Intelligent recruitment analysis method and system

Abstract

The invention provides an intelligent recruitment analysis method and system, and relates to the technical field of human resource technology and artificial intelligence. During resume evaluation, weights are dynamically re-allocated according to recruitment scenes, so that the recruitment matching precision and fairness are remarkably improved; when interview questioning quality analysis is carried out, quantitative scoring of multiple evaluation dimensions is carried out, and objective quantification of interview quality is realized; a message queue of the message middleware is combined with a state machine, so that asynchronous execution of tasks is realized, and the system stability and throughput are improved; through role difference view safety control, sensitive data difference visibility is realized, sufficient HR decision information is ensured, and psychology and data safety of interviewers are protected.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of human resources and artificial intelligence, and particularly to an intelligent recruitment analysis method and system. BACKGROUND

[0002] With the deepening of enterprise digital transformation, the recruitment management system (Applicant Tracking System, ATS) has become the core tool of human resource management. Modern ATS generally integrates AI (Artificial Intelligence) technology, and realizes functions such as resume screening, candidate matching and interview assistance through NLP (Natural Language Processing) and ML (Machine Learning) algorithms.

[0003] In the prior art, the evaluation model of ATS usually adopts a static weight system, and calculates a comprehensive score through a fixed formula for factors such as educational background, work experience and skill certificate. This unified evaluation mode is difficult to adapt to the differentiated needs of social recruitment and campus recruitment in talent evaluation dimensions - for example, campus recruitment needs to focus on academic achievements and internship experience, while social recruitment pays more attention to project experience and career continuity.

[0004] At the system architecture level, mainstream ATS adopts monolithic application design, and the evaluation calculation and result presentation are highly coupled. When different roles such as HR (Human Resources), department managers and others view the candidate evaluation results, the system will return the same AI analysis results, including directly displaying low score warnings. This presentation method lacking role adaptation is easy to cause cognitive bias or psychological resistance of interviewers, but if the evaluation details are completely hidden, it will also lead to the lack of improvement basis. Technically, complex semantic analysis (such as resume analysis and skill map construction) usually adopts a synchronous calling mode, and the front end needs to wait for all calculations to be completed before responding, which is easy to cause request timeout when the data volume is large, and lacks a task state tracking mechanism.

[0005] For the quality analysis of the interview link, the existing technology mainly relies on keyword frequency statistics or preset question template matching, which is difficult to quantitatively evaluate the structured degree, relevance to the post, neutrality in the questioning and other key dimensions of the interview questions. In terms of multi-language support, the system usually stores the original text in a single field, and the translation version lacks a unified version management mechanism, resulting in inconsistent translations in different modules (such as job descriptions and evaluation reports). In addition, the direct call of ATS and third-party services (such as video conferencing and voice transcription) has strong dependence, and a unified API (Application Programming Interface, Application Programming Interface) gateway layer is not established to handle authentication, traffic control and service degradation, which easily causes system cascade failure when the external service is abnormal. SUMMARY

[0006] The purpose of the present application is to provide an intelligent recruitment analysis method and system to at least solve one of the above problems.

[0007] In a first aspect, the present application provides an intelligent recruitment analysis method, comprising: Obtaining an analysis task request carrying a task type, generating a task identifier and an idempotent key corresponding to the analysis task request, and writing a task record into a task state storage table, and delivering a task message to a message middleware; According to the task type, the analysis process of the task consumed from the message middleware is executed, the generated structured result is written into the analysis result table, and the completion event of the corresponding task is published to the message middleware; wherein when the task type is interview question quality analysis, the analysis process includes preprocessing the interview question text, quantitatively scoring based on a plurality of predetermined evaluation dimensions, and generating classification results and improvement suggestions according to the scores of each evaluation dimension; when the task type is resume evaluation, the analysis process includes extracting key information from the position description and resume text for multi-dimensional matching, and dynamically adjusting the weight of each matching dimension according to the recruitment scene to calculate the total score and the credibility; the structured result includes a visible role set; Pushing the analysis completion notification to the user front end initiating the analysis task request in real time, so that the user front end initiates a result query request; When receiving the result query request, the analysis result corresponding to the result query request is field-cropped and desensitized according to the requester role in the result query request, and the differentiated result is returned.

[0008] In a second aspect, the present application provides an intelligent recruitment analysis system, which is used to implement the intelligent recruitment analysis method of any one of the preceding embodiments; the intelligent recruitment analysis system comprises a backend master server, a message middleware, an integrated gateway, an intelligent analysis engine, a data storage module and a real-time notification module; The backend master server is configured to obtain an analysis task request carrying a task type from the user front end, generate a task identifier and an idempotent key corresponding to the analysis task request, write a task record into a task state storage table through a data storage module, and deliver a task message to a message middleware; The intelligent analysis engine is configured to consume tasks from the message middleware, obtain corresponding text data through an integrated gateway, execute an analysis process on the consumed tasks according to a task type, write a generated structured result into an analysis result table through a data storage module, and publish a completion event of the corresponding task to the message middleware; The message middleware is configured to deliver the completion event to the backend master server; The backend master server is further configured to forward the completion event to a real-time notification module; The real-time notification module is configured to push an analysis completion notification to the user front end in real time, so that the user front end initiates a result query request to the backend master server; The backend master server is further configured to, when receiving the result query request, perform field pruning and desensitization on an analysis result corresponding to the result query request according to a requester role in the result query request, and return a differentiated result to the user front end.

[0009] The intelligent recruitment analysis method and system provided by the application, the method comprises the following steps: obtaining an analysis task request carrying a task type, generating a task identifier and an idempotent key corresponding to the analysis task request, and writing a task record into a task state storage table and delivering a task message to a message middleware; performing an analysis process on the tasks consumed from the message middleware according to the task type, writing the generated structured result into an analysis result table, and publishing a completion event of the corresponding task to the message middleware; wherein when the task type is interview question quality analysis, the analysis process comprises preprocessing the interview question text, performing quantitative scoring based on a plurality of predetermined evaluation dimensions, and generating a classification result and improvement suggestions according to the scores of the evaluation dimensions; when the task type is resume evaluation, the analysis process comprises extracting key information from the position description and the resume text for multi-dimensional matching, and dynamically adjusting the weights of the matching dimensions according to the recruitment scene to calculate the total score and the credibility; the structured result comprises a visible role set; the analysis completion notification is pushed to the user front end in real time to enable the user front end to initiate a result query request; when the result query request is received, the analysis result corresponding to the result query request is field-cropped and desensitized according to the requester role in the result query request, and the differentiated result is returned. In this way, when the resume is evaluated, the weights are dynamically redistributed according to the recruitment scene, which significantly improves the recruitment matching accuracy and fairness; when the interview question quality analysis is performed, the quantitative scoring of multiple evaluation dimensions is performed to realize the objective quantification of the interview quality; the message queue and the state machine of the message middleware are combined to realize the asynchronous execution of the task, which improves the system stability and throughput; through the role differential view security control, the sensitive data differential view is realized, which not only ensures the sufficient HR decision information, but also protects the psychological and data security of the interviewers. BRIEF DESCRIPTION OF DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0011] Figure 1 A flowchart of an intelligent recruitment analysis method provided by an embodiment of the present application; Figure 2 A structure diagram of an intelligent recruitment analysis system provided by an embodiment of the present application; Figure 3 A general architecture diagram of an intelligent recruitment analysis system provided by an embodiment of the present application; Figure 4 A timing flowchart of an intelligent recruitment analysis system provided by an embodiment of the present application.

[0012] Icons: 201 - backend master server; 202 - message middleware; 203 - integration gateway; 204 - intelligent analysis engine; 205 - data storage module; 206 - real-time notification module. DETAILED DESCRIPTION

[0013] The technical solutions of the present application will be described below in conjunction with embodiments. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0014] The existing recruitment management system has the following defects in resume screening and interview assistance: 1. Evaluation model is rigid: uniform weight is used for social recruitment and campus recruitment, and dimensions such as education, internship, scientific research, and work experience cannot be dynamically adjusted according to the scene, resulting in scene mismatch and talent misjudgment.

[0015] 2. The result presents a "single view": the system cannot differentially display AI results according to user roles (such as HR, interviewer, and administrator); directly displaying low scores to interviewers will cause resistance and psychological burden, and if not displayed, it will lose the basis for improvement.

[0016] 3. Calculation is synchronous coupling and easy to timeout: complex semantic analysis uses synchronous request-response, which blocks the front end and is easy to timeout; it lacks the engineering capabilities of task traceability, failure retry, and state visualization.

[0017] 4. Interview quality analysis dimension is single: it mainly stays in keyword counting or template matching, and it is difficult to quantify key dimensions such as structuredness, customization, effectiveness, neutrality, and clarity of questions.

[0018] 5. Multi-language operation is not friendly: text is stored in a single field, lacking unified translation management, version review, and cross-module reuse capabilities.

[0019] 6. Strong coupling of external dependencies: directly connected with external services such as meetings, speech transcription, and object storage, lacking authentication, traffic limiting, degradation, and caching strategies of integration gateway, and the system is not resilient enough.

