One-stop talent screening method and system

Abstract

The invention relates to the technical field of human resource management, and provides a one-stop talent screening method and system. The method comprises the steps of performing text analysis on a job seeker resume, identifying structured information in the job seeker resume, performing weight distribution on the analyzed information through a preset post competency model, and generating a resume analysis result containing a candidate and target post matching degree score; consultation information about enterprise information, position treatment and office environment of the job seeker is obtained, the consultation information is analyzed to obtain a question intention of the user, and response content corresponding to the question intention is generated; performing semantic analysis on the communication record corresponding to the job seeker, and identifying the job hunting willingness level of the job seeker; and dynamically generating personalized interview questions through an interview question generation model in combination with the matching degree score and the job hunting willingness level to complete talent screening. The method is suitable for full-process automatic processing of resume screening, intelligent communication and adaptive interview in the enterprise recruitment process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of human resource management, and particularly relates to a one-stop talent screening method and system. BACKGROUND

[0002] In the enterprise recruitment process, talent screening is a key link to determine the quality of recruitment. The traditional talent screening method mainly relies on manual operation, and has problems such as low efficiency, strong subjectivity, and non-uniform screening standards. With the development of information technology, some enterprises have introduced resume analysis tools and online communication platforms, but the existing technology still has the following significant defects: 1. Shallow and non-quantitative resume screening: the existing technology mainly uses keyword matching or simple rule extraction of resume information, which can only identify the surface text matching, cannot perform deep semantic analysis on information such as education background and work experience, and lacks a quantitative evaluation model based on historical recruitment data, resulting in that the post matching degree judgment depends on manual experience and the precision is insufficient.

[0003] 2. Mechanism of intelligent communication and lack of willingness recognition: the existing customer service or communication module usually responds to problems based on a fixed knowledge base, cannot understand the deep intention of the user's question, and cannot intelligently recognize the job seeker's job-seeking willingness level through communication records. The manual annotation or rule-based willingness judgment method is time-consuming and laborious, and lacks data-driven generalization ability.

[0004] 3. Fixed and non-personalized interview: the traditional interview mainly uses a fixed question bank, and the difficulty and direction of the questions cannot be dynamically adjusted according to the real-time answers of the job seekers, and it is difficult to comprehensively evaluate the candidate's ability. At the same time, the interview evaluation depends on the subjective record of the interviewer, lacks quantitative analysis of the answer content and multi-dimensional data fusion, resulting in strong subjectivity and low repeatability of the screening result.

[0005] Therefore, there is an urgent need for a method to solve at least one of the above problems. SUMMARY

[0006] The present application provides a one-stop talent screening method and system, which aims to solve the problem that although there are intelligent attempts in single links (such as resume keyword extraction and fixed question bank interview) in the existing technology, there is no technical solution to deeply integrate natural language processing, machine learning, and generation model algorithms to build a full-process intelligent linkage system from resume analysis to interview evaluation. In particular, the existing technology does not disclose the synergistic effect of the post competency model (based on historical data training), the dialogue intention recognition algorithm, and the adaptive interview question generation model to realize the closed-loop screening logic of "resume analysis-intelligent communication-dynamic interview".

[0007] In a first aspect, the embodiments of the present application provide a one-stop talent screening method, which comprises: The natural language processing algorithm is used to analyze the resume text of the job seeker, identify the structured information in the resume of the job seeker, assign weights to the parsed information through a preset job competence model, and generate a resume analysis result containing the matching degree score of the candidate and the target position; the job competence model is trained based on historical recruitment data through a machine learning algorithm, and is used to quantitatively evaluate the matching degree of the resume information and the position demand; Obtain the consultation information of the job seeker on enterprise information, position treatment and office environment, analyze the consultation information to obtain the user's question intention, and generate the answer content corresponding to the question intention according to the pre-configured knowledge base of the enterprise and the externally enabled data query interface; The semantic analysis of the communication record corresponding to the job seeker is performed through a text classification algorithm, and the job-seeking intention level of the job seeker is identified, and the training data of the text classification algorithm includes the job-seeking intention labels manually labeled in the historical communication record; Combined with the matching degree score and the job-seeking intention level, individualized interview questions are dynamically generated through an interview question generation model; during the interview process, the difficulty and direction of the subsequent questions are dynamically adjusted according to the real-time answer content of the job seeker through a semantic similarity calculation algorithm, and the key answer information in the interview process is recorded to generate an interview comprehensive evaluation report to complete talent screening.

[0008] In some embodiments, the natural language processing algorithm is used to analyze the resume text of the job seeker, identify the structured information in the resume of the job seeker, including: extracting education background entities, work experience entities, skill certificate entities and project experience entities from the resume text through a named entity recognition algorithm, classifying the extracted entities using an information classification model to form structured resume information data, and the information classification model is trained through supervised learning training of a training set containing labeled resume text data.

[0009] In some embodiments, the parsed information is assigned weights through a preset job competence model to generate a resume analysis result containing the matching degree score of the candidate and the target position, including: determining a position demand keyword set through a keyword extraction algorithm according to the position description text of the target position, performing semantic matching between the structured resume information data and the position demand keyword set, and assigning weight values to the matched resume information according to a preset weight rule; wherein the weight rule is set based on the importance level of the position demand keywords, and the matching degree score of the candidate and the target position is calculated by weighted summation.

[0010] In some embodiments, the post competency model is trained by a machine learning algorithm based on historical recruitment data, including: obtaining the resume data of candidates who pass the screening in historical recruitment as positive samples, and the resume data of candidates who do not pass the screening as negative samples, performing structured information extraction and position demand keyword labeling on the positive and negative sample resume data, and constructing a training data set; using a support vector machine algorithm or a random forest algorithm to train the training data set, obtaining a post competency model capable of outputting a resume information and position demand matching degree score, and optimizing the model parameters through a cross-validation method.

[0011] In some embodiments, the parsing consultation information obtains the user question intention, including: inputting the user consultation information into a pre-trained intention classification model, the intention classification model is constructed based on a Transformer neural network architecture, trained by historical consultation data containing enterprise information class, position treatment class, office environment class and other multi-class labels, outputting the intention category corresponding to the user question, the intention category at least includes enterprise basic information query, salary and welfare consultation, office environment understanding and position requirement confirmation.

[0012] In some embodiments, the generation of the answer content corresponding to the question intention according to the pre-configured knowledge base of the enterprise and the optional external data query interface includes: retrieving a standard answer matching the user question intention from the pre-configured knowledge base of the enterprise, if there is no matching answer in the knowledge base and the external data query interface is in the enabled state, obtaining the public information related to the question intention in the official website of the enterprise, or obtaining the industry general information through a pre-set third-party data interface, and generating the answer content after fusion processing of the knowledge base answer, the official website information of the enterprise and the third-party data.

[0013] In some embodiments, the semantic analysis of the communication record corresponding to the job seeker through a text classification algorithm to identify the job-seeking willingness level of the job seeker includes: extracting text paragraphs from the communication record as input data, converting the text paragraphs into semantic vectors through a word embedding model, inputting the semantic vectors into a pre-trained job-seeking willingness classification model, the job-seeking willingness classification model adopts a convolutional neural network or a recurrent neural network architecture, and the corresponding training data contains high, medium and low three-grade job-seeking willingness labels manually labeled in historical communication records, and the job-seeking willingness level of the current job seeker is obtained through the model output.

