AI large model-based arrival person screening method and device, medium and equipment

Abstract

The invention discloses an AI large model-based arrival person screening method and device, a medium and equipment, and belongs to the field of screening, and the method comprises the steps of firstly obtaining screening conditions input by a user, including text content, image data and qualification information requirements, and to-be-screened arrival person data; next, performing word segmentation, keyword extraction and semantic matching on the text content of the person arrival data by utilizing a preset semantic analysis model in combination with BiLSTM, CRF and LDA technologies, and outputting a semantic score; meanwhile, note styles are recognized through the text classification model, logic judgment is conducted, and styles and logic verification scores are obtained. In addition, element detection and style verification are carried out on the image data through the multi-modal recognition model, and image matching scores are output; the qualification scoring module calculates qualification scores according to a preset weight formula. And finally, fusing the multi-dimensional scores to generate a comprehensive score, and carrying out accurate screening on the arriving persons according to the comprehensive score.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of screening, in particular to a method and device for screening talents based on an AI large model, a medium and equipment. BACKGROUND

[0002] In the field of digital marketing, talent promotion has become an important way for brands to reach target users. With the continuous development of social media platforms, the number of talent accounts and note content has grown rapidly, and traditional talent screening technical solutions have gradually exposed significant limitations:

[0003] On the one hand, the manual operation process is tedious and the technical implementation path is fixed. The existing screening work mainly relies on manual linear search, and the staff needs to obtain the talent list from various channels and complete information collection in the mode of "account ID input → jump to homepage → browse notes one by one". When facing a large number of candidate accounts, each account needs to consume several minutes, and there is a lack of automatic processing mechanism, making it difficult to meet the high-frequency and multi-dimensional screening needs.

[0004] On the other hand, the evaluation method lacks objectivity. The evaluation of talent notes and talents themselves completely depends on manual subjective judgment. Different staff may have different judgments on the importance weight of the same note whether it meets the promotion needs, the tone of the talent account, etc. due to individual experience and preference differences, resulting in a lack of uniform standards for screening results and difficulty in ensuring accuracy.

[0005] In addition, the problem of low screening efficiency is prominent. Due to the limitations of manual operation, when facing hundreds or even thousands of talents, a screening may take several days or even weeks, and as the number of talents and notes grows, the screening difficulty increases exponentially, making it impossible to meet the needs of rapid start and efficient promotion of marketing projects. At the same time, long-term high-intensity work is easy to make workers tired, further reducing accuracy and efficiency.

[0006] These deficiencies result in the inability of existing technologies to accurately and efficiently screen talents in large-scale information. SUMMARY

[0007] The present application provides a talent screening method, device, medium and equipment based on an AI large model to solve the problem of inability to accurately and efficiently screen talents in large-scale information in the prior art.

[0008] In a first aspect, the present application provides a talent screening method based on an AI large model, comprising:

[0009] obtaining a user input screening condition and a to-be-screened talent data, the screening condition including a text content requirement, an image data requirement, and a qualification information requirement;

[0010] The preset semantic analysis model is constructed according to a bidirectional long short-term memory network combined with a conditional random field, and is constructed by extracting keywords through an implicit Dirichlet distribution and calculating historical user text data through cosine similarity.

[0011] The preset text classification model is used for note style type identification and logical judgment on the to-be-screened talent data, to obtain various style scores and various logical verification scores.

[0012] The preset multi-modal recognition model is used for element detection and style verification on the to-be-screened talent data and image data requirements, to output various image matching scores.

[0013] Based on the to-be-screened talent data, a preset weight formula is used to calculate various qualification scores, and the weight formula is adjusted according to the weight configuration of the qualification information requirements.

[0014] The various semantic scores, various style scores, various logical verification scores, various image matching scores and various qualification scores are fused to generate various comprehensive scores.

[0015] According to the various comprehensive scores, the talents in the to-be-screened talent data are screened.

[0016] The present application realizes efficient, accurate and objective talent screening by constructing a multi-dimensional talent screening method based on an AI large model. The present application first obtains the screening conditions input by the user and the to-be-screened talent data, including text content, image data and qualification information requirements. Then, the preset semantic analysis model is used to extract keywords and perform semantic matching on the text by combining BiLSTM, CRF and LDA technology, to output semantic scores; at the same time, the text classification model is used to identify the note style and verify the logical structure to obtain style and logical scores. In addition, the multi-modal recognition model performs element detection and style verification on the image based on YOLOv5 and DeepLabv3+ to output image scores. The present application also calculates the qualification scores according to the preset weight formula based on the qualification information requirements. Finally, the above-mentioned dimensional scores are fused to generate comprehensive scores, and the talents meeting the user's needs are accurately screened. This process not only significantly improves the screening efficiency and reduces manual operation, but also ensures the accuracy and adaptability of the screening results through multi-modal fusion and dynamic weight adjustment, effectively solving the limitations of traditional screening methods and the problem that the existing technology cannot accurately and efficiently screen talents in a large amount of information.

[0017] Further, the preset semantic analysis model is used to match and perform semantic analysis on the to-be-screened talent data and text content requirements, to output various semantic scores, specifically:

[0018] inputting the text content requirement and the to-be-screened expert data into the semantic analysis model, so that the semantic analysis model performs word segmentation processing on the to-be-screened expert data according to a bidirectional long short-term memory network (BiLSTM) and in combination with a conditional random field (CRF), to obtain each first word segmentation result;

[0019] extracting each second word segmentation result and a weight corresponding to each second word segmentation result in the each first word segmentation result through a latent Dirichlet allocation (LDA) model;

[0020] calculating a cosine similarity between the each second word segmentation result and the text content requirement according to the each second word segmentation result and the weight corresponding to the each second word segmentation result, and outputting each semantic score.