[0020] Therefore, there is an urgent need for a systematic technical solution that covers adaptive evaluation, role differential view, asynchronous task closed loop, five-dimensional interview quality analysis, multi-language operation, and decoupling of external dependencies.

[0021] The embodiment of the application provides a kind of intelligent recruitment analysis method and system, it is related to natural language processing, big language model guide template, message middleware's asynchronous task processing, authority and view control, data security and observability etc., can realize adaptive AI evaluation and role difference view.

[0022] The embodiment of the application provides an intelligent recruitment analysis method, which can be executed by an electronic device with data processing capability. Figure 1 As shown in the flowchart of an intelligent recruitment analysis method, the method mainly includes the following steps S110 to S140: Step S110, obtain an analysis task request carrying a task type, generate a task identification and an idempotent key corresponding to the analysis task request, and write a task record into a task state storage table, and deliver a task message to a message middleware.

[0023] The user can send an analysis task request through a user front end, and the request content of the analysis task request can include the following structured information: task type, business identifier, idempotent key and optional parameters. The task type can be interview question quality analysis or resume evaluation, and other task types can be further extended. The business identifier is used to uniquely identify the associated business object, such as the ID of a certain interview or the ID of a certain candidate. The idempotent key can be a unique string generated by the user front end or the backend master server, which is used to prevent the creation of the same task repeatedly. Optional parameters are some analysis options that can be customized by the user, such as the upper limit of the number of questions. In an asynchronous message processing system, the idempotent key can be used to ensure that a message will not cause repeated changes to the system state even if it is consumed multiple times.

[0024] The above step S110 can be executed by a backend master server connected to the user front end. After the request receiving unit of the backend master server receives the analysis task request, it can first perform security check, which can include identity authentication and permission check. The identity authentication refers to verifying whether the access token carried in the request is valid to confirm the user identity. The permission check refers to checking whether the user has the right to access the resource represented by the business identifier (such as whether it has the right to view this interview).

[0025] After the security check passes, a unique identifier of the task (i.e., task identifier) can be generated and idempotency control can be performed. In specific implementation, a distributed unique ID generator can be called to assign a globally unique task identifier to the task. This identifier will serve as the unique credential of the task throughout its life cycle in the system. Idempotency control can be based on the combination of (user identifier, business identifier, task type, idempotency key) as the basis for judgment, and whether the same task request exists within a short window of time (e.g., 5 minutes) can be queried. If it exists, a new task will not be created, and the existing task identifier and current state will be immediately returned to the front end. If it does not exist, the process continues, and a new task is prepared to be created. This mechanism ensures that even in the case of network instability leading to repeated user submissions, duplicate tasks will not be created.

[0026] Task record persistence is then performed. Specifically, a new record can be inserted into a task state storage table in the database. This record contains a complete snapshot of the task: core metadata, request snapshot, and management information. Among them, the core metadata includes the task identifier, idempotency key, task type, business identifier, and task state (set to "to be processed" at this time). The request snapshot includes a copy of the de-identified request parameters for subsequent consumption or auditing. The management information includes the task creator, creation time, and update time, etc. Before persistence, personal sensitive information (such as name, phone number, company name) in the request snapshot can be de-identified and replaced with placeholders to meet data privacy protection requirements.

[0027] Next, asynchronous message delivery is performed. In specific implementation, a task message is first constructed: a lightweight, standardized task message can be constructed based on the task record; this message can only contain the minimum set of information (such as task identifier, business identifier, etc.) necessary for subsequent execution analysis processes. Then it is delivered to a message middleware: the message producer can publish this task message to a designated task topic; the message middleware (such as Kafka, RocketMQ) is responsible for receiving and storing the message. Reliability can be guaranteed in the following ways: 1. Delivery failure retry: if the message delivery fails due to network problems, etc., it can be automatically retried according to an exponential backoff strategy (such as intervals of 1 second, 2 seconds, 4 seconds).

[0028] 2. Final failure handling: if the retry exceeds the maximum number of times (e.g., 3 times) and is still unsuccessful, the state of the corresponding task in the task state storage table can be updated to "failed", and the specific error reason can be recorded. This task will enter the dead letter management process and require manual intervention for troubleshooting.

[0029] Once the task record is successfully written into the database, the request receiving unit returns an HTTP (Hypertext Transfer Protocol) 202 Accepted response to the front end immediately, regardless of whether the message has been successfully delivered to the message middleware. The response content can include a task identifier and a polling address (optional), wherein the task identifier is used to inform the user front end that the system has successfully accepted the task; and the polling address is used to provide an API address that can be used by the user front end to query the task progress.

[0030] In step S120, the task consumed from the message middleware is analyzed according to the task type, the generated structured result is written into the analysis result table, and the completion event of the corresponding task is published to the message middleware; wherein when the task type is interview question quality analysis, the analysis process includes preprocessing the interview question text, quantitatively scoring based on a plurality of predetermined evaluation dimensions, and generating a classification result and improvement suggestion according to the scores of each evaluation dimension; when the task type is resume evaluation, the analysis process includes extracting key information from the position description and resume text for multi-dimensional matching, and dynamically adjusting the weight of each matching dimension according to the recruitment scenario to calculate the total score and credibility; the structured result includes a visible role set.

[0031] The above step S120 can be performed by an intelligent analysis engine. The core task of the intelligent analysis engine is to consume tasks in the message queue, perform specific AI analysis, and save and notify the results. Specifically, step S120 can include three sub-steps: Sub-step 1, task distribution and identification: the intelligent analysis engine takes a task from the message middleware, first checks its task type, and determines whether it is an interview question quality analysis task or a resume evaluation task.

[0032] Sub-step 2, execute specific analysis process: If it is an interview question quality analysis, the intelligent analysis engine can: first clean and organize the interviewer's question text, then use an AI model to score from multiple dimensions (such as whether the question is professional, neutral, and clear), and finally give an overall evaluation (such as "excellent" or "to be improved") and generate specific improvement suggestions.

[0033] If it is a resume evaluation, the intelligent analysis engine can: extract key information from the resume and recruitment requirements for comparison, and intelligently adjust the scoring focus according to the recruitment scenario (such as social recruitment or school recruitment) (for example, school recruitment places more emphasis on educational background and internship experience), and finally calculates a matching total score and credibility.

[0034] Sub-step 3, save result and notify: after the analysis is completed, the intelligent analysis engine can store the structured result with the score, suggestion and other information into the database. At the same time, it will also send an "task completed" event to the message middleware to notify other parts of the system (such as the backend master server) to obtain the result.

[0035] Step S130, push the analysis completion notification to the user front-end that initiates the analysis task request in real time, so that the user front-end initiates a result query request.

[0036] After the AI analysis is completed, the user is actively and timely notified, so as to avoid the user manually refreshing the page frequently to wait for the result, greatly improving the user experience. In specific implementation, the backend master server always listens (subscribes) to the "task completion event" topic, and after capturing the event, it immediately interprets it, such as "oh, task XXX is completed, which is initiated by user ZZZ". Subsequently, the backend master server finds the current active browser session of the user through the real-time notification module (usually based on WebSocket or similar long connection technology), and sends a short "analysis completed" push message to it. The browser front-end (i.e. the user front-end) of the user receives the push message in real time through the established WebSocket connection. The message content may be: { "event": "analysis_done", "task_id": "XXX", "result_id": "YYY"}. After receiving the message, the user front-end automatically triggers the subsequent action: according to the result_id provided in the message, it immediately initiates an HTTP request to the backend master server to query the detailed analysis result.

[0037] Step S140, when receiving the result query request, field pruning and desensitization are performed on the analysis result corresponding to the result query request according to the role of the requester in the result query request, and a differentiated result is returned.

[0038] After the result query request reaches the backend master server, the identity and role of the current user can be verified, the complete analysis result is filtered and desensitized at the field level according to the visible role set stored in the result, and a safe differentiated result customized for the user is returned to the user front-end. Finally, the front-end interface is automatically updated, and the analysis result is clearly displayed to the user.

[0039] In some possible embodiments, the above step S140 can include: predefining a basic field set and a sensitive field set; when the requester role is a human resource manager or a system administrator, returning the union set of the basic field set and the sensitive field set in the corresponding analysis result; when the requester role is an interviewer, only returning the basic field set in the corresponding analysis result, and filtering out all fields of the sensitive field set in the corresponding analysis result.

[0040] Further, the differential result can be cached with the result identifier and the role bucket as the key, and a cache expiration time is set. Through the cache mechanism, repeated queries for the same analysis result and the same role can be quickly responded, thereby significantly reducing the database pressure, improving the system response speed, and ensuring the consistency of data.

[0041] The cache key is composed of (result identifier, role bucket), which can ensure the accuracy and security of the cache. The result identifier is the main dimension of the cache, and uniquely corresponds to a complete and uncut original analysis result (stored in the analysis result table), which ensures that the data source of the cache is accurate. The role bucket is the key to realizing the differential view, which is not a specific user ID, but classifies user roles into several predefined "buckets", such as hr_bucket (including HR, ADMIN roles) and interviewer_bucket (only including INTERVIEWER role). Using (result identifier, role bucket) as a composite key, independent cache entries can be created for different views (HR view, interviewer view) of the same analysis result, so that HR queries only hit the HR cache and interviewer queries only hit the interviewer cache, perfectly realizing data security isolation. Setting a reasonable survival time (for example, 15 minutes) for the cache data can balance the real-time data and the effectiveness of the cache.