[0014] In some embodiments, the difficulty and direction of the subsequent question are dynamically adjusted according to the real-time answer content of the job seeker during the interview process through a semantic similarity calculation algorithm, including: performing semantic similarity calculation on the real-time answer text of the job seeker and a preset high-quality answer library, the semantic similarity calculation generates sentence vectors of the answer text and the high-quality answer based on a pre-trained language model, and the similarity score between the vectors is calculated through a cosine similarity algorithm; if the similarity score is higher than a first preset threshold, an advanced question of higher difficulty is selected, and if the similarity score is lower than a second preset threshold, a basic guiding question is selected, so as to adjust the difficulty and direction of the subsequent interview question.

[0015] In some embodiments, the key answer information recorded during the interview process is used to generate an interview comprehensive evaluation report to complete talent screening, including: extracting the answer content, answer duration and semantic similarity score of the job seeker for each interview question from the interview process record, combining the matching degree score and the job seeking willingness level, performing weighted calculation according to the preset evaluation rules, generating an interview comprehensive evaluation report containing the candidate comprehensive ability score, the job position matching degree analysis and the job seeking willingness evaluation, and completing the talent screening decision according to the score result of the evaluation report.

[0016] In a second aspect, the present application provides a one-stop talent screening system, the system comprising: a text analysis unit configured to analyze a resume of a job seeker by using a natural language processing algorithm, identify structured information in the resume of the job seeker, perform weight distribution on the analyzed information by using a preset job position competence model, and generate a resume analysis result containing a matching degree score of the candidate for a target position; the job position competence model is obtained by machine learning algorithm based on historical recruitment data, and is used to quantitatively evaluate the matching degree of the resume information and the position demand; an information acquisition unit configured to acquire consultation information of the job seeker about enterprise information, position benefits and office environment, analyze the consultation information to obtain the user's questioning intention, and generate answer content corresponding to the questioning intention according to a pre-configured knowledge base of the enterprise and an externally query interface that can be enabled; a semantic analysis unit configured to perform semantic analysis on the communication record corresponding to the job seeker by using a text classification algorithm, and identify the job seeking willingness level of the job seeker; the training data of the text classification algorithm contains the job seeking willingness labels manually labeled in the historical communication record; a screening completion unit configured to dynamically generate personalized interview questions by using an interview question generation model in combination with the matching degree score and the job seeking willingness level; during the interview process, the difficulty and direction of the subsequent question are dynamically adjusted according to the real-time answer content of the job seeker through a semantic similarity calculation algorithm, and the key answer information during the interview process is recorded to generate an interview comprehensive evaluation report to complete talent screening.

[0017] The one-stop talent screening method and system provided by the embodiment of the application, through natural language processing algorithm to analyze the structured information of the resume, and using the post competence model trained based on historical data to score the quantitative matching degree, solves the problem of the traditional screening of the superficialization; the intelligent communication layer: through the dialogue intention recognition and text classification algorithm to realize the automatic classification of the job seeking intention, replacing the inefficiency and subjective bias of the artificial judgment; the interview execution layer: based on the generated model to dynamically generate personalized questions, and through the semantic similarity algorithm to adjust the interview difficulty in real time, and to build the adaptive evaluation system. The above technical solution realizes the data linkage (matching degree score→intention level→interview question dynamic generation) among the algorithms, and forms the intelligent screening closed loop of the whole process.

[0018] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the application. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0020] Figure 1 is a step schematic flow chart of a one-stop talent screening method provided by an embodiment of the application; Figure 2 is a structural schematic block diagram of a one-stop talent screening system provided by an embodiment of the application; Figure 3 is a structural schematic block diagram of a computer device provided by an embodiment of the application.

[0021] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the application will be described clearly and completely in combination with the drawings in the embodiments of the application. Obviously, the described embodiments are some embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0023] The flow chart shown in the drawings is only an example description, and is not necessarily to include all the contents and operations / steps, and is not necessarily executed in the described order. For example, some operations / steps can be decomposed, combined or partially combined, so that the actual execution order can be changed according to the actual situation.

[0024] It should be understood that, in order to facilitate the clear description of the technical solutions of the embodiments of the present application, in the embodiments of the present application, the terms "first", "second", etc. are used to distinguish the same items or similar items with basically the same function and role. Those skilled in the art can understand that the terms "first", "second", etc. do not limit the quantity and execution order, and the terms "first", "second", etc. also do not necessarily mean different.

[0025] It should be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in the specification and the appended claims of the present application, the singular forms "a", "an" and "the" are intended to include the plural forms unless the context clearly dictates otherwise.

[0026] It should also be understood that the term "and / or" used in the specification and the appended claims of the present application means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0027] In the enterprise recruitment process, talent screening is a key link to determine the quality of recruitment. The traditional talent screening method mainly relies on manual operation, which has problems such as low efficiency, strong subjectivity, and non-uniform screening standards. With the development of information technology, some enterprises have introduced resume analysis tools and online communication platforms, but the existing technology still has the following significant defects: 1. Shallow and non-quantitative resume screening: the existing technology mainly uses keyword matching or simple rule extraction of resume information, which can only identify the surface text matching and cannot perform deep semantic analysis on information such as education background and work experience. Moreover, there is a lack of quantitative evaluation model based on historical recruitment data, resulting in a dependence on manual experience for job matching degree judgment and insufficient precision.

[0028] 2. Mechanical and lack of willingness recognition in intelligent communication: the existing customer service or communication module usually responds to problems based on a fixed knowledge base, cannot understand the deep intention of the user's question, and cannot intelligently recognize the job seeker's job-seeking willingness level through communication records. The manual annotation or rule-based willingness judgment method is time-consuming and labor-intensive, and lacks data-driven generalization ability.

[0029] 3. Fixed and non-personalized interview: traditional interviews mostly use fixed question banks, and the difficulty and direction of the questions cannot be dynamically adjusted according to the real-time answers of the job seekers, making it difficult to fully assess the candidate's ability. At the same time, the interview evaluation relies on the subjective record of the interviewer, lacks quantitative analysis of the answer content and multi-dimensional data fusion, resulting in strong subjectivity and low repeatability of the screening results.

[0030] Although there are single-link intelligent attempts in the prior art (such as resume keyword extraction, fixed question bank interview), there is no technical solution to deeply integrate natural language processing, machine learning, and generation model algorithms to build an intelligent linkage system from resume analysis to interview evaluation. In particular, the prior art does not disclose the synergistic effect of the post competence model (trained based on historical data), the dialogue intent recognition algorithm, and the adaptive interview question generation model to realize the closed-loop screening logic of "resume analysis-intelligent communication-dynamic interview". Therefore, how to solve the balance between efficiency and accuracy in the talent screening process through algorithm innovation has become a technical problem to be solved in this field.

[0031] To solve the above problems, please refer to Figure 1 The embodiment of the present application provides a one-stop talent screening method applied to a computer device. The computer device can be deployed on a single server or a server cluster. It can also be deployed on a handheld terminal, a notebook computer, a wearable device, or a robot, etc. At the same time, it should be noted that the method provided by the present application involves extracting each piece of information under the authorization of the relevant user and in accordance with the relevant provisions, without infringing on the privacy of the user.

[0032] The one-stop talent screening method provided includes steps S101 to S104. Details are as follows: Step S101. A natural language processing algorithm is used to analyze the text of the job seeker's resume, identify the structured information in the job seeker's resume, and assign weights to the analyzed information through a pre-set post competence model to generate a resume analysis result containing the matching degree score of the candidate and the target position; the post competence model is trained based on historical recruitment data through a machine learning algorithm and is used to quantitatively evaluate the matching degree of the resume information and the position requirements.