[0021] The preset semantic analysis model is used to accurately match and analyze the semantic relationship between the to-be-screened expert data and the text content requirement, so as to output each semantic score. Specifically, the text content requirement and the to-be-screened expert data are input into the semantic analysis model, and a bidirectional long short-term memory network (BiLSTM) is used in combination with a conditional random field (CRF) to perform word segmentation processing, so as to obtain accurate first word segmentation results. Then, a latent Dirichlet allocation (LDA) model is used to extract key second word segmentation results and their weights, and these key word segmentation results can effectively reflect the core semantics of the text. Finally, based on the key word segmentation results and their weights, the cosine similarity between the key word segmentation results and the text content requirement is calculated, so as to obtain each semantic score. This process not only accurately identifies the keywords of the text content, but also quantitatively evaluates the semantic similarity, so as to ensure that the screening result is highly matched in the semantic level, and the accuracy and efficiency of the expert screening are significantly improved.

[0022] Further, the to-be-screened expert data is input into the text classification model, so that the text classification model identifies the style of the to-be-screened expert data through a BERT model, analyzes the logical structure of the to-be-screened expert data through a syntax tree, and outputs the text style and the text logic.

[0023] inputting the to-be-screened expert data into the text classification model, so that the text classification model identifies the style of the to-be-screened expert data through a BERT model, analyzes the logical structure of the to-be-screened expert data through a syntax tree, and outputs the text style and the text logic.

[0024] comparing the text style and the text logic with the text content requirement in the screening condition, and outputting each text style score and each text logic score.

[0025] The application realizes accurate identification and evaluation of the style type and logical structure of the star note by a preset text classification model, so as to obtain a style score and a logical verification score. Specifically, the to-be-screened star data is input into the text classification model, the style type of the note is identified by using the BERT model, and the logical structure of the note is analyzed by using the syntax tree, and the results of the text style and the text logic are output. Then, the results are compared with the text content requirements in the screening conditions input by the user, and the text style score and the text logic score are calculated. This process not only accurately identifies whether the style of the note meets the requirements set by the user, but also ensures the coherence and integrity of the note content through the verification of the logical structure, thereby significantly improving the accuracy and reliability of the star screening.

[0026] Further, based on the preset multi-modal recognition model, the to-be-screened star data and the image data requirements are element detected and style verified, and each image matching score is output, specifically:

[0027] The image data requirements and the to-be-screened star data are input into the multi-modal recognition model, so that the multi-modal recognition model performs element detection on the to-be-screened star data according to a target detection algorithm, identifies products, scenes and characters, and determines the image elements of the to-be-screened star data;

[0028] Based on a semantic segmentation algorithm, the to-be-screened star data is style verified, the color distribution and composition style of the to-be-screened star data are analyzed, and the image style of the to-be-screened star data is determined;

[0029] The to-be-screened star data is converted to an HSV color space, a color histogram is calculated, the color keynote of the color histogram is judged by using the Bhattacharyya distance, and the image color of the to-be-screened star data is determined;

[0030] The image elements, the image style and the image color are compared with the image data requirements in the screening conditions, and each image matching score is output.

[0031] The present application realizes accurate element detection and style verification of the influencer image data through the preset multi-modal recognition model, thereby outputting various image matching scores. Specifically, the image data requirements and the influencer data to be screened are input into the multi-modal recognition model, the target detection algorithm (such as YOLOv5) is used to detect the elements of the influencer data, and the key elements such as products, scenes and characters in the image are accurately identified. At the same time, based on the semantic segmentation algorithm (such as DeepLabv3+), the image is verified for style, the color distribution and composition style are analyzed, and the overall style of the image is determined. Further, the image data is converted to HSV color space, the color histogram is calculated, and the color keynote is judged by Bhattacharyya distance, so as to determine the color characteristics of the image. Finally, the detected image elements, style and color are compared with the image data requirements set by the user, and various image matching scores are output. This process not only accurately identifies the key elements in the image, but also ensures that the image is highly matched with the user's demand through the verification of style and color, significantly improving the accuracy and reliability of influencer screening.

[0032] Further, the various qualification scores are calculated by a preset weight formula based on the influencer data to be screened, specifically:

[0033] The fan portrait, field verticality, historical conversion rate and interaction rate in the influencer data to be screened are extracted.

[0034] According to the qualification information requirements and the preset weight formula, the influencer data to be screened is weighted and summed to calculate the various qualification scores.

[0035] The present application realizes accurate quantitative evaluation of the influencer qualification through the preset weight formula, thereby calculating the various qualification scores. Specifically, first, the key qualification information in the influencer data to be screened is extracted, including fan portrait, field verticality, historical conversion rate and interaction rate. These indicators can comprehensively reflect the influence and promotion effect of the influencer in a specific field. Then, according to the qualification information requirements input by the user and the preset weight formula, the qualification data is weighted and summed to calculate the various qualification scores. This process not only adjusts the weight of each qualification indicator according to the specific needs of the user, but also ensures the objectivity and consistency of the qualification evaluation through quantitative methods. In this way, the present application can accurately screen high-quality influencers that meet the user's needs, significantly improving the accuracy and reliability of influencer screening.

[0036] Further, the various semantic scores, various style scores, various logic verification scores, various image matching scores and various qualification scores are fused to generate various comprehensive scores, and the influencer in the influencer data to be screened is screened according to the various comprehensive scores, specifically:

[0037] fusing the semantic score, the style score and the logic verification score into a text matching score;

[0038] weighting and summing the text matching score, the image matching score and the qualification score according to a preset weight distribution to obtain a comprehensive score of each;

[0039] screening the influencer in the to-be-screened influencer data based on the comprehensive score.

