An intelligent interview scheduling method and system based on bilateral knowledge graph alignment and RAG technology

Abstract

The application relates to an intelligent interview scheduling method and system based on bilateral knowledge graph alignment and RAG technology, which comprises the following steps: obtaining a text resume of a candidate and a post demand document, performing logical extraction and deep segmentation on the post demand document and the text resume respectively, and constructing a post side graph and a candidate side graph; performing deep comparison on the post side graph and the candidate side graph, generating a gap signal, generating different granularity problem nodes for each skill theme based on the gap signal and the skill points that need to be dug deeply, and giving a multilevel depth identifier to obtain the problem nodes; establishing a jump edge by using the jump relationship between the problem nodes, and giving an initial weight to the jump edge to cut into the interview; in the interview process, the answer quality is evaluated in multiple dimensions, the initial weight is updated according to the evaluation result, the highest node in the updated initial weight is selected as the next question.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and human resources technology, and in particular to an intelligent interview scheduling method and system based on bilateral knowledge graph alignment and RAG technology. Background Technology

[0002] The interview and assessment tasks in talent recruitment aim to achieve precise talent selection through a deep match between candidate resumes and job requirements (JD). Traditional human resource techniques mainly rely on keyword matching or manual screening. This process is not only time-consuming and labor-intensive, but also highly susceptible to the subjective factors of interviewers.

[0003] In recent years, with the development of Large Language Modeling (LLM) and Retrieval Augmentation (RAG) technologies, LLM-based resume analysis methods have made progress in automated initial screening by combining resume text with prompts. The traditional approach is to use RAG technology to break down the resume into text blocks and store them in a vector database. Then, the model evaluates the candidate unilaterally based on the job description content and outputs a matching score and evaluation summary.

[0004] These solutions typically rank candidates based on semantic similarity in their resumes and score their suitability. However, in real-world, complex recruitment environments, they still have the following shortcomings, making it difficult to simultaneously meet core requirements such as precision in hiring, industry applicability, and robustness: 1. Lack of feedback loop in execution, limiting assessment accuracy: Existing solutions mostly follow a one-way "retrieval-generation" process. The initial score generated by the system is based solely on static documents, lacking a real-time interactive feedback mechanism during the interview process. When a candidate's description of a key skill in their resume is vague, the system cannot automatically adjust its assessment strategy to verify the true level of mastery of that skill, resulting in insufficient robustness of the final matching report.

[0005] 2. Difficulty in structured semantic alignment and significant information loss: Plain text RAG solutions often treat resumes as flat strings, failing to understand the deep logical connections between skills, competencies, and projects. This leads to insufficient matching accuracy between resumes and job descriptions when faced with keywords that are not entirely identical but semantically highly related (such as synonyms or similar technologies in different project contexts), easily introducing irrelevant noise or missing key talent.

[0006] 3. Rigid questioning strategies, unable to generate targeted interview questions: Existing interview robots typically use static questions or simple logic trees. Due to the lack of real-time alignment and offsetting analysis between job requirements and candidate capabilities, the system cannot accurately identify specific skill gaps in candidates. For example, if a candidate lacks a necessary skill, traditional systems cannot automatically generate in-depth follow-up questions (gap discovery) targeting that gap, resulting in the interview process failing to effectively address the candidate's vulnerabilities.

[0007] 4. Fixed interview question bank, lack of dynamic awareness: Existing methods typically rely solely on simple sequential logic during interview scheduling, failing to incorporate multi-path fusion optimization based on the quality (Score) of candidates' past answers. This results in a lack of tiered depth in the interview process, making it impossible to dynamically adjust the weights of interview question paths based on candidates' answers, and the system's ability to detect the upper and lower limits of candidates is weak.

[0008] 5. Interview questions lack a chain of evidence and have insufficient visualization capabilities: The intermediate decision-making process of traditional solutions—why ask this question and why give this score—is a black box, lacking a visualized chain of evidence based on knowledge graphs. When it's necessary to explain the evaluation criteria to the interviewer, it cannot provide intuitive, structured support based on entity relationships.

[0009] In addition, existing technical solutions also have limitations, such as: 1. Existing technical solutions disclose inputting the resumes of candidates for interviews into a pre-trained interview question generation model to obtain at least one interview question. This approach is not flexible enough, the pre-training cost is high, the selection of interview questions is random, and there is a lack of evidence chain.

[0010] 2. Existing technical solutions add a separate AI-powered interview question generator module, which generates interview questions based on job postings. However, it cannot provide targeted, in-depth follow-up questions on specific skills (gap) missing from a candidate's resume. Furthermore, while this approach considers the quality of answers, it cannot dynamically adjust the depth and topic of subsequent questions based on the quality of the candidate's previous answer.