[0042] The intelligent recruitment analysis method provided by the embodiment of the present application dynamically reallocates weights according to the recruitment scene when performing resume evaluation, significantly improves the recruitment matching accuracy and fairness; when performing interview question quality analysis, multiple evaluation dimensions are quantitatively scored, realizing objective quantification of interview quality; the message queue of the message middleware is combined with the state machine, realizing asynchronous task execution, improving system stability and throughput; through role differential view security control, sensitive data differential visibility is realized, which not only ensures sufficient HR decision information, but also protects the psychological and data security of interviewers.

[0043] In some possible embodiments, for interview question quality analysis, the pre-processing of the interview question text can include: 1. Question extraction and positioning.

[0044] Speaker separation: first, use the speaker label (such as "interviewer A" and "candidate B") provided by the speech transcription service to filter out all sentences marked as "interviewer".

[0045] Heuristic rule assistance: for text without a speaker label, rules can be used for inference: Question words and punctuation: Identify sentences that start with question words such as "what," "how," "why," or end with a question mark "?" Contextual relevance: Identify sentences that follow the candidate's answer and open a new topic.

[0046] 2. Text cleaning and normalization.

[0047] Remove redundant information: Filter out fillers such as "um," "ah," "that," repeated words, fillers such as "just say," "right?" and long pauses.

[0048] Text standardization: Convert full-width characters to half-width, unify English case, and standardize punctuation usage.

[0049] Merge and segment: For texts where the interviewer is interrupted and repeats the question, perform intelligent merging to form a complete question.

[0050] 3. Privacy information desensitization.

[0051] Before the text enters the analysis process, all personal sensitive information can be automatically identified and replaced. For example, replace the candidate's name, specific company name, phone number, email address, etc. with general placeholders (such as [candidate's name], [company name]) to strictly protect privacy and avoid model bias.

[0052] When quantitatively scoring based on a predetermined number of evaluation dimensions, for each purified question, a combination of large language models and rule engines can be used to score in the following five dimensions (usually using a 0-10 point system): 1. Structuring.

[0053] 1.1 Evaluation goal: Does the question guide the candidate to provide a structured and in-depth answer (such as using the STAR principle: situation, task, action, result).

[0054] 1.2 Implementation: Rule matching: Check if the question contains keywords such as "give an example," "what was the situation at the time," "what actions did you take."

[0055] LLM (Large Language Model) semantic judgment: Compare the question with standard structured questions such as "please describe a challenge you encountered and the solution process" in terms of semantic similarity, and let LLM judge its potential to guide structured answers.

[0056] Weighted sum of rule matching score and semantic judgment score to get the structure dimension score.

[0057] 2. Customization.

[0058] 2.1 Evaluation goal: Is the question tailored to the duties and requirements of the current position?

[0059] 2.2 Implementation: Keyword / skill matching: Extract the key skills and job responsibility requirements from the job description and perform keyword matching and semantic similarity calculation with the question text.

[0060] LLM context understanding: Input the job description fragments and the question into the LLM, and let the model judge "to what extent is this question aimed at testing the core ability required for this position".

[0061] Weighted sum of keyword / skill matching score and context understanding score to get the score of customization dimension.

[0062] 3. Effectiveness.

[0063] 3.1 Evaluation goal: Can the question effectively tap the candidate's real ability, experience or potential?

[0064] 3.2 Implementation: LLM depth analysis: LLM based on a large interview knowledge base to judge whether the question stays on the surface (such as "do you have good teamwork skills?") or can trigger behavioral examples (such as "please share an experience of failed team cooperation and your reflection"), and give higher scores to the latter.

[0065] 4. Neutrality.

[0066] 4.1 Evaluation goal: Does the question avoid involving discriminatory, suggestive or irrelevant personal topics? It can impose penalties or determine non-compliance on questions involving age, gender, marriage, region, religion, health and other compliance-sensitive areas; additional points will be deducted for suggestive, suggestive and discriminatory language.

[0067] 4.2 Implementation: Basic confidence sub-score base N (from LLM's judgment of "whether there is inappropriate pointing"): Let p safe ∈[0,1] be the safe probability of LLM in "compliance / non-compliance" binary classification (carried in output JSON), define the basic confidence sub-score base N = p safe .

[0068] Dual penalty term of dictionary and pattern matching penalty ex : Get the sensitive word hit ratehit rate Hit rate of suggestive / inductive phrases lead flag The weighted sum of the two yields a double penalty term, namely... penalty ex = α 1· hit rate + α 2· lead flag .

[0069] The formula for calculating the neutrality dimension score is as follows: subscore N = clip ( base N - penalty ex (0, 1). After subtracting the double penalty term from the basic confidence sub-score, the clip function is used to restrict the result to the interval [0, 1].

[0070] Severe non-compliance forces downgrade: If LLM output hard violation = true (i.e., a serious violation sign, such as directly inquiring about marital status, childbirth, or religious beliefs), then subscore N =0, and mark the violation code in the result (such as N_MARRIAGE violation of marital and reproductive status or N_RELIGION violation of religious beliefs, etc.).

[0071] 5. Clarity.

[0072] 5.1 Assessment objectives: Quantify grammatical correctness, syntactic complexity, terminology readability, and ambiguity risk.

[0073] 5.2 Implementation method: Readability metrics: Calculate average sentence length (words / phrase) ASL The proportion of complex clauses (number of complex clauses / total number of sentences, denoted as...) CC ratio ), Terminology density (industry jargon / general terminology, denoted as jargon ratio There are also structural ambiguity markers (such as multiple references): amb flag ∈{0,1} etc.

[0074] The various readability metrics are linearly fused and normalized into structural sub-components: struct clar = clip (1- β 1· norm z(ASL) -β 2· CC ratio - β 3· jargon ratio - β 4. amb flag ,0,1), where, norm z(ASL) This is the Z-score standardized value of the average sentence length. β 1. β 2. β 3. β 4 are all preset weights, for example, β 1 = 0.35 β 2 = 0.35, β 3 = 0.2 β 4 = 0.1.

[0075] LLM verification sub-segments llm clar LLM is categorized by its conciseness, grammatical correctness, and single-mindedness. llm clar ∈[0,1].

[0076] The score for the clarity dimension is: subscore C = γ · struct clar +(1- γ )· llm clar , γ =0.6.

[0077] After scoring each evaluation dimension, a structured scoring result can be generated, which includes the score for each evaluation dimension, the reason for the score, and key evidence (such as hitting JD clauses or hitting sensitive words).

[0078] In some possible embodiments, when performing the interview question quality analysis, the plurality of evaluation dimensions can include structuring, customization, effectiveness, neutrality, and clarity, the neutrality is provided with a preset compliance threshold; based on this, the generation of the classification result and the improvement suggestion according to the scores of the evaluation dimensions can include: when the score of the neutrality dimension is lower than the compliance threshold, the classification result is directly determined as the to-be-improved category; when the score of the neutrality dimension is not lower than the compliance threshold, the scores of the evaluation dimensions are weighted and summed to obtain a weighted total score; according to the relationship between the weighted total score and a preset threshold interval, the classification result is determined; a low-score dimension set whose score is lower than a corresponding suggestion threshold is identified; for each evaluation dimension in the low-score dimension set, a large language model is called to generate a targeted improvement suggestion, and the improvement suggestion is a text rewriting suggestion generated based on the original question text and the characteristics of the low-score dimension.

[0079] Exemplarily, the weights of the evaluation dimensions can be set as: the weight of customization is 0.30, the weight of effectiveness is 0.30, the weight of structuring is 0.20, the weight of clarity is 0.10, and the weight of neutrality is 0.10. The score of each evaluation dimension is score X ∈[0,10], which is normalized as score X =score_X / 10∈[0,1] after which the scores of the evaluation dimensions are weighted and summed to obtain the weighted total score.

[0080] In a possible implementation manner, for the case that the score of the neutrality dimension is not lower than the compliance threshold, the classification result can be determined in the following manner: if the weighted total score is greater than or equal to a first preset threshold, and the scores of the customization and effectiveness dimensions are both greater than or equal to a second preset threshold, the classification result is excellent; if the weighted total score is less than a third preset threshold, the classification result is to be improved; otherwise, the classification result is regular.

[0081] The compliance threshold, the first preset threshold, the second preset threshold, and the third preset threshold can all be set according to actual needs, which are not limited here. For example, if the value range of the score of a single dimension is [0, 10] and the value range of the weighted total score is [0, 1], the compliance threshold, the first preset threshold, the second preset threshold, and the third preset threshold can be 6, 0.8, 8, and 0.6 respectively.

[0082] To accurately locate the problem, after scoring the five dimensions (such as structure, effectiveness, neutrality, etc.) of the interview questions (for example, each dimension is scored 10 points), a screening action is also performed, that is, the score of each evaluation dimension is compared with the preset recommended threshold (for example, 7 points); all dimensions with a score lower than the recommended threshold are automatically screened out and classified into a low-score dimension set, so that targeted improvement suggestions can be generated. It should be noted that the recommended thresholds corresponding to different evaluation dimensions can be the same or different, and the present embodiment does not limit this.