[0040] The present application realizes comprehensive evaluation and accurate screening of influencer data by fusing multi-dimensional scores. Specifically, first, the semantic score, the style score and the logic verification score are fused into a text matching score, which comprehensively considers the semantic relevance, style consistency and logic integrity of the text content, ensuring comprehensive evaluation in the text dimension. Subsequently, the text matching score, the image matching score and the qualification score are weighted and summed according to a preset weight distribution to obtain a comprehensive score of each influencer. This weighting and summing process considers the importance of different dimensions, ensuring the scientificity and rationality of the comprehensive score. Finally, the influencer in the to-be-screened influencer data is screened based on the comprehensive score, ensuring that the screening result not only meets the requirements of the user in multiple dimensions such as text, image and qualification, but also accurately identifies the best influencer through the level of the comprehensive score. This process significantly improves the accuracy and efficiency of influencer screening and provides more reliable candidate recommendations for marketing activities.

[0041] In a second aspect, the present application provides an AI large model-based influencer screening device. The AI large model-based influencer screening device comprises:

[0042] An acquisition module acquires a screening condition input by a user and to-be-screened influencer data, wherein the screening condition comprises a text content requirement, an image data requirement and a qualification information requirement;

[0043] A semantic scoring module is configured to match and analyze the to-be-screened influencer data with the text content requirement based on a preset semantic analysis model, and output a semantic score of each; the semantic analysis model is constructed by calculating historical user text data through an implicit Dirichlet distribution and cosine similarity based on a bidirectional long short-term memory network combined with a conditional random field;

[0044] A text classification module is configured to identify a note style type and perform logic judgment on the to-be-screened influencer data based on a preset text classification model, and obtain a style score and a logic verification score of each;

[0045] A modality recognition module is configured to detect elements and verify styles of the to-be-screened influencer data based on a preset multi-modal recognition model, and output an image matching score of each;

[0046] The qualification scoring module is configured to calculate each qualification score based on the to-be-screened talent data by using a preset weight formula, and the weight formula is configured and adjusted according to the qualification information requirements.

[0047] The screening module is configured to fuse each semantic score, each style score, each logic verification score, each image matching score and each qualification score to generate each comprehensive score, and screen the talents in the to-be-screened talent data according to the comprehensive scores.

[0048] The talent screening device based on the AI large model realizes the full-process automation and intelligent processing from data acquisition to final screening through modular design. Specifically, the acquisition module first collects the screening conditions input by the user and the to-be-screened talent data, ensuring that the screening process is based on clear and specific requirements. The semantic scoring module uses an advanced semantic analysis model to accurately match the text content and analyze its semantics, outputting semantic scores and providing a scientific basis for text dimension evaluation. The text classification module further identifies the style of the notes and judges the logic structure, outputting style and logic scores and enhancing the comprehensiveness of text evaluation. The modality recognition module uses multi-modal recognition technology to detect elements and verify the style of image data, outputting image matching scores and expanding the screening dimension to the visual level. The qualification scoring module calculates the qualification scores according to the preset weight formula, quantitatively evaluates the qualification information of the talents, and ensures that the screening considers the comprehensive strength of the talents. Finally, the screening module fuses the scores of each dimension to generate comprehensive scores, and screens the talents based on the scores, ensuring that the screening results are accurate and efficient. The modular design of this device not only improves the screening efficiency, but also significantly improves the accuracy and reliability of talent screening through multi-dimensional comprehensive evaluation, providing better talent recommendations for marketing activities.

[0049] Further, the to-be-screened talent data is matched with the text content requirements and semantic analysis is performed based on the preset semantic analysis model, and each semantic score is output, specifically:

[0050] The text content requirements and the to-be-screened talent data are input into the semantic analysis model, so that the semantic analysis model performs word segmentation processing on the to-be-screened talent data according to a bidirectional long short-term memory network and in combination with a conditional random field, and each first word segmentation is obtained.

[0051] Each second word segmentation in the each first word segmentation and a weight corresponding to each second word segmentation are extracted by using an implicit Dirichlet distribution model;

[0052] The cosine similarity between each second word segmentation and the text content requirements is calculated according to each second word segmentation and the weight corresponding to each second word segmentation, and each semantic score is output.

[0053] The application realizes accurate matching and semantic analysis of the to-be-screened talent data and the text content requirements through a preset semantic analysis model, thereby outputting various semantic scores. Specifically, the text content requirements and the to-be-screened talent data are input into the semantic analysis model, and a bidirectional long short-term memory network (BiLSTM) combined with a conditional random field (CRF) is used for word segmentation processing to obtain an accurate first word segmentation result. Then, a latent Dirichlet allocation (LDA) model is used to extract key second word segmentation and weights, and the key word segmentation can effectively reflect the core semantics of the text. Finally, based on the key word segmentation and the weights, the cosine similarity between them and the text content requirements is calculated to obtain various semantic scores. This process not only accurately identifies the keywords of the text content, but also ensures that the screening results are highly matched at the semantic level through quantitative evaluation of semantic similarity, thereby significantly improving the accuracy and efficiency of talent screening.

[0054] In a third aspect, the application provides a computer-readable storage medium comprising a stored computer program, wherein the computer program, when executed, controls a device in which the computer-readable storage medium is located to perform an AI large model-based talent screening method as described. The beneficial effects are the same as those of the AI large model-based talent screening method provided in the first aspect of the application.