[0011] In summary, although existing technical solutions have made some progress in exploring AI interviews, there are still many bottlenecks in the perception of candidate answer quality, execution feedback, and dynamic scheduling and adaptation of interview questions. There is an urgent need to propose an intelligent interview decision framework that combines candidate resume and job description alignment with dynamic path optimization. Summary of the Invention

[0012] To address the problems existing in the prior art, the purpose of this invention is to overcome the shortcomings of existing automated interview systems, such as a single evaluation dimension, lack of execution feedback loop, and inability to conduct precise follow-up questions on competency gaps. This invention proposes an intelligent interview scheduling method and system based on bilateral knowledge graph alignment and RAG technology. By constructing a structured job and candidate knowledge graph and combining it with a scheduling algorithm that provides real-time feedback on answer quality, this invention enables in-depth and personalized verification of candidates' professional competence.

[0013] To achieve the above objectives, the present invention provides the following solution: An intelligent interview scheduling method based on bilateral knowledge graph alignment and RAG technology includes: Obtain the candidate's text resume and job requirement document, perform extraction logic and depth segmentation on the job requirement document and the text resume respectively, and construct the job-side map and the candidate-side map; A deep comparison is performed on the job-side map and the candidate-side map to generate gap signals. Based on the gap signals and the skill points that need to be explored, problem nodes of different granularities are generated for each skill theme and assigned multi-level depth labels to obtain problem nodes. Establish jump edges using the jump relationships between the question nodes and assign initial weights to the jump edges to initiate the interview; during the interview, evaluate the quality of the answers from multiple dimensions, update the initial weights based on the evaluation results, and select the node with the highest updated initial weight as the next question.

[0014] Optionally, constructing the job profile and the candidate profile includes: The structured data objects in the job requirement document after extracting the logic are persisted to a graph database. The logic is then extracted using a knowledge graph builder, and the extracted structured data objects are mapped to nodes and relationships in the graph database to perform skill mapping, thereby obtaining the relationship between job and professional skills and constructing the job-side graph. The chapter titles in the resume are identified by regular expressions at the beginning of the lines to define block boundaries. The resume is then deeply segmented using these block boundaries to obtain text blocks. Metadata is assigned to each text block, and the text blocks with assigned metadata are stored in a MySQL database. Entity mapping nodes are created using the structured archives in the MySQL database, and relationships between candidates and professional skills, candidates and employers, and candidates and project experience are established to construct the candidate profile.

[0015] Optionally, generating the gap signal includes: By maintaining a mapping table, the non-standard descriptions of candidates in the job side map and the candidate side map are unified and normalized with the standard terminology of the skill requirements in the job. If no direct match is found during the normalization process, a fuzzy keyword overlap strategy is adopted to extract Chinese and English keywords to calculate the maximum overlap and identify potential related skills. If no match is found, LLM semantics is used to extract implicit skills from the job requirement description and the candidate's complete project background and generate an alignment report. The skill edges associated with all project nodes of the candidate's skills in the alignment report are scanned. An index is established with skills as keys and related project lists as values ​​to determine experience depth. All skill nodes marked as key gaps in the job side map are scanned. If the matching status of the skill node is missing or cannot be confirmed after secondary inference by the LLM model, it is considered as a skill that is defined as essential in the job requirements but is missing in the resume, and the gap signal is generated.

[0016] Optionally, obtaining the problem node includes: The gap-driven interview mechanism takes gap signals and gap questions in the skill points that need to be explored as the primary purpose of the assessment. Based on the hierarchical strategy of multi-branch tree structure, it uses the progressive structure of multi-branch tree and the difficulty of questions to generate question nodes of different granularities for each skill topic and assigns them multi-level depth labels.

[0017] Optionally, assigning an initial weight to the jump edge includes: ; in, The base weights are calculated at the start of the interview. Indicates the initial weights. This indicates the weight corresponding to the core level of JD. Indicates gap risk The corresponding weights, The importance value of this skill in the job description map.

[0018] Optionally, a multi-dimensional evaluation of the quality of the response includes: The current question, candidate answers, and relevant job requirements are incorporated into prompts to assess accuracy, depth, and relevance, and candidate scores are obtained.

[0019] Optionally, updating the initial weights based on the evaluation results includes: ; in, For the final weight, These are the initial weights. Indicates job gain. Indicates the historical coverage balance factor. For feedback decision factors.