[0083] For resume evaluation, the matching dimensions can include skill matching, experience matching, education background, and research / internship; based on this, dynamically adjusting the weights of the matching dimensions according to the recruitment scenario can include: determining whether the trigger condition of the campus recruitment scenario is met; wherein the trigger condition of the campus recruitment scenario includes that the position description is marked as campus recruitment, or the total work experience in the resume text is less than or equal to a preset year threshold; if it is met, a first weight set is used, wherein the weights of the education background and research / internship dimensions are set to be higher than the weight of the experience matching dimension; if it is not met, a second weight set is used, wherein the weight of the experience matching dimension is set to be the highest.

[0084] The above-mentioned year threshold can be set according to actual needs, for example, the year threshold is 1 year. Exemplarily, the first weight set can include: the weights corresponding to the education background, research / internship, skill matching, and experience matching are 0.35, 0.35, 0.20, and 0.10, respectively; the second weight set can include: the weights corresponding to the experience matching, skill matching, education background, and research / internship are 0.40, 0.30, 0.20, and 0.10, respectively.

[0085] When performing resume evaluation, information extraction and structuring are first performed. Hard requirements, skill keywords, job description, and priorities can be extracted from the position description, and the hard requirements can include education requirements (such as “bachelor’s degree or above”), minimum work experience requirements (such as “3-5 years of experience”), etc.; education background, work experience, skill list, and project / research / internship experience can be extracted from the resume text. Precise extraction and normalization can be achieved by combining named entity recognition and large language models.

[0086] Then, multi-dimensional matching and scoring are performed. The extracted information can be compared and scored in multiple dimensions, and the scoring methods of skill matching, experience matching, education background, and research / internship are introduced below.

[0087] The skill matching dimension can be achieved through semantic matching and synonym / superordinate and subordinate word hit rate; semantic matching uses a text embedding model to convert the position skill requirements and the resume skill list into vectors (i.e. Emb ( JDskills ), Emb ( CV skills The semantic similarity between the two is calculated using the following formula: sim emb = cosine ( Emb ( JD skills ), Emb ( CV skills )), sim emb ∈[0,1]; Synonym / hypernym hit rate hit syn Based on the skill map, the formula is: hit syn =(Number of synonyms / related words matched) / (Total number of skills in the job description). Skill matching score. S skill The calculation formula can be: S skill = clip ( η 1· sim emb + η 2· hit syn (0,1). For example, η 1 = 0.7 η 2 = 0.3.

[0088] Experience matching can be achieved through semantic alignment and structured strength. Semantic alignment compares the semantic similarity between the job descriptions in the resume and the job requirements in the job description, and the calculation formula is as follows: sim sem = cosine ( Emb ( CV expdesc ), Emb ( JD responsibilities The structured intensity maps role level, team / business size, and quantifiable output to [0,1], merging them into a unified whole. role strength Score for the experience matching dimension S exp The calculation formula can be: S exp =0.7· sim sem +0.3· role strength .

[0089] The education background dimension can be achieved by strong rule matching, for example, the JD (Job Description) requires a master's degree, and the candidate's education is a doctorate or a master's degree, which gets a high score, and a bachelor's degree, which gets a very low score. The score calculation method of the education background dimension can be: if the CV education is higher than the JD required education (for example: JD requires a master's degree, and the candidate is a master's or doctorate), S edu =1; if the CV education is only one level lower than the JD required education (for example: JD requires a master's degree, and the candidate is a bachelor's degree), then S edu =0.8; otherwise (for example: JD requires a master's degree, and the candidate is a junior college or below), S edu =0.4. An additional 0-0.1 reward cap can be added to "school level / major relevance / certification" ( clip to ensure that it does not exceed 1).

[0090] For the research / internship dimension, the research output score and the internship quality score can be weighted and summed to obtain the score of the research / internship dimension. The research output score (journal / conference level mapping to [0,1]) is denoted as paper score , the internship quality score (internship length, enterprise level, and job relevance) is denoted as intern score , and the score of the research / internship dimension is: S res =0.6· paper score +0.4· intern score .

[0091] Then, dynamic weight adjustment is made according to the recruitment scenario. See the foregoing related content for details.

[0092] Finally, the total score and credibility are calculated. The score of each dimension can be multiplied by its dynamically adjusted weight and then added to obtain the final total score (0-1 or 0-100). For example, the total score calculation formula for social recruitment is: Score soc =0.40· S exp +0.30· S skill +0.20· S edu +0.10· S res . The total score calculation formula for school recruitment is: Score camp =0.35· S edu +0.35·S res +0.20· S skill +0.10· S exp .

[0093] The credibility can be calculated by the following four evaluation factors ∈ [0, 1]: Completeness : Key field completeness (=1- \text{missing proportion}); Consistency : Multiple redundant information consistency, timeline self-consistency (detect conflict ratio and take 1-conflict ratio); Stability : Variance stability of split resampling or multiple prompt review (1-var_norm); ModelConf : LLM self-evaluation confidence or multi-model consistency score (normalized after integration).

[0094] The credibility calculation formula can be: Confidence =0.4· Completeness +0.2· Consistency +0.2· Stability +0.2· ModelConf .

[0095] If Confidence <0.60, it can be marked as "low credibility (need manual review)" in the results. When there is a hard constraint that is not met (such as a master's degree is required but only a bachelor's degree), a cap is applied to the total score (such as cap=0.75), and "hard constraint not met" is written in Explanations.

[0096] After generating the structured results, write them into the analysis result table. The process can include data assembly, atomic writing, and state updating. Among them, data assembly refers to the intelligent analysis engine assembling the results generated by all previous steps into a pre-defined data structure. This data structure can include: Association identification: task_id (unique ID of the task), used to associate with ai_task table (i.e. task state storage table); Evaluation data: dimension score set (such as interview five-dimensional score or resume matching each item score), final classification (such as "high quality" / "regular" / "to be improved"), improvement suggestion set, summary text, etc; Visibility control: visible role set, which is a key field that explicitly lists the role list that has the right to view the entire content of this result (such as ["HR", "ADMIN"]); Credibility and metadata: calculated credibility score, AI model version number used, prompt word version, etc. Audit information: creation timestamp, etc.

[0097] Atomic write means that all data is completely written into the ai_analysis_result table (i.e., the analysis result table) in one database transaction. Idempotency is guaranteed by taking the task_id as the primary key or unique constraint. In this way, even if the engine repeatedly receives the same message due to network jitter or the like, attempts to write multiple times, the database level will ensure that only the first one can succeed, and subsequent writes will fail due to primary key conflicts, thereby achieving "exactly-once processing in effect". State update means updating the record of the corresponding task_id in the ai_task table from "processing" to "success" in the same transaction or immediately after. This provides the final state for task state query and management.

[0098] After ensuring that the data (i.e., structured results) has been safely written to the database, the intelligent analysis engine needs to broadcast the "task completed" message to the outside world. In the specific implementation, event message construction is performed first. The intelligent analysis engine can construct a lightweight completion event message, which does not need to include the complete analysis result (because the result is already in the database), and the event content in the message can include the task identification, result identification, and user identification that initiated the task. Then, reliable publishing is performed. The intelligent analysis engine publishes the completion event message to the completion topic (e.g., ai_tasks_completed) of the message middleware. The message middleware will ensure that the completion event message is successfully delivered at least once.

[0099] In some possible embodiments, the above-mentioned intelligent recruitment analysis method further includes an idempotent and retry control mechanism: in the task creation stage, the generated idempotent key is used to ensure that the same analysis task request creates only one task within a preset time window; in the structured result writing stage, the task identification is used as the idempotent key to ensure that the structured result is written only once; the message middleware uses at least once delivery semantics and configures a retry queue and a dead letter queue for retry and fallback after processing failure. In this way, it can prevent task duplication, result duplication, or task loss caused by common problems in distributed systems such as network timeout, service restart, and message duplication, and ensure that each task is processed only once (or "exactly-once in effect").

[0100] For idempotency control of the task creation phase, see the foregoing corresponding content, which will not be repeated here. For idempotency control of the result writing phase, the application scenario can be: after the intelligent analysis engine processes the task, when writing the result, the response may be timed out due to network jitter; the intelligent analysis engine cannot determine whether the database is successfully written, so it may retry writing. The solution for this scenario can be: the ai_analysis_result table in the database sets the task_id as the primary key or unique index, and when the intelligent analysis engine tries to write the result, the database will guarantee that the record of the same task_id can be inserted only once.