[0033] Specifically, the natural language processing analyzes the resume: four core entities are extracted from the resume text through the named entity recognition (NER) algorithm - education background (school, major, graduation time), work experience (company name, position, tenure, job description), skill certificate (technical skill, certificate name, certification authority), and project experience (project name, role, achievement description); further utilize information classification model (such as BERT-based text classifier) to classify the extracted entities in a fine-grained manner (such as "work experience" is divided into "core position", "related position", and "non-related position") to form a structured data table.

[0034] Post competency model quantitative assessment: The model is trained based on historical recruitment data (positive samples are resumes that pass the screening, and negative samples are resumes that do not pass the screening), uses support vector machine (SVM) or random forest algorithm, inputs structured resume information and job requirement keywords (extracted from job description, such as "Java development" and "data analysis"), and outputs a matching score of 0-100. Scoring logic: The matched keywords are classified according to their importance (such as "core skills" with a weight of 30% and "preferred conditions" with a weight of 20%), and the comprehensive score is calculated by weighted summation.

[0035] In the recruitment scenario, the HR of the enterprise uploads the job description of the target position in the system, and the system automatically parses and generates a keyword library of job requirements (such as extracting "deep learning", "TensorFlow", and "algorithm optimization" for the position of "machine learning engineer"). After the job seeker uploads the resume, the system parses it within 10 seconds and generates a structured resume report containing an education experience timeline, a skill matrix graph, and a project achievement quantification table, and marks the matching highlights with the job requirements (such as "XX project in work experience completely matches XX technology required by the position, with a weight addition of 25%"). The HR can view the visual matching score radar chart to quickly locate high-matching candidates, replacing the traditional inefficient process of manually reading resumes line by line.

[0036] Step S102. Obtain the job seeker's consultation information about the enterprise information, job treatment, and office environment, parse the consultation information to obtain the user's question intention, and generate the answer content corresponding to the question intention according to the pre-configured knowledge base of the enterprise and the optional external data query interface.

[0037] Specifically, the question intention recognition inputs the job seeker's consultation text (such as "How is the company's overtime situation?" and "What is the promotion path for this position?") into a pre-trained intention classification model (based on the Transformer architecture, with training data containing 100,000+ historical consultation records and 15 types of intention labels such as "salary and welfare", "career development", and "office environment"). The model outputs the most probable intention category (such as "overtime situation" corresponding to "work environment" intention).

[0038] The answer content generation prioritizes the retrieval of the enterprise knowledge base (stores standardized answers such as "The company implements a flexible work system, and overtime needs to be applied in advance") and, if there is no matching answer in the knowledge base and the external interface is enabled, uses web crawlers to scrape the contents of the "Join Us" section on the company's official website or calls third-party job market data interfaces (such as the "Zhaopin" company evaluation) to generate answers by fusing multiple sources of information (such as "According to the information on the official website, the average overtime for this position is 2-3 times per month, depending on the project progress").

[0039] In the recruitment scenario, job seekers ask questions through the recruitment platform IM system, enterprise WeChat or official website customer service window. The system analyzes the intention in real time and automatically replies, reducing the workload of HR repeated answering; for high-frequency questions (such as "salary range" and "job responsibilities"), the system actively pushes common question shortcuts in the chat window to guide job seekers to quickly obtain information; the answer content supports rich text format (such as salary structure displayed in table and office environment attached with real scene picture link), which improves communication efficiency and experience.

[0040] Step S103. Perform semantic analysis on the communication record corresponding to the job seeker by a text classification algorithm, and identify the job-seeking willingness level of the job seeker, wherein the training data of the text classification algorithm includes the job-seeking willingness labels manually annotated in the historical communication records.

[0041] Specifically, communication record text processing: extract the dialogue record between the job seeker and the HR / system (such as interview invitation email, communication summary, IM chat record), convert the text paragraph into a 300-dimensional semantic vector through the Word2Vec or GloVe word embedding model, and input the job-seeking willingness classification model (using LSTM recurrent neural network, and the training data includes 20,000+ historical dialogues labeled as "high willingness", "medium willingness" and "low willingness").

[0042] Level determination logic: the model focuses on key sentences (such as "When can I start working?" and "Does this position have a clear training plan?") through an attention mechanism, and outputs a probability value corresponding to the level (probability>80% for "high", 50%-80% for "medium", and <50% for "low").

[0043] In the recruitment scenario, the system automatically extracts the text of the job seeker in each communication link (including the job-seeking intention in the resume, pre-interview question consultation, and salary negotiation dialogue), and generates a "job-seeking willingness analysis report"; for "low willingness" candidates, the system automatically triggers targeted communication strategies (such as pushing enterprise advantage scripts and adjusting salary negotiation techniques); for "high willingness" candidates, they are preferentially arranged in the interview process to reduce the loss of high-quality talents.

[0044] Step S104. Combine the matching degree score and the job-seeking willingness level, and dynamically generate personalized interview questions through an interview question generation model; during the interview process, according to the real-time answer content of the job seeker, dynamically adjust the difficulty and direction of the subsequent questions through a semantic similarity calculation algorithm, and record the key answer information in the interview process for generating an interview comprehensive evaluation report to complete talent screening.

[0045] Specifically, personalized question generation: interview question generation model (based on GPT-3.5 fine-tuning, input matching score, job application willingness level, job type) outputs 3-5 customized questions (such as for a candidate with a matching score of 85 for a research and development position, generate "Please describe the algorithm optimization challenges you encountered in the XX project and the solutions").

[0046] Dynamic adjustment of interview: semantic similarity calculation (based on BERT sentence vector generation, cosine similarity threshold set to 0.7 and 0.4) is performed on the real-time answers of job seekers and the preset high-quality answer library (formed by senior employee answer annotation) to determine whether the similarity is greater than 0.7: enter advanced questions (such as "What is your optimization idea if the solution has performance bottlenecks in the production environment?"); similarity <0.4: switch to basic guiding questions (such as "Can you first introduce the basic principles of the technology?"); evaluation report generation: extract answer content, duration (reflecting reaction speed), similarity score, combine with resume matching score (accounting for 40%) and willingness level (accounting for 20%), generate a comprehensive score through a weighted formula, and output a structured report containing "technical capability radar chart" and "job motivation analysis".

[0047] In the recruitment scenario, after the interviewer logs in to the system, an individualized interview outline is automatically obtained for the current candidate, and the difficulty of the questions changes dynamically with the answers, avoiding the mechanical nature of fixed question banks; During the interview, the system real-time labels key response highlights (such as "the candidate mentioned XX technical innovation point, which is highly matched with the job requirements"), assisting the interviewer to quickly record; within 5 minutes after the interview, an evaluation report is generated, and HR can compare multiple candidate reports to make screening decisions based on quantitative data, reducing subjective judgment bias.

[0048] In some embodiments, the natural language processing algorithm is used to analyze the text of the resume, and the structured information in the resume is identified, including: extracting education background entities, work experience entities, skill certificate entities, and project experience entities from the resume text through a named entity recognition algorithm; using an information classification model to classify the extracted entities, forming structured resume information data, and the information classification model is obtained through supervised learning training of a training set containing resume text labeled data.

[0049] Four core entities are extracted from the resume text through a named entity recognition (NER) algorithm: education background entities (school name, major, graduation time, etc.), work experience entities (company name, job title, tenure, job description, etc.), skill certificate entities (technical skill name, certificate name, certification authority, etc.), and project experience entities (project name, role, technical solution, achievement data, etc.).