[0020] To achieve the above objectives, the present invention also provides an intelligent interview scheduling system based on bilateral knowledge graph alignment and RAG technology, comprising: The data preparation module is used to obtain candidates' text resumes and job requirement documents; The bilateral knowledge graph construction module is used to extract logic and perform depth segmentation on the job requirement document and the text resume respectively, and construct the job-side graph and the candidate-side graph. The talent matching and alignment module is used to perform in-depth comparison of the job-side map and the candidate-side map, generate gap signals, and generate problem nodes of different granularities for each skill theme based on the gap signals and the skill points that need to be explored, and assign multi-level depth labels to obtain problem nodes; The problem graph construction module is used to establish jump edges using the jump relationships between the problem nodes and to assign initial weights to the jump edges in order to get into the interview. The weight scheduling module is used to evaluate the quality of answers from multiple dimensions during the interview process, update the initial weights based on the evaluation results, and select the node with the highest updated initial weight as the next question.

[0021] The beneficial effects of this invention are as follows: Significantly improves the relevance of interviews: This invention uses knowledge graph alignment technology, which enables the system to automatically identify candidates' weaknesses and ask targeted supplementary questions, avoiding the rigid "one-size-fits-all" questioning mode of traditional systems.

[0022] With dynamic adaptive capability: Based on the weighting algorithm of answer scores, this invention enables the system to simulate the thinking of a senior interviewer and dynamically adjust the depth and path of questions according to the candidate's performance, which greatly improves the efficiency and robustness of talent verification.

[0023] High interpretability of evaluation results: This invention unfolds the interview process based on a structured graph path, with each evaluation and question supported by underlying graph entities, thus achieving white-box and transparency in the evaluation process.

[0024] Low resource consumption and high versatility: The modular design of the system in this invention supports the dynamic injection of JD requirements for different positions through Prompt, without the need for model fine-tuning for specific positions, and has strong portability and implementability. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of data parsing and candidate evaluation according to an embodiment of the present invention; Figure 2 This is a diagram illustrating the alignment strategy for constructing knowledge graphs and entities in an embodiment of the present invention. Figure 3 This is a diagram illustrating the construction of the problem knowledge graph in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the planning strategy for dynamic paths in a problem knowledge graph according to an embodiment of the present invention. Figure 5 This is a flowchart of an intelligent interview scheduling method based on bilateral knowledge graph alignment and RAG technology according to an embodiment of the present invention; Figure 6 Preparing preliminary data for embodiments of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] like Figure 5As shown, this embodiment discloses an intelligent interview scheduling method based on bilateral knowledge graph alignment and RAG technology, including: acquiring candidates' text resumes and job requirement documents; performing extraction logic and depth segmentation on the job requirement documents and text resumes respectively, and constructing job-side graphs and candidate-side graphs; performing depth comparison on the job-side graphs and candidate-side graphs to generate gap signals; generating question nodes of different granularities for each skill topic based on the gap signals and the skill points that need to be explored, and assigning multi-level depth labels to obtain question nodes; establishing jump edges using the jump relationships between question nodes, and assigning initial weights to the jump edges to enter the interview; during the interview process, evaluating the quality of the answers in multiple dimensions, updating the initial weights based on the evaluation results, and selecting the node with the highest updated initial weight as the next question.

[0030] Specifically, this embodiment discloses an intelligent interview scheduling method based on bilateral knowledge graph alignment and RAG technology, including: like Figure 6 As shown, the preliminary data preparation involves: first, obtaining the candidate's original resume document and the target job requirements document; second, using ResumeNodeParser to perform context-aware deep segmentation of the resume, and then using an embedding model to vectorize the text blocks and store them in the ChromaDB vector database; simultaneously, using JDProcessor to perform structured parsing of the job requirements document; and finally, using LLM to extract structured candidate profiles from the complete resume text and store them in a MySQL database. Only after the above preparations are completed can the core process of this invention proceed.

[0031] (1) Feature extraction from resumes and construction of a two-sided knowledge graph: The first step is to construct a job graph. The core of this construction lies in transforming unstructured job descriptions into structured graph data objects and persisting them to the Neo4j graph database, such as... Figure 1 As shown. The extraction logic via the Job Knowledge Graph Builder can be divided into preprocessing segments, LLM structured extraction rules, and a fallback mechanism for failures. Specifically: 1. Preprocessing and segmentation: Read the .md format JD file, extract the job title using regular expressions, and extract content by segmenting according to chapter titles (company background, job responsibilities, job requirements), etc.

[0032] 2. LLM structured extraction: Construct the prompt word "Prompt", inject the complete JD text and recruitment source into the template, require LLM to output in strict JSON format, and then call LLM.

[0033] 3. Fallback mechanism: When LLM extraction fails (network timeout, invalid output format, parsing error, etc.), the rule will be rolled back.