[0055] In a fourth aspect, the application provides a terminal device comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements any of the AI large model-based talent screening methods described in the first aspect when executing the computer program. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 FIG. 1 is a flowchart of an embodiment of the AI large model-based talent screening method provided by the application;

[0057] Figure 2 FIG. 2 is a structural diagram of an embodiment of the AI large model-based talent screening device provided by the application. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only some of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0059] Embodiment one

[0060] Please refer to Figure 1In order to solve the problem that the prior art cannot accurately and efficiently screen talents under large-scale information, the application provides a talent screening method based on an AI large model, which comprises steps S01-S06.

[0061] S01: obtaining screening conditions input by a user and talent data to be screened, wherein the screening conditions comprise text content requirements, image data requirements and qualification information requirements.

[0062] As a preferred embodiment of the present embodiment, the screening conditions input by the user and the talent data to be screened comprise text content requirements, image data requirements and qualification information requirements, and specifically:

[0063] The user completes the structured input of the screening conditions through an interactive interface, and meanwhile, the application automatically associates the full data of the talent to be screened to form a complete analysis data set. The screening conditions input by the user need to clearly include three types of core requirements: the text content requirements need to include theme keywords (such as "children's safety education" and "parent-child picture book recommendation"), style preferences (such as "practical and scientific popular science style" and "daily vlog style") and logical structure rules (such as "question introduction→solution→case explanation"); the image data requirements need to specify the required elements (such as "parent-child interaction scene" and "book cover close-up"), picture style (such as "bright and warm color tone" and "indoor home scene") and composition standards (such as "personality proportion≥40%" and "no watermark interference"); the qualification information requirements need to limit the account attributes of the talent, including field verticality (such as "education content proportion≥80%"), fan portrait (such as "3-6 year-old children's parents proportion≥60%"), interaction indicators (such as "average like amount≥500") and historical conversion data (such as "past promoted commodity click rate≥3%").

[0064] The talent data to be screened is obtained in two ways: one is that the user imports a list of talent account IDs in an Excel file in batches, automatically connects to a talent database, and pulls the historical note data (including text content, picture materials and publishing time) of the corresponding account; the other is that the application directly synchronizes the candidate talent list from the MCN agency interface, and automatically associates the qualification archives (fan details, field labels and historical cooperation records) thereof; finally, the application binds the user screening conditions and the talent data to be screened, and constructs an "talent-note-qualification" associated data set, thereby providing standardized input for subsequent multi-dimensional matching of the AI model.

[0065] S02: based on a preset semantic analysis model, matching and performing semantic analysis on the talent data to be screened and the text content requirements, and outputting respective semantic scores; the semantic analysis model is constructed by calculating historical user text data through an implicit Dirichlet distribution according to a bidirectional long short-term memory network combined with a conditional random field and a cosine similarity.

[0066] As a preferred embodiment of the present embodiment, the preset semantic analysis model is used to match the to-be-screened talent data with the text content requirements and perform semantic analysis, and output respective semantic scores, specifically:

[0067] Firstly, the preset semantic analysis model (which is constructed based on a bidirectional long short-term memory network (BiLSTM) combined with a conditional random field (CRF), and is optimized by a cosine similarity calculation logic based on a latent Dirichlet allocation (LDA) model to extract keywords) is called by the present application. The historical note text (such as product evaluation, scene experience, etc. published by the talent) in the to-be-screened talent data is input into the model, and the text content requirements set by the user (such as the theme keyword “infant food safety” and the core semantic “no additive recommendation”) are also input into the model.

[0068] The model first performs word segmentation on the note text: the BiLSTM extracts the context features of the text, the CRF model is used to mark the word boundaries, the continuous text is split into independent words (first word segmentation) such as “infant”, “food”, “no additive”, etc. (such as splitting “this infant food has no additives, suitable for babies over 6 months” into ["this", "infant", "food", "no", "additives", ",", "suitable", "for babies over 6 months"]), and filtering out stop words such as “this” and “,”. The second step is to extract the core keywords (second word segmentation) by the LDA model: based on the text theme distribution probability, the words with the highest correlation to the theme are selected (such as extracting “infant food”, “no additives”, and “over 6 months” as core keywords from the above word segmentation result), and converting them into a TF-IDF vector (the vector dimension corresponds to the keyword weight, such as “no additives” having a higher weight in the recommendation text);

[0069] Finally, the model calculates the cosine similarity between the note keyword vector and the text content requirement keyword vector set by the user: the formula is: to calculate the semantic overlap (such as the note keyword being “infant food no additives” and the user requirement keyword being “infant food safety no additives”, and the similarity being 0.85). According to the similarity result, the semantic score is output: following the linear mapping rule, a similarity of 1.0 corresponds to 60 points (full score), a similarity of 0.7 corresponds to 42 points, and a similarity lower than 0.7 corresponds to 0 points (such as the similarity of 0.85 corresponding to 51 points). Each note of each to-be-screened talent generates an independent semantic score, and the overall semantic performance data of the talent is finally summarized, which provides a basis for subsequent text matching score calculation.

[0070] Further, the semantic analysis model is constructed according to a bidirectional long short-term memory network combined with a conditional random field, and by extracting keywords through a latent Dirichlet allocation and calculating a cosine similarity of historical user text data.