[0050] The extracted entities are classified into fine-grained categories using an information classification model (such as a BERT-based text classifier) (e.g., distinguishing "work experience" into "core related positions", "marginal related positions", "non-related positions", and distinguishing "skill certificates" into "hard technical skills" and "soft general skills"), forming a structured resume information data table.

[0051] The information classification model is trained through supervised learning, and the training set contains resume text data with annotated entity categories (such as 100,000+ historical resumes, with each entity's category label manually annotated).

[0052] The HR of an enterprise uploads a job seeker's resume (supports PDF, Word format) to the recruitment system, and the system automatically triggers the NER algorithm to highlight the education, work, skills, and project content in the resume within 10 seconds, and generates a structured data table (such as a table column containing "entity type-subcategory-specific content-timeline"). For the "work experience" entity, the system automatically identifies whether it is a "core related position" of the target position (for example, if the target position is "Java development engineer", then the "Java backend development" position is marked as core related, and the "frontend development" position is marked as marginal related), and the different color labels are visualized to assist the HR to quickly locate the key information.

[0053] In some embodiments, the parsed information is assigned a weight by a preset job competence model to generate a resume analysis result containing a matching score of the candidate and the target position, including: determining a position requirement keyword set through keyword extraction algorithm according to the position description text of the target position, performing semantic matching between the structured resume information data and the position requirement keyword set, and assigning a weight value to the matched resume information according to a preset weight rule; wherein the weight rule is set based on the importance level of the position requirement keywords, and the matching score of the candidate and the target position is calculated by weighted summation.

[0054] Keyword extraction extracts a position requirement keyword set (such as "deep learning", "TensorFlow", "model optimization" for "machine learning engineer" position) from the position description text of the target position through TF-IDF or TextRank algorithm, and divides them into core keywords (such as necessary skills) and secondary keywords (such as preferred skills) according to their importance.

[0055] Semantic matching and weight assignment: semantic matching of structured resume information with a set of keywords (cosine similarity based on Word2Vec word vector calculation, threshold set to 0.6), and the matched resume information is assigned a weight value according to the importance of the keyword (core keyword corresponds to weight 30%-50%, secondary keyword corresponds to 10%-20%). Matching degree calculation: the comprehensive score is calculated by the weighted summation formula, the formula is: matching degree score =∑(matching keyword weight x resume information matching degree score).

[0056] When the enterprise releases a new position, the system automatically parses the job description to generate a set of keywords (supports manual adjustment of weight by HR), for example, in the "data analyst" position, "SQL" and "Python" are set as core keywords (weight 25% each), and "Tableau" is set as a secondary keyword (weight 15%).

[0057] After parsing the resume of the job seeker, the system generates a "keyword matching matrix table", which lists the matched keywords, corresponding resume content, weight value and score (for example, a candidate's resume "master Python data cleaning" matches the core keyword "Python", weight 25%, score 25 points) line by line, and finally outputs a matching degree score of 0-100 points, and marks "no matching core keyword" (red alert) to remind HR to pay attention.

[0058] In some embodiments, the job competence model is trained by machine learning algorithm based on historical recruitment data, including: obtaining the resume data of candidates who pass the screening in historical recruitment as positive samples, and the resume data of candidates who do not pass the screening as negative samples, performing structured information extraction and position demand keyword labeling on the positive and negative sample resume data to construct a training data set; using support vector machine algorithm or random forest algorithm to train the training data set to obtain a job competence model capable of outputting resume information and position demand matching degree score, and optimizing the model parameters through cross-validation method.

[0059] Training data construction includes: positive samples: resume data of candidates who pass the screening in historical recruitment (labeled "pass"); negative samples: resume data of candidates who do not pass the screening (labeled "not pass"). Perform structured information extraction and position demand keyword labeling (label whether each resume contains core / secondary keywords of the target position) on the positive and negative samples to form a training data set containing thousands to tens of thousands of samples.

[0060] Model training uses support vector machine (SVM) or random forest algorithm, inputs structured resume information and keyword labeling data, and outputs a matching degree score of 0-100 points. Optimize model parameters (such as kernel function type of SVM, tree depth of random forest) through k-fold cross-validation (such as 10-fold) to ensure the generalization ability of the model.

[0061] By importing historical recruitment results (pass / fail resumes) into the system regularly (such as every quarter) by enterprise HR, the system automatically cleanses the data (removes sensitive information) and labels positive and negative samples, triggering the model training process. When a new position is released, the system automatically calls the corresponding position competence model (such as the SVM model for technical positions and the random forest model for sales positions) according to the position type (such as technical positions and sales positions), avoiding "one-size-fits-all" evaluation, such as sales positions focusing more on "communication skills" and "customer resources" keywords, and the model weights are tilted towards such keywords.

[0062] In some embodiments, the parsing consultation information obtains the user's question intention, including: inputting the user's consultation information into a pre-trained intention classification model, the intention classification model is constructed based on a Transformer neural network architecture, trained by historical consultation data containing enterprise information class, position treatment class, office environment class and other multi-class labels, outputting the intention category corresponding to the user's question, the intention category at least includes enterprise basic information query, salary and welfare consultation, office environment understanding and position requirement confirmation.

[0063] By constructing an intention classification model (based on the Transformer architecture, such as BERT-base), the training data contains tens of thousands of historical consultation records, and 4 or more intention labels are labeled (such as "enterprise basic information query", "salary and welfare consultation", "office environment understanding", and "position requirement confirmation"). Input the user's consultation information into the model, the model captures the text semantics through self-attention mechanism (such as "accumulation fund payment ratio" corresponds to "salary and welfare consultation" intention), and outputs the intention category with the highest probability (the probability threshold is set to 70%, and below is marked as "other intention").

[0064] Job seekers ask "how is the annual bonus calculated for this position?" through the chat window of the recruitment platform, the system analyzes the intention in real time as "salary and welfare consultation", and retrieves the corresponding answer from the knowledge base (such as "the company's annual bonus is based on performance appraisal, ranging from 1-3 months of monthly salary"). For "position requirement confirmation" intention (such as "can an intern apply?"), the system automatically associates the "employment requirements" field in the position description to generate accurate answers (such as "this position welcomes outstanding interns, and requires XX basic skills"), reducing the repeated explanation work of HR.

[0065] In some embodiments, the generating the response content corresponding to the interview intention according to the pre-configured knowledge base of the enterprise and the optional external data query interface comprises: retrieving a standard answer matching the interview intention of the user from the pre-configured knowledge base of the enterprise, if there is no matching answer in the knowledge base and the external data query interface is in an enabled state, obtaining public information related to the interview intention in the official website of the enterprise, or obtaining industry general information through a preset third-party data interface, and generating the response content after fusing and processing the knowledge base answer, the official website information of the enterprise and the third-party data.

[0066] The knowledge base retrieval retrieves a standard answer (such as "office environment" intention corresponding to "the company is located in XX business circle, open office area, equipped with rest area") matching the user intention from the pre-configured knowledge base (Excel table or database) of the enterprise in priority.

[0067] The external data fusion obtains the content of the "recruitment information" block (such as the "employee welfare" page) of the official website of the enterprise through a web crawler or calls a third-party interface to obtain industry general information (such as "the average overtime length of the industry to which the post belongs") if there is no matching answer in the knowledge base and the external interface is enabled, and generates the response content by fusing multiple source information through a text summary algorithm (to avoid direct copying and ensure compliance).

[0068] When the job seeker asks "does the company have a training system?", if there is no clear answer in the knowledge base, the system automatically crawls the content of the "talent training" page of the official website of the enterprise, and extracts the core information (such as "the company provides systematic training such as new employee training, technical salon and management course") as the answer.