[0034] After the logic extraction is complete, the process of building the knowledge graph is actually mapping the extracted structured data objects to nodes and relationships in the graph database. Then, the JobGraphStore component will execute the following persistence logic: Skill mapping: Establish (Job)-[:REQUIRES_SKILL]→(Skill) relationship.

[0035] Furthermore, the candidate graph is constructed based on pre-stored structured archives in MySQL, mapping them to a complex relationship network, while also incorporating fallback rules. First, entity mapping nodes are created: Candidate, Skill, Project, and Company. Then, the relationship logic is constructed. Skill association: Establish a (Candidate)-[:HAS_SKILL]→(Skill) relationship; Work experience: Establish a relationship between (Candidate) and (Company); Project experience: Establishing the (Candidate)-[:DELIVERED]→(Project) relationship; The rollback rules include: when the LLM extraction chain throws any exception, including network timeout, LLM returning invalid JSON, Pydantic parsing and validation failure, the system will not be interrupted, but will enter a rollback, using pure rules (without LLM) to generate structured data from the pre-processed JD segment text. That is, firstly, text cleaning and line splitting are performed: blank lines are removed, and plain text content is retained. Each line represents an independent requirement or responsibility, and heuristic classification is based on position: in the JD text, the job requirements are arranged from high to low importance by default, so the first item is marked as "must have", and the later item is marked as "bonus", etc. Here, JD (Job Description) represents the job requirements document, RAG (Retrieval-Augmented Generation) represents retrieval-enhanced generation, and LLM (Large Language Model) represents the large language model.

[0036] (2) Talent matching and alignment strategy based on graph association: This strategy is the core decision-making layer of the entire system, implemented through the JobCandidateAligner class. Its purpose is to conduct in-depth comparison between the job-side knowledge graph and the candidate-side knowledge graph, mainly including: a. Deep Semantic Alignment Mechanism: The system employs a multi-level matching strategy, surpassing existing methods that rely solely on literal keyword matching. The system maintains a mapping table to normalize non-standard descriptions provided by candidates against the standard terminology required for the position. If no direct match is found, a fuzzy keyword overlap strategy is used to extract Chinese and English keywords, calculate the maximum overlap, and identify potential related skills. If no match is found again, an LLM semantic assessment is used as a fallback. This provides the job requirements description and the candidate's complete project background to the LLM, allowing it to determine whether there are implicit skills supporting the alignment, providing more specific details. 1. Input: The list of (Job)-[:REQUIRES_SKILL]->(Skill) nodes in the job side map and the list of (Candidate)-[:HAS_SKILL]->(Skill) nodes in the candidate side map.

[0037] 2. Alignment operation: Traverse each skill node in the job side graph, find the corresponding entity in the candidate side graph, and perform entity alignment through a three-layer strategy (mapping table normalization → fuzzy keyword overlap → LLM semantic inference).

[0038] 3. Output: A structured alignment report, including the matching status and coverage metrics for each skill. This alignment report is reflected in the system, primarily showing skill coverage and the core skill missing rate.

[0039] b. Practical Evidence Chain Verification Mechanism: Experience depth is determined by the connecting edges in the graph, identifying any inflated or superficial aspects of a candidate's skills. All project nodes and their associated skill edges in the candidate's Neo4j graph database are scanned, creating an index with the skill as the key and the list of related projects as the value. If the candidate possesses the skill and has applied it in at least one real project, the matching level is set to Strong. Otherwise, it is defined as Basic.

[0040] c. Risk and Gap Identification: To prevent subsequent interviews from deviating from the JD's theme, the system will focus more on gap issues. Specifically, it will list skills defined as essential in the JD but missing from the resume as "missing_must_have" critical gaps. These will be passed to the QuestionGenerator module as the basis for generating interview questions. After completing the semantic alignment (a) and experience depth determination (b) above, the system scans all skill nodes marked with importance as must_have in the job side map. If the matching status of the skill in the alignment result is Missing or cannot be confirmed after LLM secondary inference, the skill is marked as missing_must_have gap signal, which is used as the priority input for generating subsequent interview questions.

[0041] (3) Dynamic generation of guided interview question graph: The generation of guided question graph is the key and innovative point of the entire system design, because it is necessary to set reasonable difficulty to facilitate a comprehensive assessment of the candidate's abilities, while also preventing the AI ​​interviewer from going off-topic. The core task of this module is to use the reasoning ability of LLM to generate highly targeted questions based on the specific circumstances of the candidates.