[0101] For message processing and retry mechanism, the application scenario can be: when the message middleware delivers the task message, the processing may fail due to temporary failure of the consumer (intelligent analysis engine) or network problem. The solution for this scenario can be: "at least once" delivery semantics + retry queue + dead letter queue. "At least once" delivery means that the message middleware will guarantee that the message is delivered to the consumer at least once (may be repeatedly delivered due to retry). Retry queue means that when the consumer fails to process, the message will not be discarded immediately, but will be transferred to the retry queue, and the system will delay and deliver it to the consumer again according to the exponential backoff strategy (such as interval 1s, 2s, 4s, 8s...), giving the service recovery time. Dead letter queue means that if the message fails after the maximum number of retries (such as 3 times), it means that a serious error that cannot be automatically repaired is encountered, and the message will be transferred to the dead letter queue, waiting for manual intervention by the operation and maintenance personnel to troubleshoot.

[0102] In some possible embodiments, the intelligent recruitment analysis method further includes: collecting and recording key performance indicators during the processing of the analysis task request, the key performance indicators at least including task end-to-end processing latency, task failure rate, queue lag value, and dead letter rate; and automatically triggering an alarm when the key performance indicators exceed preset indicator thresholds in a continuous preset time period. By continuously monitoring the core indicators of the system, the stability, high performance, and maintainability of the system can be guaranteed.

[0103] The task end-to-end processing latency, task failure rate, queue lag value, and dead letter rate respectively reveal the health status of different levels of the system: the task end-to-end processing latency is the total time from when an analysis task is initiated by a user to when the user finally receives the results. The task end-to-end processing latency is the most direct indicator of user experience, and high latency means that users have to wait too long, which will directly lead to a decrease in user satisfaction; it reflects the comprehensive performance of the entire system link, including the network, business logic processing, and AI inference speed. The task failure rate is the proportion of tasks whose final state is "failed" among all processed tasks. The task failure rate directly reflects the reliability of the system, and a sharp increase in the failure rate usually means that there is a serious problem with a core component of the system (such as the AI service, database, or external dependency), resulting in a large number of tasks being unable to complete and business functions being impaired. The queue lag value is the number of messages in the message queue that have been produced but have not yet been consumed, also known as the consumption delay. The queue lag value can reflect whether the system throughput is balanced. The dead letter rate is the number of messages entering the dead letter queue per unit time. The dead letter rate can reflect deep system failures, and the entry of messages into the dead letter queue means that they have failed after multiple retries. A sudden increase in the dead letter rate usually points to a serious error that cannot be automatically recovered, such as a code logic defect (Bug), a major change in an external service API causing a call failure, a database schema (i.e., schema) change causing a write failure, etc.

[0104] The automatic alarm mechanism can operate in the following way: collect the latest values of the above-mentioned key performance indicators at a fixed frequency (such as every 15 seconds); compare the real-time data with the preset indicator thresholds, for example: the P95 quantile (i.e., the 95th percentile delay) of the end-to-end latency > 10 seconds, the task failure rate > 3%, the queue lag value > 1000, and the dead letter rate > 5 per minute; the triggering condition of the alarm is set to exceed the threshold for a continuous preset period of time (for example: continuously for 5 minutes, and the detection value of each minute exceeds the threshold), which can avoid false positives and improve the accuracy of the alarm; once the condition is met, the alarm information can be automatically sent to the operation and maintenance team through the preset channel (such as Dingding, WeChat for Enterprise, SMS, email, or one or more of them). The alarm information can include key context such as which indicator is abnormal, the current value, the start time of the abnormality, the possible scope of impact, etc., to help the operation and maintenance team quickly locate the problem.

[0105] In some possible embodiments, the task records, structured results, and differentiation results mentioned above are all stored through a data storage module. This module dynamically selects a storage scheme from a first and a second storage scheme based on the business characteristics of the fields. The first storage scheme is suitable for fields serving only a single entity. It stores multilingual text as embedded key-value pairs within the business entity's fields, with the format being a mapping from language identifiers to corresponding translated text. The second storage scheme is suitable for fields requiring unified cross-entity review, version management, or high-frequency reuse. It uses a separate multilingual translation table, which includes at least the entity category, unique entity identifier, field name, language identifier, and translated text content fields. This addresses the balance between storage efficiency, management complexity, and cross-module reusability for multilingual data in enterprise applications. It avoids the drawbacks of a "one-size-fits-all" storage approach through an intelligent, on-demand selection strategy.

[0106] The first storage scheme: embedded key-value pairs. The working method is as follows: within a master data record (such as a "Job" record), a dedicated area (field) is allocated to directly store the translations of all supported languages ​​in the format {"en-US":"Software Engineer","zh-CN":"Software Engineer"}. The advantages are: high read performance and simple structure.

[0107] The second storage solution: a separate translation table. This works by creating a separate database table specifically for storing all text fragments that need translation. Each record does not contain a complete business object, but rather a "term". Example record: (entity_type=job title, entity_id=P1001, field=name, locale=en-US, text=Software Engineer). Example record: (entity_type=job title, entity_id=P1001, field=name, locale=zh-CN, text=Software Engineer). Advantages include: easy unified management and reuse, support for advanced features, and reduced data redundancy.

[0108] In some possible embodiments, when consuming tasks from the message middleware, the corresponding text data is obtained through an integration gateway. The integration gateway module can provide a unified entry point for external dependencies accessing the conferencing system, speech-to-text service, and object storage service, and integrate at least one of the following resilience strategies: Circuit breaker strategy: Within a preset time window, when the error rate of calls to an external service exceeds a preset error rate threshold, temporarily stop sending requests to that external service and switch to a degradation scheme; Fallback strategy: automatically enable fallback solution when primary service is unavailable, in which, when speech transcription service is unavailable, fallback to read historical transcription text or prompt user to upload text manually; Cache strategy: cache external data whose read frequency is higher than a preset first frequency threshold and whose change frequency is lower than a preset second frequency threshold.

[0109] In this embodiment, all calls to external services are managed by a unified "smart agent" (integrated gateway); the core responsibility of this agent is not simply forwarding requests, but enhancing the "resilience" of the system to ensure the continuity of core business processes. The following introduces three resilience strategies: 1. Fuse strategy: fail fast, avoid avalanche.

[0110] The integrated gateway continuously monitors the request results of calling a certain external service (such as an ASR (Automatic Speech Recognition) service). Within a preset time window (such as 30 seconds), if the call error rate (such as timeout, 5xx error) exceeds the preset threshold (such as 50%), the integrated gateway will trigger the fuse to the "open" state. During the fuse period, all new requests to the service will immediately fail (intercepted by the integrated gateway) and will not be sent to the server that has failed. This avoids the system resources being occupied by a large number of pending requests, preventing the spread of failure. After a period of time, the fuse will enter the "half-open" state, and a request will be released tentatively. If successful, the fuse is closed and normal service is restored; if failed, the fuse state continues.

[0111] 2. Fallback strategy: bottom solution, guarantee core process.

[0112] When the gateway fails to call through the fuse strategy or directly, it is determined that a service is unavailable, and a previously defined fallback solution is automatically switched to. Specific case (speech transcription service unavailable): Primary service: call ASR service to convert conference recording to text.

[0113] Fallback solution: Solution A (read historical text): if the meeting has been transcribed before, directly read the stored historical transcription text for analysis; this ensures that the analysis work can still be performed; Solution B (manual upload): if there is no historical text, notify the user that "the speech transcription service is temporarily unavailable, please upload the text record directly", and convert the task to manual assistance mode.

[0114] 3. Cache strategy: reduce repeated requests, improve performance and resilience.

[0115] The gateway caches data that is frequently read and infrequently changed. Suitable data can include: High read frequency (higher than a first frequency threshold): For example, a JD description of a popular position is repeatedly used to analyze a large number of resumes; Low change frequency (lower than a second frequency threshold): For example, the JD description does not change every minute.

[0116] Subsequent requests directly obtain data from the cache, which is extremely fast and significantly reduces latency. When an external service (such as an object storage storing the JD) temporarily fails, the system can directly read the popular JD data from the cache, thereby being unaffected for a certain period of time and giving time for the recovery of the external service.

[0117] The embodiments of the present application also provide an intelligent recruitment analysis system for implementing the intelligent recruitment analysis method described above. As shown in Figure 2 The intelligent recruitment analysis system includes a backend master server 201, a message middleware 202, an integrated gateway 203, an intelligent analysis engine 204, a data storage module 205, and a real-time notification module 206.

[0118] The backend master server 201 is configured to obtain an analysis task request carrying a task type from a user front end, generate a task identifier and an idempotent key corresponding to the analysis task request, and write a task record into a task state storage table through the data storage module 205 and deliver a task message to the message middleware 202; The intelligent analysis engine 204 is configured to consume tasks from the message middleware 202, obtain corresponding text data through the integrated gateway 203, execute an analysis process on the consumed tasks according to a task type, write a generated structured result into an analysis result table through the data storage module 205, and publish a completion event of a corresponding task to the message middleware 202; The message middleware 202 is configured to deliver the completion event to the backend master server 201; The backend master server 201 is further configured to forward the completion event to the real-time notification module 206; The real-time notification module 206 is configured to push an analysis completion notification to the user front end in real time, so that the user front end initiates a result query request to the backend master server 201; The backend master server 201 is further configured to, when receiving the result query request, perform field pruning and desensitization on an analysis result corresponding to the result query request according to a requester role in the result query request, and return a differentiated result to the user front end.