[0069] For "industry prospect" type questions (non-knowledge base regular content), the system calls a third-party industry report interface to generate a concise answer (such as "according to XX data, the talent demand growth rate of the field to which the post belongs is 15% in the past three years") to improve the professionalism of communication.

[0070] In some embodiments, the semantic analysis of the communication record corresponding to the job seeker through a text classification algorithm to identify the job-seeking intention level of the job seeker comprises: extracting a text paragraph from the communication record as input data, converting the text paragraph into a semantic vector through a word embedding model, inputting the semantic vector into a pre-trained job-seeking intention classification model, the job-seeking intention classification model adopts a convolutional neural network or a recurrent neural network architecture, the corresponding training data contains high, medium and low three-grade job-seeking intention labels manually labeled in historical communication records, and the job-seeking intention level of the current job seeker is obtained through model output.

[0071] Text pre-processing extracts continuous text passages (e.g., more than 50 words) from communication records (IM chats, emails, interview records) and converts them into 300-dimensional semantic vectors using Word2Vec or GloVe models.

[0072] The job-seeking intention classification model uses a convolutional neural network (CNN) or recurrent neural network (RNN) architecture, and the training data includes manually labeled high / medium / low intention labels (e.g., "high intention" labels include dialogues containing keywords such as "hope to start work as soon as possible" and "very consistent with my career planning"). The model focuses on key sentences through an attention mechanism and outputs the current job seeker's intention level (probability values correspond to: high ≥ 80%, medium 50%-80%, and low < 50%).

[0073] The system automatically extracts all communication records between the job seeker and the HR (e.g., "What is the process for the physical examination for new employees?" "Is there a probationary examination for this position?") and generates an "intention analysis heat map" that highlights high-intention sentences (red highlights) and low-intention sentences (grayed out).

[0074] For "low-intention" candidates (e.g., frequently asking about "remote work policies" but without specific intentions to join), the system automatically triggers a reminder and suggests that the HR send an "enterprise advantage data package" or arrange a phone call to improve intention conversion; for "high-intention" candidates, they are prioritized as "to-be-interviewed" status.

[0075] In some embodiments, during the interview process, the difficulty and direction of subsequent questions are dynamically adjusted based on the job seeker's real-time answers, including: performing semantic similarity calculation between the job seeker's real-time answer text and a pre-set high-quality answer library. The semantic similarity calculation is based on pre-trained language models to generate sentence vectors for answer texts and high-quality answers, and calculates the similarity score between vectors through a cosine similarity algorithm; if the similarity score is higher than a first pre-set threshold, an advanced question with higher difficulty is selected, and if it is lower than a second pre-set threshold, a basic guiding question is selected, to adjust the difficulty and direction of subsequent interview questions.

[0076] The high-quality answer library provides high-quality answers to typical questions for each position by senior employees or interviewers, and generates sentence vectors through BERT and stores them (e.g., "How to handle data skew?" corresponds to the sentence vector of the high-quality answer).

[0077] The semantic similarity calculation converts the real-time answer of the job seeker into a sentence vector, and calculates the cosine similarity with the high-quality answer library vector, and sets double thresholds (such as a first threshold of 0.7 and a second threshold of 0.4): score≥0.7: determined as "high-quality answer", and the advanced question (such as "What are the optimization ideas besides the methods you mentioned?") is pushed; score<0.4: determined as "weak foundation", and the basic guiding question (such as "Please explain what data skew is?") is switched.

[0078] During the interview process, when the job seeker answers "How to optimize the efficiency of SQL query?", the system calculates the similarity between the answer and the high-quality answer in real time (such as a score of 0.85), and automatically pushes the advanced question "What is the difference in optimization strategy if it involves multi-table association?", which deepens the technical depth.

[0079] If the answer similarity is only 0.3 (such as the job seeker only repeats the theoretical definition), the system prompts the interviewer to switch to the basic question: "Can you give an example of a SQL case you have optimized?", avoiding problem difficulty discontinuity and ensuring comprehensive evaluation.

[0080] In some embodiments, the key answer information recorded during the interview process is used to generate an interview comprehensive evaluation report to complete talent screening, including: extracting the answer content, answer duration and semantic similarity score of the job seeker for each interview question from the interview process record, combining the matching degree score and the job seeking willingness level, and performing weighted calculation according to the preset evaluation rules to generate an interview comprehensive evaluation report containing candidate comprehensive ability score, job matching degree analysis and job seeking willingness evaluation, and completing talent screening decision according to the score result of the evaluation report.

[0081] Key information extraction: three core data are extracted from the interview record: answer content: text summary (such as extracting key words and sentences); answer duration: record the time from the question being raised to the answer being submitted (reflecting reaction speed, timeout red warning); semantic similarity score: the calculation result of the corresponding semantic similarity in the above embodiments.

[0082] Comprehensive evaluation model: a weighted formula is used to calculate the comprehensive score: comprehensive score=0.4x matching degree score+0.2x willingness level (high=100, medium=70, low=40)+0.4x interview answer score; a structured report containing ability radar chart (technical ability, communication ability, adaptability), matching degree analysis table (core keyword standard situation), and willingness evaluation conclusion is generated.

[0083] After the interview, the system generates a "candidate comprehensive assessment report" within 5 minutes, which can be directly viewed by HR: the radar chart shows that the candidate's "technical ability is 90 points, communication ability is 75 points"; the matching degree analysis marks "does not meet the core keyword 'distributed architecture', suggest rechecking"; the willingness evaluation prompts "high willingness, suggest prioritizing the offer process". The report supports exporting PDF for sharing and checking by the recruitment team, avoiding decision bias caused by differences in subjective records by interviewers.

[0084] In some embodiments, by adopting GPT-4 model for domain fine-tuning, input structured resume data, job competence score and historical high-quality candidate recommendation corpus (labeling "technical advantage", "experience matching", "potential point" and other dimensions), train to generate recommendation reasons conforming to the style of enterprise recruitment rhetoric.

[0085] Generation logic: the model extracts key highlights in the resume (such as "leading a project with tens of millions of users" and "collaborating across departments to optimize processes") through attention mechanisms, combines with the core requirements of the job (such as "must have experience in cloud native architecture"), and automatically generates 3-5 personalized recommendation reasons, each corresponding to a different evaluation dimension (technical ability, project experience, soft skills), and labels the matching weight of the recommendation reason and the job requirement (such as "cloud computing experience matching degree 90%").

[0086] When HR views the candidate list, the system automatically generates a structured recommendation reason module below the resume card, including: core advantage: "the candidate, during his tenure as a backend developer at XX company, led a containerization transformation project, resulting in a 40% increase in deployment efficiency, which highly matches the job requirement of 'cloud native implementation experience' (weight 35%)"; Potential supplement: "although there is no direct AI algorithm experience, he completed an NLP course through self-study and implemented a text classification project, showing strong learning ability (weight 20%)"; risk prompt: "management experience is only 1 year, which matches the 'team collaboration' secondary requirement of the job requirement (weight 15%)". The recommendation reason supports one-click copying to the interview assessment table or candidate communication email, reducing the time for HR to manually summarize, while ensuring that the recommendation logic is traceable and reducing subjective judgment bias.