[0042] First, this module receives the gap signals and skill points requiring further exploration from the previous stage. It employs a gap-driven interview mechanism, prioritizing gap-related questions as the primary objective of assessment. The skill points requiring further exploration are those marked as "requiring further exploration" during the alignment phase when the matching status of a skill is Partial (partial semantic matching) or Basic (only mentioned literally but without project support). These points are not missing, but rather lack projects to verify the candidate's claims; the candidate's depth of understanding is questionable, requiring verification through follow-up interview questions.

[0043] Secondly, a hierarchical strategy based on a multi-branch tree structure was designed. The progressive structure of the multi-branch tree closely matches the difficulty setting of the problem, and the generated problems are also set to three levels of depth: Depth 1 (Basic Verification): Verify whether the candidate possesses the basic knowledge claimed in their resume; Depth2 (Advanced Verification): Requires candidates to describe specific workflows or technical details; Depth3 (In-depth exploration): Examines the depth of problem-solving and thinking patterns in complex scenarios; Here, Depth 1 / 2 / 3 represent the problem depth levels, corresponding to basic confirmation / advanced verification / in-depth exploration, respectively.

[0044] Secondly, the received gap signals and the list of skills requiring in-depth exploration are combined with the candidate's resume background and the job description requirements to construct a structured Prompt for injection into the LLM. This requires generating three levels of questions (Depth 1 / 2 / 3) for each skill topic, using the candidate's relevant project experience in the Prompt as context to ensure the generated questions are targeted rather than from a generic question bank, thus obtaining scattered question nodes. After obtaining these scattered question nodes, QuestionGraphBuilder is responsible for the logical connections between nodes and establishing jump relationships between questions. This includes vertical in-depth exploration within the same topic, establishing a progressive examination path from shallow to deep (Depth 1 → Depth 2 → Depth 3). It also includes horizontal connections between topics, establishing jump edges between nodes in different topics based on skill relevance (e.g., the relationship between Java and Spring Boot). Each edge is assigned an initial weight, which is determined by the first question before the answer. In other words, the AI ​​decides which side (question) to start the interview from. The system comprehensively evaluates the priority of each candidate starting question using an initial weight formula. This is derived from the importance value of the skill in the job description graph (by default, the must_have skill receives a higher value). The gap risk value is derived from the gap status of the skill in the alignment report (skills marked as missing receive the highest gap risk value). The problem graph logic is designed as a multi-branch tree structure. The system selects the problem with the highest W_base among all entry nodes (in-degree is 0, indicating that it is not a predecessor of any problem, i.e., a basic problem) as the first interview question.

[0045] ; Its calculation logic comprehensively considers the importance of JD and the risk of gaps.

[0046] in, Indicates the initial weights. This indicates the weight corresponding to the core level of JD. Indicates gap risk The corresponding weights.

[0047] (4) Dynamic weight scheduling and real-time evaluation based on feedback mechanism: This step utilizes the semantic understanding capability of LLM to solve the problem of "how to quantify candidate performance". It does not rely on simple keyword matching. After the candidate submits an answer on the front-end page, the GuidedInterviewSession module is called to inject the current question, the candidate's answer and the relevant requirements of the JD into the prompt word Prompt, requiring a score based on accuracy, depth and relevance. Then, a floating-point number between 0.0 and 1.0 is output as the candidate's score, with specific evaluation reasons and suggestions for follow-up questions.

[0048] After receiving the score, before selecting the next question, the candidate updates the weighting logic and introduces... Candidate response feedback mechanism: ; in, These are the base weights calculated at the start of the interview. The feedback from the candidate's answers will be taken into account in subsequent interview questions.

[0049] The weight of cross-topic edges is selected from all outgoing edges of the current node using the `_select_best_next_question` method. The highest-scoring question is used as the next question in the round, thereby achieving adaptive path planning.

[0050] In one embodiment, this embodiment provides an intelligent interview scheduling method based on bilateral knowledge graph alignment and RAG technology. The method includes three key steps: JD semantic parsing, candidate resume segmentation and extraction, and vectorized storage and evaluation.

[0051] Resumes and job descriptions are often unstructured documents, especially since the descriptions of skills or technologies in resumes may not be as standardized and accurate as required by the job requirements. Existing keyword-based matching methods have limitations; for example, when faced with candidates' non-standard descriptions, keyword-based methods may miss candidates whose abilities match but whose descriptions are not standardized. To address this issue, this invention employs a large-scale model semantic parsing approach to identify and standardize non-standard descriptions in resumes. A mapping table is maintained, recording the mapping between non-standard descriptions in resumes and standardized terms in job descriptions. The system first searches the mapping table; if no match is found, large-scale model semantic parsing is used to determine the candidate's potential skills. On the job side, the "Job Responsibilities" and "Job Requirements" modules are parsed, and a unique key is generated using the MD5 algorithm. The first eight characters are used as the unique identifier for the subsequent job node in Neo4j. The generated key is passed to JobGraphStore to create a root node of type Job in the Neo4j database. This node serves as the center, linking subsequently extracted elements such as Skills, Competencies, and Project Scenarios. The Competency and Project Scenario relationships are: Competency mapping: establishing the (Job)-[:REQUIRES_COMPETENCY]→(Competency) relationship; Project Scenario mapping: establishing the (Job)-[:HAS_SCENARIO]→(ProjectScenario) relationship.