[0071] In the embodiment, the semantic analysis model is a deep learning model fused with multiple natural language processing algorithms and trained and optimized based on historical user text data, and the construction process is specifically as follows: a bidirectional long short-term memory network (BiLSTM) combined with a conditional random field (CRF) is used as a model basic framework: the BiLSTM is responsible for processing bidirectional context information of a text sequence, and extracts semantic features (such as an association relationship between “infant complementary food” and “no additive”) of words through a forward and backward hidden layer, and the CRF optimizes the annotation accuracy of word boundaries based on these features, and ensures that the segmentation result conforms to the grammar logic (for example, avoiding missegmentation of “no additive ingredients” into “no / additive / ingredient”). In the keyword extraction link, the model embeds a latent Dirichlet allocation (LDA) algorithm, performs topic modeling on historical user text data (such as topic descriptions in past promotion needs, high-quality influencer note texts, etc.), calculates the probability distribution of words in different topics, selects keywords (such as “6-month-old baby” and “no additive complementary food”) with the highest association degree with the core topic from “suitable for 6-month-old baby no additive complementary food recommendation”, and assigns weights. At the same time, the model introduces cosine similarity calculation logic, compares the keyword vectors of “user text requirements” and “influencer note text” in historical data, and iteratively optimizes the hidden layer parameters of the BiLSTM, the transition probability matrix of the CRF, and the topic distribution threshold of the LDA, and finally forms an analysis model that can accurately capture semantic associations. The model can not only handle ambiguous text (such as the semantic equivalence of “safe complementary food” and “no additive complementary food”), but also quantify the matching degree, providing reliable algorithm support for subsequent semantic score calculation.

[0072] S03: Based on the preset text classification model, the note style type identification and logical judgment of the to-be-screened influencer data are performed to obtain each style score and each logical verification score.

[0073] As a preferred embodiment of the embodiment, based on the preset text classification model, the note style type identification and logical judgment of the to-be-screened influencer data are performed to obtain each style score and each logical verification score, and the specific process is as follows:

[0074] Firstly, the preset text classification model (which is based on the BERT pre-training model and trained by historical note text samples labeled with style tags such as "evaluation body", "story body", and "practical popular science style") is called to input the historical note text (such as product recommendation copy and experience sharing published by the influencer) in the to-be-screened influencer data into the model. The model extracts the style features of the text (such as the "evaluation body" common "parameter comparison → advantage and disadvantage analysis → summary and recommendation" sentence structure, and the narrative vocabulary and emotional expression commonly used in the "story body") through deep semantic coding, and matches them with the style preference in the user-set text content requirements (such as "professional evaluation style is required"); if the recognition result is consistent with the set style, the note gets 20 points (the maximum score for style); if it is not consistent, it gets 0 points;

[0075] At the same time, the model analyzes the logical structure of the note text through a syntax tree analysis tool (such as spaCy) to extract key logical chains such as "premise → conclusion" and "problem → solution". For example, according to the user's requirement of "product introduction → use experience → summary", the model will analyze whether the note contains necessary links such as "product ingredient description", "skin state description after use", and "summary of suitable people": if the logical chain is complete and the order meets the requirements, it gets 20 points (the maximum score for logical verification); if there are missing links (such as only introducing the product without mentioning the use experience) or logical confusion (such as summarizing first and then introducing the product), it gets 0 points;

[0076] Each note of each to-be-screened influencer will generate an independent style score and a logical verification score, and the final score will be the overall performance of the influencer in style adaptation and logical integrity, which will provide the basis for the calculation of the subsequent text matching score.

[0077] S04: Based on the preset multi-modal recognition model, the to-be-screened influencer data and the image data requirements are detected and verified for elements and style, and each image matching score is output.

[0078] As a preferred embodiment of the present embodiment, the to-be-screened influencer data and the image data requirements are detected and verified for elements and style based on the preset multi-modal recognition model, and each image matching score is output, specifically:

[0079] Firstly, the preset multi-modal recognition model (which integrates the YOLOv5 target detection algorithm and the DeepLabv3+ semantic segmentation algorithm and is trained by historical influencer note pictures labeled with element categories, scene style, and color tone) is called to input the historical note pictures (such as product display pictures and scene real picture published by the influencer) in the to-be-screened influencer data into the model, and the user-set image data requirements (such as "need to contain outdoor backpack close-up, mountain and forest scene, and warm color tone proportion ≥60%") are input into the model;

[0080] The first step of the model is to perform pixel-level scanning on the image using the YOLOv5 algorithm, identify and label the core elements in the image (such as "backpack", "mountain forest", "person", etc.), and calculate the confidence of each element (such as "backpack" confidence 0.95, "mountain forest" confidence 0.92). Compare the detection results with the required element list of the image data: if the element categories are completely matched and the confidence is ≥0.7, score 30 points (element detection score full marks); if there are missing items (such as "mountain forest" not detected) or low-confidence elements (such as "backpack" confidence 0.6), deduct points in proportion (such as deduct 15 points for one missing item);

[0081] The second step is to perform semantic segmentation on the image using the DeepLabv3+ algorithm to analyze the scene type (such as "outdoor" "indoor") and composition style (such as "close-up" "panoramic scene"). If the scene type is consistent with the required "mountain forest scene" and the composition meets the requirement of "backpack proportion ≥40%", score 30 points (scene and composition score full marks, of which scene matching accounts for 15 points and composition compliance accounts for 15 points); if the scene is not consistent or the composition is not up to standard, the corresponding item scores 0 points;

[0082] The third step is to convert the image to HSV color space, calculate the histogram of hue, saturation, and brightness, and use the Bhattacharyya distance formula Where p and q are the histograms of the image and the target color respectively, and n is the number of bins in the histogram; judge the matching degree of the required "warm color proportion ≥60%": if the distance ≤0.3 (complete match), score 40 points (color base score full marks); if it is a critical value (such as warm color proportion 60%), score 28 points; if it is not matched, score 0 points;

[0083] Finally, add the element detection score, scene and composition score, and color base score to get the image matching score of a single image (full score 100 points). All historical note images of each person to be screened are scored according to this logic, and the scores are aggregated to provide a basis for subsequent content matching degree calculation.

[0084] S05: Based on the data of the person to be screened, calculate each qualification score through a preset weight formula, and the weight formula is adjusted according to the weight configuration of the qualification information requirements.