[0087] In some embodiments, by designing a talent implicit association mining system based on graph neural network, including: knowledge graph construction: taking candidate entities (education background, work experience, skills, projects) as nodes, and taking association relationships (such as “XX University Computer Major → Java Skill” “XX Project → Cloud Computing Technology”) as edges, a talent knowledge graph is constructed. The graph neural network (GNN, such as GraphSAGE algorithm) is used to train the node embedding vector to capture the implicit association (such as “financial industry data analysis experience → risk modeling potential”) not explicitly labeled in the resume. Implicit matching calculation: the path similarity between the candidate node and the job requirement node is calculated by the graph walk algorithm (Random Walk), and the potential advantages (for example, the candidate has “supply chain management” experience, although it is not labeled “data analysis”, but the graph shows that SQL and Excel skills are often required in this field, and it is automatically associated to the data job basic ability) not directly mentioned in the resume but associated through the knowledge graph are identified.

[0088] In addition to the resume matching degree score, the system additionally generates an “implicit association score” to mark the candidate’s ability not explicitly mentioned but inferred through the knowledge graph. For example: a certain candidate’s resume does not write “Python”, but the knowledge graph shows that “bioinformatics” major often needs to use Python to process data, and the system prompts “potential Python application experience (confidence 75%)”; for the “cross-border e-commerce operation” position, the system finds that the candidate has “international trade” background + “English six-level”, and through the graph association “cross-border payment process” “multilingual communication” implicit ability, the matching degree score is automatically improved by 15%. HR can click “implicit association details” to view the derivation path, avoiding missing “atypical matching” candidates, especially suitable for innovative positions or cross-domain talent recruitment.

[0089] In some embodiments, by designing an interview stress test dynamic evaluation system based on sentiment computing, including: multi-modal sentiment analysis: in the video interview link, the voice signal (extracted through MFCC feature) and facial expression data (micro-expression recognized using OpenFace) are synchronously collected, input into the sentiment computing model (multi-modal network integrating CNN and LSTM), and the emotional fluctuation (such as anxiety value, speech speed change, pupil contraction frequency) of the job seeker under stress is analyzed in real time.

[0090] Stress scenario generation: based on reinforcement learning (RL), stress problems are dynamically generated, when the model detects that the emotion of the job seeker is stable (such as increased smile frequency, smooth tone), the aggressiveness of the problem is increased (such as “What responsibility do you take for the project failure in your previous company?”); if the anxiety value exceeds the threshold (such as blink frequency > 20 times / minute + stuttering), automatically switch to a mild problem to balance the evaluation intensity and user experience.

[0091] The video interview system integrates camera and microphone data collection (with user authorization), displays the emotional state heat map in real time during the interview, and labels the "stress tolerance" dimension score: when the candidate answers "What do you think is your biggest weakness?", the system detects that the voice frequency fluctuation + frown lasts for 3 seconds, marks "moderate stress reaction", and records whether there is a logical gap in the subsequent answer; For positions such as "customer service manager" that require high stress tolerance, the system automatically increases the proportion of stress problems (such as 40%), and generates a "stress situation response report" containing emotional recovery speed, language logic change curve and other quantitative indicators. The evaluation report is provided to the interviewer at the same time, which helps to judge the candidate's communication ability and psychological resilience in high-pressure situations and fills the gap in traditional interviews.

[0092] In some embodiments, a recruitment process dynamic optimization system based on reinforcement learning is designed, including: process state modeling: abstract the recruitment process as a state machine (states include "resume screening", "first interview", "second interview", "offer negotiation"), actions include "send assessment link", "arrange emergency interview", "adjust salary plan", and reward function is defined as "high-quality candidate conversion rate" "process time reduction rate". Use deep reinforcement learning (DRL, such as PPO algorithm) to train the strategy network, and dynamically select the optimal process strategy according to real-time recruitment data (matching degree, willingness level, historical process data).

[0093] Dynamic strategy generation: for high-matching but low-willingness candidates, the strategy network automatically generates differentiated processes such as "priority arrangement of executive interviews" and "open flexible work locations"; for low-matching but potential candidates, trigger the "arrangement of skill assessment + mentor communication" training process, rather than direct elimination.

[0094] The system automatically recommends the optimal operation scheme when the candidate enters each process node: when detecting a "matching degree 85 points + willingness level low" technical position candidate, a prompt pops up: "It is recommended to arrange a CTO phone call immediately, and the historical similar case conversion rate has increased by 60%"; for "matching degree 60 points but novel project experience" fresh graduates, automatically generate a "first participate in enterprise open day + technical test" process, rather than directly enter the interview, reduce the loss rate of high-quality potential stocks. HR can choose to accept system recommendations or manually adjust, and operation data flows back to the reinforcement learning model to form a "process strategy - result feedback - model optimization" closed loop, continuously improving the conversion efficiency of different types of candidates.

[0095] In some embodiments, by designing a personalized recruitment content generation engine based on a multi-modal large model, including: multi-modal input processing: input candidate resume text, past communication records (text + IM expressions), job requirement documents, use FLAVA multi-modal large model to extract cross-modal features (such as "open source project link" in the resume corresponding to technical enthusiasm, "positive expression" in communication corresponding to job seeking willingness). Personalized content generation: generate text that meets the candidate's preferences for different touch scenarios (interview invitation email, salary negotiation tactics, welcome letter). For example: for technical candidates, generate an invitation containing technical stack keywords: "Notice your practice on Kubernetes scheduling algorithm, invite you to explore our challenges in cloud native field"; for candidates who pay attention to career development, emphasize: "Our technical committee mechanism supports you to lead XX direction research and development, matching your "technical management" career goal mentioned in your resume.

[0096] HR selects the touch scenario in the system (such as "send interview invitation"), and the system automatically generates 3 versions of personalized text for selection: technical resonance version: highlight the matching between the technical difficulties in the candidate's project and the technical stack of the position; career development version: associate the "promotion expectation" in the candidate's resume with the job training system; cultural fit version: extract the candidate's "remote work preference" mentioned in the communication, and emphasize the enterprise's flexible work system. The text is accompanied by data annotation (such as "reference to candidate XX project experience, matching degree 92%"), ensuring that the communication content accurately hits the candidate's focus and improves response rate. For example, historical data shows that personalized invitation has a 45% higher response rate than general template.

[0097] In some embodiments, by designing a federated learning architecture, including: jointly with multiple enterprises (need to sign a data sharing agreement), without revealing the original resume data, share talent evaluation model parameters through federated average algorithm (FedAvg). Each enterprise locally trains a "job competence sub-model", uploads the gradient update to the central server, and aggregates to generate an industry-wide evaluation model, solving the problem of insufficient data for a single enterprise. Privacy protection mechanism: use homomorphic encryption technology to encrypt sensitive fields of resumes (name, contact information), add noise protection data features through differential privacy, to ensure compliance with the "Data Security Law".

[0098] After joining the federal learning alliance, small and medium-sized enterprises can use evaluation models that integrate industry data to improve the accuracy of screening for niche positions such as "quantum computing engineers." For example, an enterprise recruiting "new energy battery research and development positions" may lack local data. By using a federal model, it can obtain core competency labels for the position in the industry (such as "material characterization" and "cycle life test" with a 30% weight increase), avoiding screening biases due to lack of experience. The system automatically labels the candidate's ability and compares it to the industry average level (such as "battery cycle life test experience is better than 85% of industry candidates"), assisting HR in determining the market competitiveness of scarce talents. The federal model is updated regularly to keep up with the latest changes in industry talent demand (such as the addition of "ESG compliance" skill labels), ensuring that the enterprise evaluation criteria are synchronized with market trends.

[0099] Please refer to Figure 2 as shown, Figure 2 is a structural schematic diagram of a one-stop talent screening system 200 provided by an embodiment of the present application. The one-stop talent screening system 200 is used to execute the steps of the one-stop talent screening method shown in each of the above embodiments. The one-stop talent screening system 200 can be a single server or a server cluster, or the one-stop talent screening system 200 can be a terminal, which can be a handheld terminal, a notebook computer, a wearable device, or a robot, etc.

