain identification and routing to the domain: account analysis / viral article imitation. The system extracts influencer information (account name, platform keywords) based on the natural language input of the user using an intelligent recommendation model, and then uses a KOL query tool to verify the accuracy of the information. If the natural language input by the user is not sufficient to identify the influencer information, the default guided intelligent recommendation model will simulate a social media marketing expert according to the corpus reserve, guide the user to reask questions, and let the user supplement the persona, the example price, the specific fan level, the fan portrait preference, etc. If the influencer information is successfully identified, the influencer information query tool is called to check the accuracy of the influencer information, and the analysis portrait query tool and the user information query tool are called to obtain the fan portrait and the user post, which are input into a dedicated account analysis / explosive post analysis model node to output content elements in multiple dimensions such as topic angle, appearance person, scene, and pain point, as well as content label analysis.

[0032] As a preferred embodiment, the present embodiment also provides an influencer recommendation method, which comprises the following steps as shown in the figure: Figure 3 According to the natural language input by the user, the field is identified; Further, the parameter construction intelligent recommendation model is used, and the recommendation filtering condition template is preset to identify platform, account type, emotion type, gender, fan portrait, example price, brand, etc. from natural language data; If it is field-influencer recommendation-persona: Further, the persona query tool is called to query the persona data constructed based on the user and the post; then the persona identified by the new media data processing expert based on the intelligent recommendation simulation of the natural language input by the user is added to the previously queried persona in the library, and input into the vector semantic similarity processing based on Sentence-BERT (sentence vector transformer) to extract the word groups with a similarity higher than 0.2 to the recommendation query condition; If it is field-influencer recommendation-no persona: Then skip the persona similarity processing and integration process; The above analyzed conditions and data are input into the recommendation condition construction intelligent recommendation model after the prompt word arrangement, and the final influencer filtering conditions are output according to the knowledge base recall; Further, the condition query is input into the influencer basic and index information refined by the intelligent recommendation; If there is no related influencer information returned, the user is guided to reask questions and supplement the persona, the example price, and the specific account information; If there is, the influencer data is input into the summary and recommendation after the prompt arrangement, and the final recommended commercial influencer list and the recommendation reason are output according to the natural language semantics input by the user.

[0033] As a preferred embodiment, the method further comprises the following steps: ​Step S5: Train the pre-set risk assessment model based on the pre-processed multi-dimensional account data to obtain the trained risk assessment model; Step S6: The risk assessment model performs a risk assessment on the content of the domain process of the user's natural language request routing. The intelligent recommendation model marks the recommended content according to the risk assessment results and outputs recommended content containing risk identifiers.

[0034] Furthermore, in step S6, the process by which the risk assessment model performs a risk assessment on the content of the domain process of the user's natural language request routing includes: A risk assessment is performed on the information to be processed, which is identified from the natural language input by the user, using a pre-set encyclopedia information query tool. If the identified information to be processed is an article posted by an influencer, then risk tags and a sensitive word library are set according to the content posted by the influencer, and a first risk threshold, a second risk threshold, and a third risk threshold are set according to the risk tags and the sensitive word library. If the content posted by the influencer falls within the first risk threshold, it is marked as a red light risk; if the content posted by the influencer falls within the second risk threshold, it is marked as a yellow light risk; if the content posted by the influencer falls within the third risk threshold, it is marked as a green light risk.

[0035] In an exemplary embodiment, this implementation also provides a risk assessment method, such as... Figure 4 As shown, it includes the following steps: The system identifies the expert information to be processed based on the natural language input from the user. This involves three types of risk assessment standards: Firstly: Yellow light standard for public officials: those identified through encyclopedia entries are marked with a yellow light, while others are marked with a green light; Secondly: The red and yellow light standard for online influencer posts: based on encyclopedia data and verification of negative tags related to the influencer's posts, if the influencer is confirmed to have engaged in criminal or illegal activities, received administrative penalties, or engaged in morally corrupt behavior, regardless of age, it is designated as a red light; if the influencer may have engaged in criminal or illegal activities, received administrative penalties, or engaged in morally corrupt behavior, regardless of age, it is designated as a yellow light; all others are designated as green lights. Thirdly: Red and yellow light standard for negative comments: In the negative emotional comments of the posts by the influencers, the total proportion of the specific risk type will be determined based on the proportion of the number of risk tags identified (political, legal, moral, and commercial risks). If it is greater than 5%, it is a red light (i.e., the first risk threshold); if it is greater than 1% (i.e., the second risk threshold), it is a yellow light; and if it is otherwise (i.e., the third risk threshold), it is a green light. Further, by the whole network negative information query tool, the encyclopedia information query tool, and the full quantity library document query tool, the conditions are input to the risk assessment model node after the arrangement of the prompt (prompt word). According to whether the above three types of risk assessment standards reach the specific political, legal, moral, commercial, and other public opinion risks, the specific risk label, risk type (red, yellow, and green light), influencer name, original post, comment, and other information are output to judge the risk situation of the influencer. Among them, the intelligent recommendation model marks the recommended content according to the risk assessment result, takes the highest correlation for the level and reason, merges the details, and outputs the recommended content containing the risk identification.