[0119] The intelligent recruitment analysis system provided by the embodiment of the application can dynamically redistribute weights according to a recruitment scene when evaluating a resume, significantly improving recruitment matching accuracy and fairness; when analyzing interview question quality, multiple evaluation dimensions are quantitatively scored, achieving objective quantification of interview quality; a message queue of a message middleware is combined with a state machine, asynchronous task execution is achieved, and system stability and throughput are improved; through role differential view security control, sensitive data differential visibility is achieved, which guarantees sufficient HR decision information and protects the psychological and data security of interviewers.

[0120] Further, the multiple evaluation dimensions include structure, customization, effectiveness, neutrality and clarity, the neutrality is provided with a preset compliance threshold; the intelligent analysis engine 204 is specifically configured to: when the score of the neutrality dimension is lower than the compliance threshold, directly determine the classification result as an improvement-to-be-improved category; when the score of the neutrality dimension is not lower than the compliance threshold, weighted sum of the scores of each evaluation dimension is performed to obtain a weighted total score; according to a relationship between the weighted total score and a preset threshold interval, the classification result is determined; a low-score dimension set whose score is lower than a corresponding suggestion threshold is identified; for each evaluation dimension in the low-score dimension set, a large language model is called to generate a targeted improvement suggestion, and the improvement suggestion is a text rewriting suggestion generated based on an original question text and a characteristic of the low-score dimension.

[0121] Further, the multiple evaluation dimensions include structure, customization, effectiveness, neutrality and clarity, the neutrality is provided with a preset compliance threshold; the intelligent analysis engine 204 is specifically configured to: when the score of the neutrality dimension is lower than the compliance threshold, directly determine the classification result as an improvement-to-be-improved category; when the score of the neutrality dimension is not lower than the compliance threshold, weighted sum of the scores of each evaluation dimension is performed to obtain a weighted total score; according to a relationship between the weighted total score and a preset threshold interval, the classification result is determined; a low-score dimension set whose score is lower than a corresponding suggestion threshold is identified; for each evaluation dimension in the low-score dimension set, a large language model is called to generate a targeted improvement suggestion, and the improvement suggestion is a text rewriting suggestion generated based on an original question text and a characteristic of the low-score dimension.

[0122] Further, the backend master server 201 is specifically configured to: predefine a basic field set and a sensitive field set; when the requester role is a human resource administrator or a system administrator, return a union set of the basic field set and the sensitive field set in the corresponding analysis result; when the requester role is an interviewer, only return the basic field set in the corresponding analysis result, and filter all fields of the sensitive field set in the corresponding analysis result.

[0123] Further, the backend master server 201 is further configured to: cache the differential result with a result identifier and a role bucket as keys, and set a cache expiration time.

[0124] Further, the backend master server 201 is further configured to: in the task creation stage, ensure that only one task is created for the same analysis task request within a preset time window through the generated idempotent key; in the structured result writing stage, use the task identification as the idempotent key to ensure that the structured result is written only once; and the message middleware 202 is further configured to use at least one delivery semantics and configure a retry queue and a dead letter queue to handle retries and bottom out after processing failures.

[0125] Further, the backend master server 201 is further configured to: in the process of processing the analysis task request, collect and record key performance indicators, the key performance indicators at least including a task end-to-end processing delay, a task failure rate, a queue lag value and a dead letter rate; and automatically trigger an alarm when the key performance indicators exceed preset indicator thresholds within a continuous preset time period.

[0126] Further, the intelligent recruitment analysis system further comprises a data storage module, the task record, the structured result and the differentiated result are stored through the data storage module, and the data storage module is configured to dynamically select a storage scheme from a first storage scheme and a second storage scheme according to the business characteristics of a field; the first storage scheme is applicable to a field serving only a single entity, and the first storage scheme stores multilingual text in the form of an embedded key-value pair in the field of the business entity, and the format is a mapping of a language identifier to a corresponding translated text; and the second storage scheme is applicable to a field that needs cross-entity unified proofreading, version management or high-frequency reuse, and the second storage scheme stores using an independent multilingual translation table, and the multilingual translation table at least includes an entity category, an entity unique identifier, a field name, a language identifier and a translated text content field.

[0127] Further, the intelligent recruitment analysis system further comprises an integration gateway, the intelligent analysis engine 204 obtains corresponding text data through the integration gateway when consuming from the message middleware 202, the integration gateway module provides a unified entrance for accessing external dependencies such as conference systems, speech transcription services and object storage services, and integrates at least one of the following resilience strategies: a fuse strategy: when the error rate of a call to an external service exceeds a preset error rate threshold within a preset time window, temporarily stop sending requests to the external service, and switch to a degradation scheme; a degradation strategy: when the main service is unavailable, automatically enable a backup scheme, wherein when the speech transcription service is unavailable, degrade to read historical transcription text or prompt manual upload of text; a cache strategy: cache external data with a read frequency higher than a preset first frequency threshold and a change frequency lower than a preset second frequency threshold.

[0128] The intelligent recruitment analysis system provided by the embodiment has the same implementation principle and technical effects as the foregoing intelligent recruitment analysis method embodiment, and for brief description, the parts not mentioned in the intelligent recruitment analysis system embodiment can refer to the corresponding contents in the foregoing intelligent recruitment analysis method embodiment.

[0129] For the convenience of understanding, the foregoing intelligent recruitment analysis method and system are introduced in detail below.

[0130] The embodiment of the application realizes a high-concurrency traceable AI recruitment analysis process through an asynchronous task architecture, and realizes precise evaluation, safe display and stable execution in interview quality analysis and resume matching by combining a scene-adaptive weight algorithm and a role differential view control, thereby effectively improving the intelligence and compliance level of enterprise recruitment.

[0131] The embodiment provides an adaptive AI evaluation method, which can dynamically redistribute evaluation weights according to scenes such as school recruitment and social recruitment; a multi-dimensional semantic interview quality analysis method, which realizes quantitative scoring and suggestion generation for the structure, customization, effectiveness, neutrality and clarity of questions; a role-based differential view method, which outputs differential fields to different roles under the premise of ensuring data security and compliance; a robust asynchronous AI task processing architecture, which has idempotency, retry, state machine, message notification, real-time push and observability; a multi-language data storage scheme and external dependency integration gateway, which improves cross-regional collaboration and system resilience.

[0132] The intelligent recruitment analysis method provided by the embodiment of the application mainly includes the following processes: S1 request acceptance: a request receiving unit receives an analysis task request (task type contains “interview question quality analysis” or “resume evaluation”) from a user front end, verifies an identity authentication credential and a permission, generates a task identifier and an idempotent key.

[0133] S2 task queuing: a task record is written into an AI task state storage table (containing state = to be processed, retry parameter, request snapshot), and the task message is delivered to the task topic of the message middleware; “has accepted” information (containing the task identifier) is immediately returned to the user front end.

[0134] S3 model arrangement: an intelligent analysis engine subscribes to the task topic and consumes the task: When the type is interview question quality analysis, the pre-processing-five-dimensional scoring-three classification-suggestion generation process is executed; When the type is resume evaluation, the JD / resume key information extraction-multi-dimensional matching-scene-adaptive weight adjustment-total score and reliability calculation process is executed.

[0135] S4 Result write: write the algorithm result in a predetermined format data structure to the AI analysis result table, including fields such as "visible role set", "dimension score set", "credibility", "suggestion set", etc.; task status is set to success.

[0136] S5 Complete event: the intelligent analysis engine publishes a "task completion event" (including task identification, result identification, and initiator user) on a completion topic.

[0137] S6 Real-time notification: the backend master server subscribes to the completion topic and pushes "analysis completion" and result identification to the initiator user through the real-time notification module.

[0138] S7 Differential query: the user front end initiates a result query after receiving the notification; the role view control module of the backend performs field-level pruning and desensitization return according to the requester role and the "visible role set" in the result record; at the same time, cache and audit records are executed.

[0139] The key enhancement mechanisms of the above process design include: Idempotent and retry: the task creation requires carrying an idempotent key; the result is stored in the database with the task identification as the idempotent key; the message uses "at least once" semantics, and the failure enters the retry queue, exceeding the maximum retry is failed; Observation and SLA (Service Level Agreement): record request_id / task_id in the whole link, and statistics reasoning delay, success rate, and timeout rate; set the P95 delay target (such as interview ≤8s, resume ≤5s), and trigger alarm when threshold is reached; External dependency degradation: the integration gateway provides authentication, flow limiting, retry, cache, and degradation; when ASR is unavailable, it can fall back to the stored text for further analysis; Security and privacy: desensitization before storing the transcribed and resume texts; access to the original model output is hierarchical; query returns the minimum necessary; full audit.

[0140] The system composition of the above intelligent recruitment analysis system can be: Request receiving unit, task arrangement unit (in the backend master server); message middleware (task topic and completion topic); intelligent analysis engine (carrying algorithm process); result storage unit and task status storage table (database); real-time notification module (persistent connection push); role view control module (field pruning and cache); integration gateway (conference, speech transcription, object storage, and unified integration of external interfaces); multi-language management unit (translation management and version proofreading).