[0085] As a preferred embodiment of the present embodiment, based on the data of the person to be screened, calculate each qualification score through a preset weight formula, and the weight formula is adjusted according to the weight configuration of the qualification information requirements, specifically:

[0086] First, four core qualification indicators are extracted from the data of the to-be-screened KOLs: fan portrait matching degree (the overlap ratio of KOL fans and target users, such as "25-35 year-old female proportion" and "maternity product follower proportion"), field verticality (the relevance of KOL published content to the screening requirement field, such as "cosmetic note proportion" and "whether irrelevant content is published in the past three months"), historical conversion rate (the conversion efficiency of KOL past promotion, such as "product link click rate" and "order conversion rate"), and interaction rate (the user engagement of KOL notes, such as "like + comment + forward total quantity / fan quantity").

[0087] Subsequently, a preset weight formula is called: qualification score = w1 x fan portrait matching degree + w2 x field verticality + w3 x historical conversion rate + w4 x interaction rate. Wherein, the default weight parameters are w1 = 0.35, w2 = 0.3, w3 = 0.2, and w4 = 0.15 (based on industry average data training optimization), but dynamic adjustment according to user set qualification information requirements is supported. For example, if the user emphasizes "precise touch of target population", the weight of fan portrait matching degree can be increased (such as w1 adjustment to 0.45); if the promotion effect is emphasized, the weight of historical conversion rate can be increased (such as w3 adjustment to 0.3).

[0088] Each indicator is quantified in percentage (such as fan portrait matching degree 80% corresponding to 80 points, field verticality 90% corresponding to 90 points), and the final qualification score (full score 100 points) is obtained after substitution. For example, the four indicators of a KOL are 85 points, 90 points, 70 points, and 80 points, and the default weight calculation is: qualification score = 0.35 x 85 + 0.3 x 90 + 0.2 x 70 + 0.15 x 80 = 29.75 + 27 + 14 + 12 = 82.75 points. Each to-be-screened KOL corresponds to an independent qualification score, which is the core quantitative basis for evaluating whether the account qualification meets the promotion requirements.

[0089] S06: fuse the respective semantic scores, respective style scores, respective logic verification scores, respective image matching scores, and respective qualification scores to generate respective comprehensive scores; and screen KOLs in the to-be-screened KOL data according to the respective comprehensive scores.

[0090] As a preferred embodiment of the present embodiment, the respective semantic scores, respective style scores, respective logic verification scores, respective image matching scores, and respective qualification scores are fused to generate respective comprehensive scores; and KOLs in the to-be-screened KOL data are screened according to the respective comprehensive scores, specifically:

[0091] First, the score fusion calculation is performed: in the first step, the text correlation score is integrated into the text matching score, that is, the semantic score (full score 60 points), the style score (full score 20 points), and the logic verification score (full score 20 points) of each influencer note are summed to obtain the text matching score (full score 100 points) of a single note, and then the average value of all historical notes of the influencer is taken as the final text matching score. In the second step, the content matching degree is calculated, and the text and image dimensions are fused according to the preset weight formula "content matching degree = 0.6 x text matching score + 0.4 x image matching score" (where the image matching score takes the average score of the influencer's historical note pictures, with a full score of 100 points). In the third step, the comprehensive score is generated by fusing the content performance and account quality through the formula "comprehensive score = content matching degree x a + quality score x b", where the default weights a = 0.4 and b = 0.6 (supporting users to adjust according to promotion needs, such as increasing a to 0.5 if focusing on content quality);

[0092] For example, an influencer's text matching score is 85 points, the image matching score is 90 points, and the quality score is 82 points. According to the default weights, the content matching degree is 0.6 x 85 + 0.4 x 90 = 51 + 36 = 87 points, and the comprehensive score is 87 x 0.4 + 82 x 0.6 = 34.8 + 49.2 = 84 points.

[0093] In the screening stage, a double threshold is required: only influencers with a content matching degree ≥ 60 points and a quality score ≥ 55 points enter the candidate pool, and then the comprehensive score is sorted from high to low, and finally a screening list containing influencer account information, score details, and recommendation priority is output. Users can perform secondary screening based on this list (such as filtering by domain verticality, fan profile, etc.), or directly select the top N influencers as promotion partners to achieve a closed loop of the screening process.

[0094] To further illustrate the specific implementation of the present application, the specific process of the present application will be described in detail in the following with reference to a scenario of a mother and baby brand promoting new infant complementary food:

[0095] First, the data acquisition step is performed: the brand party clearly inputs the screening conditions, which require the text content to contain keywords such as "infant complementary food", "no additives", "6 months + applicable", etc., the style to be professional evaluation, and the logic structure to cover "ingredient analysis → safety verification → feeding suggestions"; the image data requires that it must contain "complementary food packaging close-up" and "baby eating scene" elements, the style is bright indoor light, close-up shooting, and the complementary food proportion is not less than 50%, the color tone is warm, and the yellow / orange proportion is ≥ 60%; the quality information requires that the domain verticality of the mother and baby content is ≥ 90%, the fan profile of the 3-6 year old children's parents is ≥ 70%, the interaction rate is ≥ 3%, and the historical conversion rate is ≥ 4%. At the same time, the data of 200 mother and baby field influencers to be screened is collected, including their note texts and pictures published in the last 6 months, fan profiles, and historical promotion data, etc.