[0036] As a preferred embodiment, the method further comprises the following steps: Step S7: Based on the risk assessment model, the content of the domain process routed by the user natural language request is risk assessed, and the risk assessment result is evaluated for the score calculation: Step S8: Evaluation score = (CPE / 5 + CPM / 5 + completion rate / 5 + interaction amount / 5 + browsing volume / 5) / 4x25 + cooperation competitor matching degree / 4x25 + fan portrait matching degree / 4x25 + person setup consistency / 4x25. Among them, CPE represents the cost per participation, CPM represents the cost per thousand impressions, and if a certain dimension data is missing, the remaining dimension weight is proportionally allocated.

[0037] For example, the present embodiment provides an evaluation summary method, as shown in Figure 5 The method comprises the following steps: According to the preconditions of the user input delivery platform, influencer type, batch specified influencer, fan portrait, influencer setup, influencer area, target CPE, CPM, play volume, interaction amount, and browsing volume range value; Firstly, if the CPE, CPM, play volume, interaction amount, and browsing volume conditions are specified, the specific index value will be matched according to the influencer list tool. If matched, the score is obtained. Secondly, if the cooperation competitor is specified, the corresponding brand will be matched according to the brand query tool. If not matched, the score is obtained. Thirdly, if the fan portrait is specified, the fan portrait query tool will be used to query the age distribution, gender distribution, and city level distribution. If the specified portrait condition meets the standard, the score is obtained. Fourthly, if the person setup keyword or person setup exclusion word is specified, the person setup query tool will be used to query the person setup of the influencer. The person setup keyword and the person setup exclusion word are input to the intelligent recommendation model to judge whether the user's desired person setup is consistent with the expected one. If consistent, the score is obtained. Further, the four types of evaluation dimension score conditions and specific index values are input to the evaluation summary node, and a summary report and a final evaluation reference score are output from the CPE, CPM specific performance, whether a cooperative competitor exists, fan image compliance, and consistency of the person set.

[0038] It can be understood that, based on the risk label and the sensitive word library, the risks in metaphors, irony, and network language can be recognized, and the false negative rate is reduced.

[0039] In the embodiment, the Moe method is used to process different field processes by using model nodes corresponding to the specialized field, including: The account analysis field: the intelligent recommendation model is used for content element extraction; The risk assessment field: the risk assessment model is used to process text and image risk information; The Moe represents a mixed expert mode.

[0040] It can be understood that, by using the Moe (mixed expert mode) concept, the dynamic routing mechanism based on field recognition + modular intelligent recommendation model node arrangement (workflow engine coordination), different tasks (account analysis, explosive article imitation, talent recommendation, risk assessment) are processed by intelligent recommendation model nodes specialized in the field, and the illusion rate is reduced.

[0041] In the embodiment, by preprocessing the obtained account multidimensional data, the practicability and reliability of the data can be effectively improved, and the requirement for data computing performance is effectively reduced; the intelligent recommendation model is trained based on the preprocessed account multidimensional data, so that the intelligent recommendation model can perform end-to-end deep analysis and representation learning on the original text, to further improve the accuracy of the text analysis task (such as topic classification, intent understanding, etc.), thereby effectively performing deep semantic understanding, context association and reasoning ability processing on the account multidimensional data, to significantly improve the accuracy and semantic understanding depth of text analysis and recommendation / evaluation.

[0042] By the method of the application, the accuracy and semantic understanding depth of text analysis and recommendation / evaluation are significantly improved, the intelligent recommendation model node driven by the preset Prompt (prompt word) is used to replace the traditional algorithm (such as rule engine, classification model), and the intelligent recommendation model is used for deep semantic understanding, context association and reasoning ability processing of multidimensional account data.

[0043] By using the Moe (mixed expert mode) concept, the dynamic routing mechanism based on field recognition + modular intelligent recommendation model node arrangement (workflow engine coordination), different tasks (account analysis, explosive article imitation, talent recommendation, risk assessment) are processed by intelligent recommendation model nodes specialized in the field, and the illusion rate is reduced.

[0044] Based on the sensitive word library, the risks in metaphors, irony and network language can be recognized, and the false negative rate is reduced.

[0045] AI replaces manual work, greatly reducing the participation of human labor costs.

[0046] Compared with the prior art: relying on manual code table, statistical features or shallow ML model, the accuracy is limited when dealing with complex and unstructured text, it is difficult to capture deep semantics and implicit information, the present application significantly improves the analysis accuracy and depth of the recommendation / evaluation task through the intelligent recommendation model core node.