[0141] Compared with the prior art, the embodiments of the present application have the following significant technical effects: Adaptive algorithm mechanism: the model can dynamically reallocate weights according to the scene, significantly improving the accuracy and fairness of recruitment matching. Multi-dimensional intelligent evaluation: through five-dimensional semantic analysis and neutral guardrails, the objective quantification and compliance constraints of interview quality are realized. High-concurrency asynchronous architecture: the combination of message queues and state machines realizes AI task asynchronous execution, idempotent database writing, retry, and real-time feedback, improving system stability and throughput. Role differential view security control: sensitive data differential visualization is realized, ensuring sufficient HR decision-making information while protecting the psychological and data security of interviewers. Full-link traceability and audit compliance: the system has built-in logs, indicators, and access audit mechanisms, providing verifiable evidence for model optimization and compliance supervision. Extensibility and strong versatility: modular architecture suitable for multi-language, multi-position, and multi-model fusion intelligent recruitment scenarios.

[0142] Figure 3 The system's 7 core submodules are shown: user front-end, back-end main server, message middleware, intelligent analysis engine, integration gateway, data storage module, and real-time notification module. All AI analysis results are asynchronously decoupled through message middleware. Differential results are dynamically generated by the back-end aggregation engine based on user roles. The data link forms a closed loop: initiation → analysis → notification → display → audit.

[0143] Figure 4 The time sequence flowchart shows the entire process from the user clicking "Start Analysis" to the user front-end receiving the results. The asynchronous mechanism is reflected in the task queuing (5) and event notification (11-14). All links have idempotent mechanisms and state writing to ensure consistency. The notification link has two channels: WebSocket real-time push + HTTP polling fallback.

[0144] The overall architecture and module functions of the system are introduced below.

[0145] (1) User front-end.

[0146] 1) Responsibilities: page presentation (position / candidate / interview), initiating analysis tasks, receiving real-time notifications, and rendering HR / interviewer views according to differential rules.

[0147] 2) Inputs: user operations and form data; authentication credentials (access tokens) are read from browser secure storage and sent with the request.

[0148] 3) Outputs: Submit analysis task request data structure (including task type, business identifier, optional parameters) to "Request Receiving Unit".

[0149] Subscription to the "real-time notification module" push complete event data structure (including task identification, result identification, timestamp).

[0150] 4) State and constraints: Backoff retry (1s→2s→4s, max 5 times) when network is interrupted to submit requests; WebSocket disconnection automatically reconnects (heartbeat 20-30s), and pulls the latest unread complete events (by timestamp) after reconnection.

[0151] 5) Failure mode: Request 401 / 403: Trigger re-authentication; 202 accepted but not completed: Show "task in progress" and optional polling button.

[0152] 6) Observation index: End-to-end latency from user front-end initiation to receiving completion notification; failure popup times; reconnection times.

[0153] (2) Backend master server.

[0154] 1) Responsibilities: Identity authentication and authorization, request reception and parameter verification, task orchestration (generate task identification / idiom key, write task table, deliver message), result aggregation and view clipping, multi-language management, audit and log.

[0155] 2) Input: Analysis task request data structure submitted by the front end, including: task_type ∈ {INTERVIEW_QA, RESUME_EVAL}, that is, task type includes interview question quality analysis or resume evaluation; Business identification: interview_id / candidate_id / job_id (that is, interview ID / candidate ID / job ID, take one or more according to type); Optional options (such as whether to generate abstracts, maximum question number limit, etc.); Access token (parse to obtain user identification and role set).

[0156] 3) Output: Task acceptance receipt (202, including task_id, optional polling address); Write ai_task task record; Deliver task message to message middleware "task topic"; After handling the completion event, submit completion notification (including result_id) to the "real-time notification module".

[0157] 4) Key constraints: Idempotency: Same (requester, business id, task_type) + idempotency_key (i.e. idempotency key) creates a task only once in a "short time window"; Authentication: Must check that the requester has access to the business object (position / interview / candidate); Sanitization: Replace PII (Personally Identifiable Information) fields with placeholders before writing to the database.

[0158] 5) Failure mode: Parameter error returns 4xx (does not create a task); Message delivery failure enters the retry channel, exceeds the upper limit, writes "dead letter queue", and puts the task in "failure".

[0159] 6) Observation index: QPS (Queries Per Second), task creation success rate, entry failure rate, P95 handling delay, 403 / 422 ratio.

[0160] (3) Intelligent analysis engine.

[0161] 1) Responsibility: Consume tasks, execute "interview quality analysis" or "resume evaluation" algorithm chain, write back results and publish completion events.

[0162] 2) Input: Task message data structure in "task topic" (including task_id, business id, options, requester, and link tracking identifier); Pull transcribed text (including speaker labels) / position JD / resume text through "integration gateway module".

[0163] 3) Output: ai_analysis_result structured result record (including dimension score, classification, confidence, visible role set, model version); Update ai_task status to "success / failure"; Publish completion events to "completion topic".

[0164] 4) Execution constraints: Algorithm processing must be completed within a timeout threshold (e.g. 90s); Retry and degrade LLM / embedded services; Strictly follow the output format constraints (predefined data structure), if not met, judge as failure and write error_code.

[0165] 5) Failure mode: External dependency unavailable, model returns unqualified structure, input missing.

[0166] 6) Observation index: Model inference time-consuming, success rate, retry rate, number of downgrades, result writing library failure rate.

[0167] (4) Message middleware.

[0168] 1) Duty: Provide task topics and complete topics; "at least once" delivery, dead letter and retry queue.

[0169] 2) Parameter suggestion: Task topic: The number of partitions is configured according to the target concurrency x 2 (the target concurrency refers to the highest task processing amount per second that the system expects to reach in design, and in this embodiment, the number of partitions should be set to twice the target concurrency value); Retention period ≥ 24h; Retry strategy: Exponential backoff (1s, 2s, 4s, … to the upper limit); Maximum number of retries = 3; Dead letter queue: Carrying task_id, error_code, last_attempt_at (i.e. last attempt time).

[0170] 3) Consistency explanation: Achieve exactly once processing under the effect of at least once delivery through idempotent writing library (with task_id as the unique key).

[0171] (5) Integration gateway.

[0172] 1) Duty: Unified access to conference system, speech transcription, object storage, etc.; Implement authentication, flow limiting, retry, fuse, downgrade, cache.

[0173] 2) Strategy: Authentication: Token refresh and key rotation; Flow limiting: QPS per dependency; Fuse: Open when error rate > 15%, close after N successful half-open probes; Downgrade: ASR unavailable → read historical transcription; Object storage unavailable → local staging + asynchronous retransmission; Cache: Short-term cache JD / resume standardized text, TTL 15-30 minutes.

[0174] 3) Audit: Record external dependency call time-consuming, error and retry times.

[0175] (6) Real-time notification module.

[0176] 1) Duty: Push service based on persistent connection (such as WebSocket), route "complete event" to the user session initiator.

[0177] 2) Constraints: Heartbeat (20-30s) connection, after the service side to retain the last completed event 5-10 minutes for reissue; Push failure into the queue, the next time the client is online to push.

[0178] 3) Observation index: online connection number, message delivery success rate, average delivery delay.

[0179] (7) Data storage module.

[0180] 1) Contains: ai_task, ai_analysis_result, translation table and embedded multi-language field, audit log and index library (record access and model / system index).

[0181] 2) Physical design: it is recommended to partition by time or organization dimension; large fields (such as raw_model_output) and hot fields are divided into tables / columns to avoid affecting transaction path; archiving and retention policy (such as results for 180 days, audit for 365 days).

[0182] (8) Operation guarantee.

[0183] Idempotent control: the front end provides idempotency_key; the back end judges by (requester, business key, task_type, idempotency_key).

[0184] Retry limit: up to 3 times of consumption failure; Timeout scanning: background inspection ai_task, exceeding the threshold is "timeout"; Link tracking: throughout request_id / task_id; SLA index: P95 delay and failure rate monitoring; Alarm strategy: alarm when P95> target threshold for 5 consecutive minutes, or failure rate> threshold.

[0185] The following introduces the core data structure and storage scheme.

[0186] (1) Multi-language storage scheme.

[0187] 1) Scheme A (embedded key-value pair set): Field: process_name_i18n (predefined format: key-value pair of language→text).

[0188] Consistency: atomic update in place in the same transaction; Retrieval: suitable for business flow reading by primary key; weak cross-entity query.

[0189] 2) Solution B (independent translation table): Fields: unique identifier, entity category, entity identifier, field name, language identifier, text content, version number, creation / modification time; Index: (entity_type, entity_id, field, locale) combined index; optional full-text index; Review and rollback: ver field drives multi-version preservation and rollback; Cache strategy: hot words are cached for 15-60 minutes according to (entity, field, locale), and invalidation uses random jitter to avoid avalanches.

[0190] 3) Selection suggestion: fields that need to be reused / audited across modules use solution B; the rest use solution A, and keep A B's migration path (migration tools import and export in batches according to primary keys, and preserve ver).

[0191] (2) AI task status storage table (ai_task).

[0192] 1) Key and foreign key: Primary key / identifier: task_id (unique), idempotency_key (idempotent), trace_id (link tracking); Business foreign key: position, candidate, interview association identifier (nullable on demand).