[0100] As Figure 2 shown, the one-stop talent screening system 200 includes: A text analysis unit 201 is configured to perform text analysis on a job seeker's resume using a natural language processing algorithm, identify structured information in the job seeker's resume, and distribute weights to the analyzed information through a pre-set job competence model to generate a resume analysis result containing a candidate's matching score with a target position. The job competence model is trained based on historical recruitment data through a machine learning algorithm and is used to quantitatively evaluate the matching degree of resume information and job requirements; An information acquisition unit 202 is configured to acquire the job seeker's consultation information about enterprise information, job treatment, and office environment, analyze the consultation information to obtain the user's questioning intention, and generate response content corresponding to the questioning intention according to the enterprise's pre-configured knowledge base and the externally available data query interface; A semantic analysis unit 203 is configured to perform semantic analysis on the communication records corresponding to the job seeker through a text classification algorithm, and identify the job seeker's job-seeking intention level. The training data of the text classification algorithm includes manually labeled job-seeking intention labels in historical communication records. The screening completion unit 204 is configured to combine the matching degree score and the job-seeking willingness level, and dynamically generate personalized interview questions through the interview question generation model; during the interview process, the difficulty and direction of subsequent questions are dynamically adjusted according to the real-time answer content of the job seeker through a semantic similarity calculation algorithm, and key response information in the interview process is recorded to generate an interview comprehensive evaluation report to complete talent screening.

[0101] It should be noted that, for the convenience and brevity of description, the specific working processes of the one-stop talent screening system and each module described above can refer to the corresponding content in the embodiments of the one-stop talent screening method described above, and will not be described here.

[0102] The one-stop talent screening method described above can be implemented in the form of a computer program, which can run on the device as shown in Figure 2 .

[0103] Please refer to Figure 3 , Figure 3 is a structural schematic block diagram of a computer device provided by the embodiment of the present application. The computer device includes a processor, a memory and a network interface connected through a device bus, wherein the memory can include a storage medium and an internal memory.

[0104] The storage medium can store an operating device and a computer program. The computer program includes program instructions, which, when executed, can cause the processor to execute any one of the one-stop talent screening methods.

[0105] The processor is configured to provide computing and control capabilities to support the operation of the entire computer device.

[0106] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium, which, when executed by the processor, can cause the processor to execute any one of the one-stop talent screening methods.

[0107] The network interface is configured to perform network communication, such as sending assigned tasks. Those skilled in the art can understand that the structure shown in Figure 3 , only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the terminal to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0108] It should be appreciated that the processor can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0109] In one embodiment, the processor is configured to run a computer program stored in the memory to implement the following steps: The resume of the job seeker is parsed by using a natural language processing algorithm to identify structured information in the resume of the job seeker, and the parsed information is weighted by using a preset job competence model to generate a resume analysis result containing a matching degree score of the candidate and the target position. The job competence model is trained by using a machine learning algorithm based on historical recruitment data, and is used to quantitatively evaluate the matching degree of the resume information and the job requirements. The consultation information of the job seeker about enterprise information, job treatment, and office environment is obtained, the consultation information is parsed to obtain the user's question intention, and the answer content corresponding to the question intention is generated according to the pre-configured knowledge base of the enterprise and the externally query interface that can be enabled; The semantic analysis of the communication record corresponding to the job seeker is performed by using a text classification algorithm to identify the job-seeking intention level of the job seeker, and the training data of the text classification algorithm contains the job-seeking intention labels manually annotated in the historical communication records. The personalized interview questions are dynamically generated by using a combination of the matching degree score and the job-seeking intention level, and a face-to-face interview question generation model. In the interview process, the difficulty and direction of the subsequent questions are dynamically adjusted according to the real-time answer content of the job seeker by using a semantic similarity calculation algorithm, and the key answer information in the interview process is recorded to generate an interview comprehensive evaluation report to complete the talent screening.

[0110] In some embodiments, the resume of the job seeker is parsed by using a natural language processing algorithm to identify structured information in the resume of the job seeker, including: extracting education background entities, work experience entities, skill certificate entities, and project experience entities from the resume text by using a named entity recognition algorithm, and classifying the extracted entities by using an information classification model to form structured resume information data. The information classification model is trained by using a training set containing annotated data of the resume text.

[0111] In some embodiments, the pre-set post competency model is used to assign weights to the parsed information to generate a resume analysis result including a matching score of the candidate and the target position, including: determining a set of position requirement keywords from the position description text of the target position by a keyword extraction algorithm, performing semantic matching between the structured resume information data and the set of position requirement keywords, and assigning a weight value to the matched resume information according to a pre-set weight rule; wherein the weight rule is based on the importance level of the position requirement keywords, and the matching score of the candidate and the target position is calculated by weighted summation.

[0112] In some embodiments, the post competency model is trained by a machine learning algorithm based on historical recruitment data, including: obtaining candidate resume data that passed the screening in historical recruitment as positive samples, and candidate resume data that did not pass the screening as negative samples, performing structured information extraction and position requirement keyword labeling on the positive and negative sample resume data to construct a training data set; using a support vector machine algorithm or a random forest algorithm to train the training data set to obtain a post competency model that can output a matching score of resume information and position requirements, and optimizing the model parameters by a cross-validation method.

[0113] In some embodiments, the parsed consultation information obtains the user's question intention, including: inputting the user's consultation information into a pre-trained intention classification model, the intention classification model is based on a Transformer neural network architecture and is trained by historical consultation data containing multiple labels such as enterprise information, position benefits, office environment, etc., to output the intention category corresponding to the user's question, the intention category at least includes enterprise basic information query, salary and welfare consultation, office environment understanding and position requirement confirmation.

[0114] In some embodiments, the question intention corresponding answer content is generated according to the pre-configured knowledge base of the enterprise and the optional external data query interface, including: retrieving a standard answer matching the user's question intention from the pre-configured knowledge base of the enterprise, if there is no matching answer in the knowledge base and the external data query interface is in the enabled state, obtaining the public information related to the question intention in the official website of the enterprise, or obtaining the industry general information through a pre-set third-party data interface, and generating the answer content after fusion processing of the knowledge base answer, the official website information of the enterprise and the third-party data.

[0115] In some embodiments, the semantic analysis of the communication record corresponding to the job seeker by the text classification algorithm to identify the job-seeking willingness level of the job seeker comprises: extracting a text passage from the communication record as input data, converting the text passage into a semantic vector by a word embedding model, inputting the semantic vector into a pre-trained job-seeking willingness classification model, the job-seeking willingness classification model adopts a convolutional neural network or a recurrent neural network architecture, the corresponding training data contains high, medium and low three grades of job-seeking willingness labels manually annotated in historical communication records, and the job-seeking willingness level of the current job seeker is obtained by model output.

[0116] In some embodiments, the difficulty and direction of the subsequent question are dynamically adjusted according to the real-time answer content of the job seeker during the interview process by a semantic similarity calculation algorithm, which comprises: performing semantic similarity calculation on the real-time answer text of the job seeker and a preset high-quality answer library, the semantic similarity calculation is based on a pre-trained language model to generate sentence vectors of the answer text and the high-quality answers, and the similarity score between the vectors is calculated by a cosine similarity algorithm; if the similarity score is higher than a first preset threshold, an advanced question of higher difficulty is selected, and if the similarity score is lower than a second preset threshold, a basic guiding question is selected, so as to adjust the difficulty and direction of the subsequent interview question.