[0096] Next, multi-dimensional analysis is carried out in turn: in the text analysis link, the preset semantic analysis model is used to match and analyze the semantics of the note text of the influencer and the text content requirements, the semantic score is output by word segmentation, keyword extraction and calculation of cosine similarity, for example, an influencer gets 55 points because the target keyword appears frequently in the note; with the help of the preset text classification model, the style type of the note is identified and the logic is judged, which meets the style of the professional evaluation body for 20 points, and the logic structure is complete for 20 points, and the text matching score is obtained by adding the three scores. In the image analysis link, based on the preset multi-modal recognition model, the elements detection and style verification of the influencer note picture and image data requirements are carried out, the elements meet the requirements for 30 points, the scene and composition meet the standards for 30 points, and the color tone matching for 40 points, and the image matching score is obtained. In the qualification evaluation link, the fan portrait matching degree, field verticality, historical conversion rate, interaction rate and other indicators are extracted from the to-be-screened influencer data, and the qualification score is calculated through the preset weight formula (the weight configuration can be adjusted according to the qualification information requirements).

[0097] Finally, score fusion and screening are carried out: the text matching score and the image matching score are fused into the content matching degree according to the weight ratio of 6:4; then the content matching degree and the qualification score are fused according to the weight ratio of 4:6 to generate the comprehensive score of each influencer. The influencer with content matching degree ≥60 points and qualification score ≥55 points is screened out, and the comprehensive score is sorted from high to low, and the screening result is finally output, and the brand party can select the appropriate influencer for promotion cooperation.

[0098] In summary, the multi-dimensional influencer screening method based on AI large model is constructed, and efficient, accurate and objective influencer screening is realized. First, the screening conditions input by the user and the to-be-screened influencer data are obtained, including text content, image data and qualification information requirements. Next, the preset semantic analysis model is used to extract keywords and match semantics from the text by combining BiLSTM, CRF and LDA technology, and output the semantic score; at the same time, the note style is identified and the logic structure is verified by the text classification model to obtain the style and logic scores. In addition, the multi-modal recognition model detects elements and verifies the style of the image based on YOLOv5 and DeepLabv3+ to output the image score. According to the qualification information requirements, the qualification score is calculated by the preset weight formula. Finally, the comprehensive score is generated by fusing the scores of each dimension to accurately screen the influencer who meets the user's needs. This process not only significantly improves the screening efficiency and reduces manual operation, but also ensures the accuracy and adaptability of the screening results through multi-modal fusion and dynamic weight adjustment, effectively solving the limitations of traditional screening methods and the problem that the existing technology cannot accurately and efficiently screen influencers in large-scale information.

[0099] Embodiment two

[0100] Reference is made to Figure 2 The AI large model-based talent screening device is provided for the embodiments of the present application.

[0101] In the present embodiment, the AI large model-based talent screening device comprises an acquisition module 10, a semantic scoring module 20, a text classification module 30, a modality recognition module 40, a qualification scoring module 50, and a screening module 60.

[0102] The acquisition module 10 acquires the screening conditions and the to-be-screened talent data input by a user, wherein the screening conditions comprise text content requirements, image data requirements, and qualification information requirements.

[0103] The semantic scoring module 20 is configured to match and perform semantic analysis on the to-be-screened talent data and the text content requirements based on a preset semantic analysis model, and output respective semantic scores; the semantic analysis model is constructed based on a bidirectional long short-term memory network combined with a conditional random field, and by extracting keywords through a hidden Dirichlet distribution and calculating historical user text data based on cosine similarity.

[0104] The text classification module 30 is configured to perform note style type recognition and logical judgment on the to-be-screened talent data based on a preset text classification model, and obtain respective style scores and respective logical verification scores.

[0105] The modality recognition module 40 is configured to perform element detection and style verification on the to-be-screened talent data and the image data requirements based on a preset multi-modal recognition model, and output respective image matching scores.

[0106] The qualification scoring module 50 is configured to calculate respective qualification scores based on the to-be-screened talent data by using a preset weight formula, wherein the weight formula is adjusted based on the weight configuration of the qualification information requirements.

[0107] The screening module 60 is configured to fuse the respective semantic scores, the respective style scores, the respective logical verification scores, the respective image matching scores, and the respective qualification scores, generate respective comprehensive scores, and screen the talents in the to-be-screened talent data based on the respective comprehensive scores.

[0108] For the convenience and brevity of description, the device embodiments of the present application include all the implementation manners of the AI large model-based talent screening method embodiments described above, which will not be described herein again.

[0109] Embodiment Three

[0110] The embodiment of the application provides a computer readable storage medium, the computer readable storage medium comprises a stored computer program, wherein the computer readable storage medium is controlled by the computer program to execute the AI large model based talent screening method when the computer program is running.

[0111] The AI large model based talent screening method can be stored in a computer readable storage medium if it is realized in the form of a software function unit and used as an independent product. Based on this understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. When the processor executes the computer program, the steps of the above-mentioned various method embodiments can be realized. The computer program includes computer program code, which can be in the form of source code, object code, an executable file or some intermediate form, etc. The computer readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a U disk, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal and a software distribution medium, etc.

[0112] Embodiment four

[0113] The embodiment provides a terminal device, comprising a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, and the processor executes the computer program to realize the AI large model based talent screening method according to any one of the embodiment one.

[0114] The above-mentioned specific embodiments further illustrate the purpose, technical solutions and advantages of the application. It should be understood that the above-mentioned embodiments are only specific embodiments of the application and are not used to limit the protection scope of the application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principle of the application should be included in the protection scope of the application.