[0193] 2) State machine field: pending / running / success / failed / timeout / canceled; 3) Fault tolerance and control: priority, current retry count, maximum retry count, expiration time (timeout threshold); 4) Snapshot: task load snapshot (predefined format, de-identification storage); 5) Audit: creator, timestamp, error code and error information, time consumption in milliseconds; 6) Index: (status, task_type, updated_at) composite index to support scheduling and inspection; task_id unique index; (created_at) partition key can be added to support time partitioning and archiving; 7) Constraints: set check constraints for retry count and expiration time; make unique constraints for idempotency_key + business key (short window table or unique index).

[0194] 8) Query mode: Scheduler: Scan status in (pending, failed) and updated_at within window. Trace page: Exact query by trace_id / task_id.

[0195] (3) AI analysis result table (ai_analysis_result).

[0196] 1) Association: result_id (primary key), task_id (foreign key), result_type (aligned with task type).

[0197] 2) Visibility control: view_role (visible role set, such as HR / INTERVIEWER / ADMIN), set type field (array / bit map / set) optional.

[0198] 3) Evaluation data: dimension score set (predefined format), classification (excellent / regular / needs improvement or matching level), suggestion set, summary text, credibility value; 4) Model metadata: model_ver, prompt_ver, embedding_ver; 5) Original output: optional raw_model_output (limited access / table storage / object storage reference); 6) Audit fields: creator and timestamp; 7) Index: task_id (foreign key retrieval); (result_type, created_at) (reporting / playback); dimension score structured index, can use JSONB+GIN (or column store) to support "filtering by dimension interval" (e.g. find samples with neutrality <6 for compliance playback); 8) Partition and retention: monthly partition; results retained for 180 days, audits for 365 days; original output can be migrated to cold storage.

[0199] 9) Privacy compliance: field-level desensitization for PII before storage; original text can only be viewed by HR / ADMIN through audit authorization.

[0200] 10) Query mode: Front-end details: exact retrieval by task_id / result_id; Report: filter by time and result_type; Training / review: filter by "low credibility / compliance hit".

[0201] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to perform the intelligent recruitment analysis method in the foregoing method embodiment. The computer readable storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a RAM, a magnetic disk or an optical disk, and various media capable of storing program codes.

[0202] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent recruitment analysis method, characterized in that, include: Obtain an analysis task request carrying a task type, generate a task identifier and idempotent key corresponding to the analysis task request, write the task record into the task status storage table, and deliver the task message to the message middleware. The analysis process is executed on tasks consumed from the message middleware according to the task type, the generated structured results are written into the analysis result table, and the completion event of the corresponding task is published to the message middleware. Specifically, when the task type is interview question quality analysis, the analysis process includes preprocessing the interview question text, quantitatively scoring it based on multiple predetermined evaluation dimensions, and generating classification results and improvement suggestions based on the scores of each evaluation dimension. When the task type is resume evaluation, the analysis process includes extracting key information from the job description and resume text for multi-dimensional matching, and dynamically adjusting the weights of each matching dimension according to the recruitment scenario to calculate the total score and credibility. The structured results include a set of visible roles. A notification of analysis completion is pushed to the user frontend that initiated the analysis task request in real time, so that the user frontend can initiate a result query request. When the result query request is received, the analysis results corresponding to the result query request are cropped and anonymized according to the requester role in the result query request, and the differential results are returned.

2. The intelligent recruitment analysis method according to claim 1, characterized in that, The multiple evaluation dimensions include structure, customization, effectiveness, neutrality, and clarity, with the neutrality set with a preset compliance threshold; The process of generating classification results and improvement suggestions based on the scores of each evaluation dimension includes: When the score of the neutrality dimension is lower than the compliance threshold, the classification result is directly identified as the category to be improved. When the score of the neutrality dimension is not lower than the compliance threshold, the scores of each evaluation dimension are weighted and summed to obtain a weighted total score; the classification result is determined based on the relationship between the weighted total score and the preset threshold range. Identify the set of low-scoring dimensions whose scores are below the corresponding suggested threshold; For each evaluation dimension in the set of low-scoring dimensions, a large language model is invoked to generate targeted improvement suggestions. These improvement suggestions are text rewriting suggestions generated based on the original question text and the characteristics of that low-scoring dimension.

3. The intelligent recruitment analysis method according to claim 1, characterized in that, The matching dimensions include skills matching, experience matching, educational background, and research / internship; the dynamic adjustment of the weights of each matching dimension according to the recruitment scenario includes: Determine whether the triggering conditions for the campus recruitment scenario are met; wherein, the triggering conditions for the campus recruitment scenario include the job description being marked as campus recruitment, or the total years of work experience in the resume text being less than or equal to a preset years threshold. If satisfied, the first weight set is adopted, wherein the weights of the educational background and the research / internship dimension are set higher than the weight of the experience matching dimension. If the conditions are not met, a second set of weights is used, wherein the weight of the experience matching dimension is set to the highest.

4. The intelligent recruitment analysis method according to claim 1, characterized in that, The step of cropping and anonymizing the analysis results corresponding to the result query request according to the requester role in the result query request, and returning differential results, includes: Predefined sets of basic fields and sensitive fields; When the requester's role is a human resources administrator or a system administrator, return the union of the basic field set and the sensitive field set in the corresponding analysis results; When the requester is an interviewer, only the set of basic fields in the corresponding analysis results is returned, and all fields in the set of sensitive fields in the corresponding analysis results are filtered out.

5. The intelligent recruitment analysis method according to claim 1, characterized in that, The intelligent recruitment analysis method also includes: The differential results are cached using the result identifier and role bucket as keys, and a cache expiration time is set.

6. The intelligent recruitment analysis method according to claim 1, characterized in that, The intelligent recruitment analysis method also includes idempotency and retry control mechanisms: During the task creation phase, the generated idempotent key ensures that only one task is created for the same analysis task request within a preset time window. During the structured result writing phase, the task identifier is used as an idempotent key to ensure that the structured result is written only once. The message middleware employs at least one delivery semantics and is configured with a retry queue and a dead letter queue for retries and fallbacks after processing failures.

7. The intelligent recruitment analysis method according to claim 6, characterized in that, The intelligent recruitment analysis method also includes: During the processing of the analysis task request, key performance indicators are collected and recorded. These key performance indicators include at least the task end-to-end processing latency, task failure rate, queue lag value, and dead letter rate. An alarm is automatically triggered when the key performance indicator exceeds a preset threshold within a consecutive preset time period.

8. The intelligent recruitment analysis method according to claim 1, characterized in that, The task record, the structured result, and the differential result are all stored through a data storage module, which is used to dynamically select a storage scheme from the first storage scheme and the second storage scheme according to the business characteristics of the field. The first storage scheme is applicable to fields that serve only a single entity. The first storage scheme stores multilingual text in the field of the business entity in the form of embedded key-value pairs, with the format being a mapping from language identifier to corresponding translated text. The second storage scheme is applicable to fields that are uniformly reviewed across entities, version managed, or frequently reused. The second storage scheme uses an independent multilingual translation table for storage. The multilingual translation table at least includes the entity category, entity unique identifier, field name, language identifier, and translated text content field.

9. The intelligent recruitment analysis method according to claim 1, characterized in that, When consuming data from the message middleware, the corresponding text data is obtained through the integration gateway. The integration gateway module provides a unified entry point for external dependencies such as the conferencing system, speech-to-text service, and object storage service, and integrates at least one of the following resilience strategies: Circuit breaker strategy: Within a preset time window, when the error rate of calls to an external service exceeds a preset error rate threshold, temporarily stop sending requests to that external service and switch to a degradation scheme; Degradation strategy: When the primary service is unavailable, the backup plan is automatically activated. When the speech-to-text service is unavailable, the service is downgraded to reading historical transcribed text or prompting manual text upload. Caching strategy: External data whose read frequency is higher than a preset first frequency threshold and whose change frequency is lower than a preset second frequency threshold is cached.

10. An intelligent recruitment analysis system, characterized in that, The intelligent recruitment analysis system is used to implement the intelligent recruitment analysis method according to any one of claims 1-9; the intelligent recruitment analysis system includes a backend main server, a message middleware, an integration gateway, an intelligent analysis engine, a data storage module, and a real-time notification module; The backend master server is used to obtain analysis task requests carrying task types from the user frontend, generate task identifiers and idempotent keys corresponding to the analysis task requests, and write task records into the task status storage table through the data storage module and deliver task messages to the message middleware. The intelligent analysis engine is used to consume tasks from the message middleware, obtain corresponding text data through the integration gateway, perform analysis processes on the consumed tasks according to the task type, write the generated structured results into the analysis result table through the data storage module, and publish the completion event of the corresponding task to the message middleware. The message middleware is used to transmit the completion event to the backend master server; The backend master server is also used to forward events to the real-time notification module; The real-time notification module is used to push analysis completion notifications to the user front-end in real time, so that the user front-end can initiate a result query request to the back-end main server. The backend master server is also used to, when receiving the result query request, perform field trimming and anonymization on the analysis results corresponding to the result query request according to the requester role in the result query request, and return the differential results to the user frontend.

Images