[0047] The above only describes the embodiments of the present application, and does not limit the patent scope of the present application, any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A KOL-based intelligent recommendation method, characterized in that, The method comprises the following steps: Obtain the account multi-dimensional data of the user, and preprocess the account multi-dimensional data; Based on the preprocessed account multi-dimensional data, the pre-set intelligent recommendation model is subjected to deep analysis and representation learning training to obtain a trained intelligent recommendation model; Obtain the natural language data input by the user, and use the intelligent recommendation model to perform domain recognition on the natural language data to determine the domain process of the user natural language request routing; The intelligent recommendation model performs content recommendation in the corresponding domain for the user according to the domain process of the user natural language request routing. 2.The KOL-based intelligent recommendation method of claim 1, wherein, The preprocessing process of the account multi-dimensional data comprises: The obtained account multi-dimensional data comprises account basic data, article data and comment data; The account basic data, article data and comment data are subjected to sentiment analysis, text classification, entity recognition, interaction quantity merging, persona identification and fan portrait construction processing. 3.The KOL-based intelligent recommendation method of claim 2, wherein, The persona identification process of the account basic data, article data and comment data comprises: Based on the intelligent recommendation model, the account basic data, article data and comment data are subjected to theme extraction, style analysis and value inclination identification to generate a persona file containing several core labels. 4.The KOL-based intelligent recommendation method of claim 1, wherein, The domain recognition process of the natural language data using the intelligent recommendation model comprises: If the natural language request input by the user involves single social media account analysis, it is routed to the account analysis domain; If the natural language request input by the user involves content generation or script writing, it is routed to the viral article imitation domain; If the natural language request input by the user involves KOL recommendation and persona analysis, it is routed to the influencer recommendation - persona domain; If the natural language request input by the user involves KOL recommendation and does not require persona analysis, it is routed to the influencer recommendation - no persona domain; If the natural language request input by the user involves negative evaluation or public opinion analysis, it is routed to the risk assessment domain; The domain process comprises account analysis, viral article imitation, influencer recommendation and risk assessment. 5.The KOL-based intelligent recommendation method according to claim 4, characterized in that, The content recommendation process in the corresponding domain for the user by the intelligent recommendation model according to the domain process of the user natural language request routing comprises: Influencer recommendation - persona domain: Use a pre-set persona query tool to obtain influencer pre-set persona data; Use the intelligent recommendation model to identify the persona demand in the natural language data input by the user, combine the pre-set persona data, calculate the semantic similarity, and set a similarity threshold value according to the semantic similarity, and when the value of the semantic similarity is greater than or equal to the similarity threshold value, the corresponding word group is taken as a recommendation condition; Query the pre-set influencer database based on the recommendation condition, and output an influencer list containing the recommendation reasons through the intelligent recommendation model. 6.The KOL-based intelligent recommendation method according to claim 5, characterized in that, The process of calculating the semantic similarity further comprises: Perform synonym expansion on the persona demand in the natural language data input by the user, supplement related words through the WordNet word library, and then calculate the similarity.

7. The KOL-based intelligent recommendation method according to claim 1, wherein, The method further comprises the following steps: Based on the preprocessed account multi-dimensional data, the pre-set risk assessment model is subjected to risk assessment training to obtain a trained risk assessment model; The risk assessment model performs risk assessment on the content of the domain process routed by the user natural language request, the intelligent recommendation model performs marking processing on the recommended content according to the risk assessment result, and outputs the recommended content containing risk identification. 8.The KOL-based intelligent recommendation method of claim 7, wherein, The process of the risk assessment model performing risk assessment on the content of the domain process routed by the user natural language request includes: Adopting a preset encyclopedia information query tool to perform risk assessment on the to-be-processed information recognized by the user input natural language; If the recognized to-be-processed information is a post published by a celebrity, setting a risk label and a sensitive word library according to the content published by the celebrity, and setting a first risk threshold, a second risk threshold and a third risk threshold according to the risk label and the sensitive word library; If the content published by the celebrity is within the first risk threshold, it is marked as a red light risk; if the content published by the celebrity is within the second risk threshold, it is marked as a yellow light risk; and if the content published by the celebrity is within the third risk threshold, it is marked as a green light risk. 9.The KOL-based intelligent recommendation method of claim 8, wherein, The method further includes the following steps: Based on the risk assessment model performing risk assessment on the content of the domain process routed by the user natural language request, performing evaluation score calculation on the risk assessment result: Evaluation score = (CPE / 5 + CPM / 5 + complete play rate / 5 + interaction amount / 5 + browsing volume / 5) / 4*25 + cooperation competitor matching degree / 4*25 + fan portrait matching degree / 4*25 + person set consistency / 4*25; Wherein, CPE represents pay-per-participation, CPM represents cost per thousand impressions, and if a certain dimension data is missing, the remaining dimension weights are proportionally allocated.

10. The KOL-based intelligent recommendation method according to any one of claims 1-9, characterized in that, Adopting the Moe method to process different domain processes by using model nodes corresponding to specialized fields, including: Account analysis field: adopting an intelligent recommendation model to extract content elements; Risk assessment field: adopting a risk assessment model to process text and image risk information; Wherein, the Moe represents a hybrid expert mode.

Images