Claims

1. A method for selecting influencers based on a large AI model, characterized in that, include: Obtain the user-input filtering criteria and the data of the talents to be filtered. The filtering criteria include text content requirements, image data requirements, and qualification information requirements. Based on a pre-defined semantic parsing model, the data of the selected influencers is matched and semantically analyzed with the text content requirements, and various semantic scores are output. The semantic parsing model is constructed by combining a bidirectional long short-term memory network with a conditional random field, extracting keywords through a hidden Dirichlet distribution, and calculating cosine similarity on historical user text data. Based on a preset text classification model, the note style type identification and logical judgment are performed on the data of the experts to be screened to obtain scores for each style and scores for each logical verification. Based on a preset multimodal recognition model, the data of the talents to be screened and the image data requirements are subjected to element detection and style verification, and each image matching score is output. Based on the data of the talents to be screened, each qualification score is calculated using a preset weighting formula, and the weighting formula is adjusted according to the qualification information requirements. The semantic scores, style scores, logical verification scores, image matching scores, and qualification scores are fused together to generate a comprehensive score. Based on the comprehensive scores, the influencers in the data to be screened are then selected.

2. The talent selection method based on AI large model according to claim 1, characterized in that, The preset semantic parsing model matches and performs semantic analysis on the data of the influencers to be screened with the text content requirements, and outputs various semantic scores, specifically: The text content requirements and the data of the influencers to be selected are input into the semantic parsing model, so that the semantic parsing model can perform word segmentation on the data of the influencers to be selected based on the bidirectional long short-term memory network and combined with the conditional random field to obtain each first word segment; Using the Latent Dirichlet Distribution Model, extract each second word and its corresponding weight from each first word segmentation. Based on each second word and its corresponding weight, the cosine similarity between each second word and the text content is calculated, and each semantic score is output.

3. The talent selection method based on AI large-scale model according to claim 1, characterized in that, The preset text classification model is used to identify note style types and make logical judgments on the data of the experts to be screened, so as to obtain style scores and logical verification scores, specifically: The data of influencers to be screened is input into the text classification model, so that the text classification model can identify the style of the data of influencers to be screened through the BERT model, and analyze the logical structure of the data of influencers to be screened through the syntax tree, and output the text style and text logic. The text style and text logic are compared with the text content requirements in the filtering conditions, and the scores for each text style and text logic are output.

4. The talent selection method based on AI large-scale model according to claim 1, characterized in that, The preset multimodal recognition model performs element detection and style verification on the data of the talents to be screened and the image data requirements, and outputs the matching scores for each image, specifically: The image data requirements and the data of influencers to be screened are input into the multimodal recognition model, so that the multimodal recognition model can perform element detection on the data of influencers to be screened according to the target detection algorithm, identify products, scenes and people, and determine the image elements of the data of influencers to be screened. Based on semantic segmentation algorithms, style verification is performed on the data of influencers to be screened, and the color distribution and composition style of the data are analyzed to determine the image style of the data of influencers to be screened. The data of the influencers to be screened is converted to the HSV color space, a color histogram is calculated, and the color tone of the color histogram is determined by the Bach distance to determine the image color of the influencer data to be screened. The image elements, image style, and image color are compared with the image data requirements in the filtering criteria, and a matching score is output for each image.

5. The talent selection method based on AI large model according to claim 1, characterized in that, The process involves calculating each qualification score based on the data of the talents to be screened using a preset weighting formula, specifically as follows: Extract the fan profile, domain verticality, historical conversion rate, and interaction rate from the data of the influencers to be screened; Based on the qualification information requirements and the preset weighting formula, the data of the talents to be screened are weighted and summed to calculate the qualification score for each qualification.

6. The talent selection method based on AI large model according to claim 1, characterized in that, The process involves fusing the semantic scores, style scores, logic verification scores, image matching scores, and qualification scores to generate comprehensive scores. Based on these comprehensive scores, influencers in the influencer data are then selected. Specifically: The semantic score, style score, and logical verification score are combined into a text matching score. According to the preset weight allocation, the text matching score, image matching score and qualification score are weighted and summed to obtain the comprehensive score. Based on the aforementioned comprehensive scores, the influencers in the data to be screened are then selected.

7. A talent selection device based on an AI large-scale model, characterized in that, include: The acquisition module acquires the user-input filtering conditions and the data of the talents to be filtered. The filtering conditions include text content requirements, image data requirements, and qualification information requirements. The semantic scoring module is used to match and semantically analyze the data of the selected experts with the text content requirements based on a preset semantic parsing model, and output various semantic scores. The semantic parsing model is constructed by combining a bidirectional long short-term memory network with a conditional random field, extracting keywords through a hidden Dirichlet distribution, and calculating historical user text data using cosine similarity. The text classification module is used to identify the note style type and make logical judgments on the data of the experts to be screened based on a preset text classification model, so as to obtain the scores of each style and the scores of each logical verification. The modality recognition module is used to perform element detection and style verification on the data of the talents to be screened and the image data requirements based on a preset multimodal recognition model, and output the matching score of each image. The qualification scoring module is used to calculate each qualification score based on the data of the talents to be screened, using a preset weight formula. The weight formula is adjusted according to the qualification information requirements. The filtering module is used to integrate the semantic scores, style scores, logical verification scores, image matching scores, and qualification scores to generate comprehensive scores, and to filter the influencers in the influencer data based on the comprehensive scores.

8. The talent selection device based on an AI large model according to claim 7, characterized in that, The preset semantic parsing model matches and performs semantic analysis on the data of the influencers to be screened with the text content requirements, and outputs various semantic scores, specifically: The text content requirements and the data of the influencers to be selected are input into the semantic parsing model, so that the semantic parsing model can perform word segmentation on the data of the influencers to be selected based on the bidirectional long short-term memory network and combined with the conditional random field to obtain each first word segment; Using the Latent Dirichlet Distribution Model, extract each second word and its corresponding weight from each first word segmentation. Based on each second word and its corresponding weight, the cosine similarity between each second word and the text content is calculated, and each semantic score is output.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the talent selection method based on an AI large model as described in any one of claims 1 to 6.

10. A terminal device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the AI-based large model-based expert selection method as described in any one of claims 1 to 6.

Images