[0052] Furthermore, the system uses the `ResumeNodeParser` class to perform deep segmentation of unstructured resumes, ensuring the accuracy of subsequent retrieval. This class is not a simple length cutter, but a context-aware parser. It first defines block boundaries by identifying section headings such as core skills, basic information, work experience, project experience, and education background in the resume using regular expressions at the beginning of each line. During the segmentation process, the system automatically assigns metadata to each text block, such as using `chunk_type` to mark whether the block belongs to skills, experience, or basic information. Afterwards, the extracted structured information from the resumes is saved to a local MySQL database and simultaneously stored in a ChromaDB vector database along with the extracted job description information.

[0053] Finally, the system utilizes the `extract_profile` method to transform from plain text to a structured model. The system defines strict prompt word templates using a structured extraction chain. RAG is used to compare the similarity of the two-sided vectors, and the defined prompt words guide the large language model for parsing. The LLM output is required to contain a JSON object including the candidate's name, total years of work experience, a summary of strengths / weaknesses, project details, and an overall match score. The system does not blindly trust the LLM's scores. The `validate_recommendation` method automatically verifies different scores based on the model's `overall_match_score` and adjusts the recommendation level accordingly.

[0054] 2. Example of a job-candidate alignment algorithm: like Figure 2 As shown, the job-candidate alignment algorithm is the core logic layer connecting the job requirements side (Job Graph) and the candidate competency side (Candidate Graph). Through the JobCandidateAligner class and its related utility functions, the system implements deep alignment from literal matching to semantic reasoning. The following is a detailed technical introduction to this section: The alignment concept mentioned in this invention is an automated logical mapping process that maps and calculates the matching degree between the skills, abilities, project scenarios, and risk points defined on the job side and the skill items, project experience, and work background recorded on the candidate side.

[0055] Alignment encompasses skills verification, competency verification, and project scenario alignment. First, skills verification distinguishes between essential and bonus skills, with matching status categorized as Strong (supported by projects), Basic (only mentioned literally), Partial (partial semantic matching or relevance), and Missing (completely missing). Competency verification assesses whether a candidate's overall qualities support the core capabilities required for the position, calculating competency support rate by measuring the coverage of relevant skills. Project scenario alignment verifies whether the candidate has played a relevant role and solved problems in typical business scenarios similar to those described in the job description. The alignment process ultimately produces a structured report containing quantitative indicators, gap signals, and graph aggregation relationships. The alignment result serves as the starting point for subsequent AI interviews, primarily in driving question generation and interview path planning. It ensures the system generates targeted questions based on the gaps identified through alignment and uses alignment signals to assign weights to different paths in the question graph, ensuring the interview prioritizes covering risk points with lower alignment levels.

[0056] After aligning the job position with the candidate, gap verification is performed based on LLM semantics. For skills initially identified as Missing or Partial and marked as "must-have" in importance, the system introduces LLM for secondary deep inference. For example, if the job description requires high-concurrency processing, and the candidate did not use the term but described handling systems with millions of QPS, the matching level is adjusted. If skills are still found to be missing after secondary inference, the system assigns a status: "open" indicates an uncovered risk point, which will be a key focus for subsequent AI interviews; "Addressed" indicates that the candidate already possesses the relevant skills and the risk point has been covered. Finally, key indicators such as overall coverage and core matching degree are calculated and output. This approach addresses the previous technology's problem of tracing the candidate's evaluation criteria and provides a theoretical basis for targeted and focused AI interviews that address the candidate's weaknesses and avoid divergent thinking.