[0117] In some embodiments, the key answer information recorded in the interview process is used to generate an interview comprehensive evaluation report to complete talent screening, which comprises: extracting the answer content, answer duration and semantic similarity score of the job seeker for each interview question from the interview process record, combining the matching degree score and the job-seeking willingness level, performing weighted calculation according to a preset evaluation rule, generating an interview comprehensive evaluation report containing the candidate comprehensive ability score, the post matching degree analysis and the job-seeking willingness evaluation, and completing the talent screening decision according to the score result of the evaluation report.

[0118] The application also provides a computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to make the processor realize the steps of the one-stop talent screening method provided by any one of the embodiments of the application.

[0119] The computer readable storage medium can be an internal storage unit of the computer device, such as a hard disk or a memory of the computer device. The computer readable storage medium can also be an external storage device of the computer device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc.

[0120] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A one-stop talent screening method, characterized in that: include: A natural language processing algorithm is used to parse the text of job applicant resumes, identifying structured information within them. The parsed information is weighted using a pre-defined job competency model to generate a resume analysis result that includes a score indicating the candidate's match with the target position. This job competency model is trained using a machine learning algorithm based on historical recruitment data and is used to quantitatively assess the degree of match between resume information and job requirements. Obtain job seekers' inquiries about company information, job benefits, and office environment, parse the inquiries to understand the user's question intent, and generate responses corresponding to the question intent based on the company's pre-configured knowledge base and the optionally enabled external data query interface; Performing semantic analysis on communication records corresponding to the job seeker using a text classification algorithm to identify the job seeker's level of job-seeking intention, wherein the training data for the text classification algorithm includes manually annotated job-seeking intention labels in historical communication records; Combining the matching score and the job-seeking intention level, personalized interview questions are dynamically generated through the interview question generation model; during the interview process, the difficulty and direction of subsequent questions are dynamically adjusted based on the real-time answers of the job seekers through the semantic similarity calculation algorithm, and the key response information during the interview process is recorded to generate a comprehensive interview evaluation report to complete the talent screening.

2. The method according to claim 1, characterized in that The method of using a natural language processing algorithm to parse the resume of a job seeker and identify structured information in the resume includes: The named entity recognition algorithm is used to extract education background entities, work experience entities, skill certificate entities and project experience entities from the resume text. The extracted entities are classified into categories using an information classification model to form structured resume information data. The information classification model is obtained through supervised learning training with a training set containing annotated resume text data.

3. The method according to claim 1, characterized in that The weighting of the parsed information is assigned through the preset job competency model to generate a resume analysis result containing a score of the candidate's match with the target position, including: Based on the job description text of the target position, a keyword extraction algorithm is used to determine the set of job requirement keywords, and the structured resume information data is semantically matched with the set of job requirement keywords. The matched resume information is assigned a weight value according to the preset weighting rules; The weighting rule is set based on the importance level of the job requirement keywords, and the matching score between the candidate and the target position is obtained through weighted sum calculation.

4. The method according to claim 1, wherein The job competency model is trained using a machine learning algorithm based on historical recruitment data and includes: Obtain the resume data of candidates who passed the screening in historical recruitment as positive samples, and the resume data of candidates who failed the screening as negative samples. Extract structured information and annotate the job requirement keywords from the positive and negative sample resume data to build a training dataset. The training data set is trained using a support vector machine algorithm or a random forest algorithm to obtain a job competency model that can output a score of the matching degree between resume information and job requirements, and the model parameters are optimized through a cross-validation method.

5. The method according to claim 1, wherein The analysis of the consultation information to obtain the user's question intention includes: The user consultation information is input into a pre-trained intent classification model. The intent classification model is built based on the Transformer neural network architecture and is trained with historical consultation data containing multiple categories of labels such as corporate information, job benefits, and office environment. The model outputs the intent category corresponding to the user's question. The intent category includes at least corporate basic information query, salary and benefits consultation, office environment understanding, and job requirements confirmation.

6. The method according to claim 1, characterized in that The generation of response content corresponding to the question intention based on the enterprise's pre-configured knowledge base and the optionally enabled external data query interface includes: Retrieve standard answers that match the user's question intent from the company's pre-configured knowledge base. If there is no matching answer in the knowledge base and the external data query interface is enabled, obtain public information related to the question intent from the corresponding company's official website, or obtain industry-wide information through the preset third-party data interface. After integrating the knowledge base answer, company official website information and third-party data, generate the response content.

7. The method according to claim 1, characterized in that The semantic analysis of the communication records corresponding to the job seeker by using a text classification algorithm to identify the job seeker's job application intention level includes: Text paragraphs are extracted from communication records as input data, and the text paragraphs are converted into semantic vectors through a word embedding model. The semantic vectors are input into a pre-trained job-seeking intention classification model. The job-seeking intention classification model adopts a convolutional neural network or a recurrent neural network architecture. The corresponding training data contains three levels of job-seeking intention labels of high, medium, and low manually annotated in historical communication records. The job-seeking intention level of the current job seeker is obtained through the model output.

8. The method according to claim 1, characterized in that During the interview process, the difficulty and direction of subsequent questions are dynamically adjusted based on the applicant's real-time answers through a semantic similarity calculation algorithm, including: The semantic similarity between the job seeker's real-time answer text and a preset library of high-quality answers is calculated. The semantic similarity calculation generates sentence vectors for the answer text and the high-quality answers based on a pre-trained language model, and the similarity score between the vectors is calculated using the cosine similarity algorithm; If the similarity score is higher than the first preset threshold, more difficult advanced questions will be selected. If it is lower than the second preset threshold, basic guiding questions will be selected to adjust the difficulty and direction of subsequent interview questions.

9. The method according to claim 1, characterized in that The key response information recorded during the interview process is used to generate a comprehensive interview assessment report to complete the talent screening, including: The content, duration and semantic similarity score of the applicant's answer to each interview question are extracted from the interview process record. Combined with the matching score and the level of job-seeking willingness, a weighted calculation is performed according to the preset evaluation rules to generate an interview comprehensive evaluation report that includes the candidate's comprehensive ability score, job matching analysis and job-seeking willingness assessment. Talent screening decisions are made based on the scoring results of the evaluation report.

10. A one-stop talent screening system, characterized by: include: A text parsing unit, configured to parse applicant resumes using a natural language processing algorithm, identify structured information in the resumes, assign weights to the parsed information using a preset job competency model, and generate a resume analysis result containing a score of the candidate's match with the target position. The job competency model is trained using a machine learning algorithm based on historical recruitment data and is used to quantitatively assess the degree of match between resume information and job requirements. The information acquisition unit is used to obtain the job seekers' consultation information about the company, job benefits, and office environment, parse the consultation information to obtain the user's question intention, and generate the corresponding response content based on the company's pre-configured knowledge base and the optionally enabled external data query interface; A semantic analysis unit, configured to perform semantic analysis on communication records corresponding to the job seeker using a text classification algorithm to identify the job seeker's level of job-seeking intention, wherein the training data for the text classification algorithm includes job-seeking intention labels manually annotated in historical communication records; The screening completion unit is used to dynamically generate personalized interview questions through the interview question generation model based on the matching score and the job-seeking willingness level; during the interview process, the difficulty and direction of subsequent questions are dynamically adjusted according to the real-time answers of the job seekers through the semantic similarity calculation algorithm, and the key response information during the interview process is recorded to generate a comprehensive interview evaluation report to complete the talent screening.

Images