[0057] 3. Example of constructing a multi-branch tree-like problem graph: like Figure 3 As shown, the construction of a multi-branch tree-like problem graph is the core step in transforming static alignment results into dynamic interview logic. This is mainly achieved by the system building a tree structure with semantic depth and logical coherence in Neo4j. The following is a detailed introduction to the technical implementation of this part: First, the system utilizes a Large Language Model (LLM) to generate question nodes of varying granularity for each skill topic, assigning them depth labels from 1 to 3, with increasing difficulty to progressively assess the candidate's true abilities. Depth 1 is the skill verification level, aiming to verify the candidate's basic knowledge and understanding of related skills. Depth 2 is the application level, aiming to assess the candidate's ability to implement skills in relevant projects. By combining the candidate's project background, it requires describing the implementation path of specific technologies to verify the candidate's depth of mastery of the skill. Depth 3 is the core logic level, primarily probing the candidate's decision-making, optimization, and systematic thinking and problem-solving abilities in complex real-world application scenarios. The construction and selection of questions fully consider the gap signals mentioned in the above steps, and the questions are progressively more difficult, avoiding the randomness and blindness of existing technology-based questions, thus enabling a more comprehensive assessment of the candidate's abilities.

[0058] Secondly, the construction of logically connected edges is crucial. After the question nodes are generated, directed edges need to be established between them to explore potential interview trends. This system defines four question themes: skills verification, experience mining, gap detection, and competency assessment, with each theme represented by a different color.

[0059] Under the same topic, the system establishes connecting edges from Depthn to Depthn+1. This tiered, question-driven connection method ensures that interviewers can conduct progressively deeper follow-up questions on a particular issue, guaranteeing the relevance and logical rigor between questions. If two question nodes involve the same related skills, the system establishes cross-topic jump edges, automatically switching topics as needed. If question A and question B both originate from the candidate's same core project, the system establishes a connection, allowing the interview logic to switch technical dimensions within the same project context, preventing excessive follow-up questions on the same topic.

[0060] The system is designed based on the principle of utilizing gap signals. However, when establishing edge relationships, if the target problem node involves the essential skills or gaps in the JD, the system will identify them by backtracking the core requirements of the JD and assign them a higher initial weight to ensure that the core requirements in the JD have a higher weight.

[0061] Furthermore, after constructing the question graph, similar to the candidate knowledge graph and the job knowledge graph, it is stored in the Neo4j database. Each question node stores relevant information such as question type, depth, and metadata. The edges store the initial weight and real-time weight (used for real-time updates by the subsequent scheduling algorithm) pointing to that question. The thicker the edge, the greater the weight of the path, and the more likely it is to be the next interview question.

[0062] The significance of this approach lies in the fact that it's not merely a collection of questions, but rather the establishment of a set of probing paths. Through hierarchical layering, the system can quickly verify whether a candidate is making false statements via D1, and then uncover their actual technical skills through D2 / D3. Furthermore, through logically connected edges, the interview process can seamlessly transition between different skill points, avoiding the negative impact on the interactive experience caused by overly abrupt jumps.

[0063] 4. Dynamic weight scheduling Example: like Figure 4 As shown, the dynamic weight scheduling engine is the core of the guided interview system. It is responsible for adjusting the logical path of questions and answers in real time during the interview process, ensuring that the interview can both deeply probe the candidate's upper limit of ability and accurately cover the risk points required for the position. It can be said that the strategy of using knowledge graphs and dynamic weight scheduling compensates for the shortcomings of existing technologies in question-answering systems. The following is a detailed introduction to the dynamic weight scheduling based on system logic: After a candidate submits their answer to the current question, the large model first evaluates the answer's quality from multiple dimensions, outputting a floating-point score S representing the answer's professionalism, completeness, and practicality. Then, based on the answer's performance, suggestions for further exploration are given (e.g., "deeper" suggests in-depth exploration, "gap" suggests verifying gaps, and "skills" suggests switching skill points). Simultaneously, the scheduler iterates through all outgoing edges of the current question and applies a formula to calculate the final weight of each path. .

[0064] ; in, These are the initial weights. This indicates job gain, derived from extracting missing skills and open risk points from earlier alignment reports. If the target node of the next question involves these missing items or risk points, the system will... The increased value ensures that the system plays a goal-oriented role in interviews, prioritizing the verification of candidates' weaknesses. This represents the historical coverage balancing factor. If a topic has been asked multiple times, this factor will be reduced to prevent the system from getting too bogged down in a single skill point and to force the interview path to evolve in a more diversified direction, so as to showcase the candidate's abilities more comprehensively. To provide feedback on decision-making factors—the core of the AI ​​interviewer's decision-making interaction—if the interviewee's answer is excellent, the system will significantly increase the weight of edges leading to higher-depth nodes on the same topic to probe the candidate's depth of mastery of that skill. If the interviewee's answer is poor, the system will prioritize activating nodes involving similar skills but at lower depths, confirming whether the candidate truly possesses the ability through horizontal comparison or basic verification, thus preventing the interview from reaching a stalemate. Finally, all outgoing edges are processed according to... Finally, the weights are sorted and selected. The highest-ranking node serves as a candidate for the next question. Through this dynamic weighted scheduling engine, the system transitions from a fixed question bank to an adaptive dialogue. The interview path is no longer pre-set, but rather the optimal probing path calculated in real-time based on each candidate's response. The result is that the system consistently tends to select questions that "provide the most unconfirmed information," thereby constructing the most complete candidate competency profile within the limited interview time.

[0065] This embodiment also provides an intelligent interview scheduling system based on bilateral knowledge graph alignment and RAG technology, including: a data preparation module for acquiring candidates' text resumes and job requirement documents; a bilateral knowledge graph construction module for extracting logic and performing depth segmentation on the job requirement documents and text resumes respectively, and constructing job-side and candidate-side graphs; a talent matching and alignment module for performing depth comparison on the job-side and candidate-side graphs, generating gap signals, and generating question nodes of different granularities for each skill topic based on the gap signals, and assigning multi-level depth labels to obtain question nodes; a question graph construction module for establishing jump edges using the jump relationships between question nodes, and assigning initial weights to the jump edges to enter the interview; and a weight scheduling module for performing multi-dimensional evaluation of the answer quality during the interview process, updating the initial weights based on the evaluation results, and selecting the node with the highest updated initial weight as the next question.

[0066] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An intelligent interview scheduling method based on bilateral knowledge graph alignment and RAG technology, characterized in that, include: Obtain the candidate's text resume and job requirement document, perform extraction logic and depth segmentation on the job requirement document and the text resume respectively, and construct the job-side map and the candidate-side map; A deep comparison is performed on the job-side map and the candidate-side map to generate gap signals. Based on the gap signals and the skill points that need to be explored, problem nodes of different granularities are generated for each skill theme and assigned multi-level depth labels to obtain problem nodes. Establish jump edges using the jump relationships between the problem nodes, and assign initial weights to the jump edges to get into the interview; During the interview process, the quality of the answers is evaluated from multiple dimensions. The initial weights are updated based on the evaluation results, and the node with the highest updated initial weight is selected as the next question.

2. The intelligent interview scheduling method based on bilateral knowledge graph alignment and RAG technology according to claim 1, characterized in that, Constructing the job profile and the candidate profile includes: The structured data objects in the job requirement document after extracting the logic are persisted to a graph database. The logic is then extracted using a knowledge graph builder, and the extracted structured data objects are mapped to nodes and relationships in the graph database to perform skill mapping, thereby obtaining the relationship between job and professional skills and constructing the job-side graph. The chapter titles in the resume are identified by regular expressions at the beginning of the lines to define block boundaries. The resume is then deeply segmented using these block boundaries to obtain text blocks. Metadata is assigned to each text block, and the text blocks with assigned metadata are stored in a MySQL database. Entity mapping nodes are created using the structured archives in the MySQL database, and relationships between candidates and professional skills, candidates and employers, and candidates and project experience are established to construct the candidate profile.

3. The intelligent interview scheduling method based on bilateral knowledge graph alignment and RAG technology according to claim 1, characterized in that, Generating the gap signal includes: By maintaining a mapping table, the non-standard descriptions of candidates in the job side map and the candidate side map are unified and normalized with the standard terminology of the skill requirements in the job. If no direct match is found during the normalization process, a fuzzy keyword overlap strategy is adopted to extract Chinese and English keywords to calculate the maximum overlap and identify potential related skills. If no match is found, LLM semantics is used to extract implicit skills from the job requirement description and the candidate's complete project background and generate an alignment report. The skill edges associated with all project nodes of the candidate's skills in the alignment report are scanned. An index is established with skills as keys and related project lists as values ​​to determine experience depth. All skill nodes marked as key gaps in the job side map are scanned. If the matching status of the skill node is missing or cannot be confirmed after secondary inference by the LLM model, it is considered as a skill that is defined as essential in the job requirements but is missing in the resume, and the gap signal is generated.

4. The intelligent interview scheduling method based on bilateral knowledge graph alignment and RAG technology according to claim 1, characterized in that, The problem nodes obtained include: The gap-driven interview mechanism takes gap signals and gap questions in the skill points that need to be explored as the primary purpose of the assessment. Based on the hierarchical strategy of multi-branch tree structure, it uses the progressive structure of multi-branch tree and the difficulty of questions to generate question nodes of different granularities for each skill topic and assigns them multi-level depth labels.

5. The intelligent interview scheduling method based on bilateral knowledge graph alignment and RAG technology according to claim 1, characterized in that, Assigning initial weights to the jump edges includes: ； in, The base weights are calculated at the start of the interview. Indicates the initial weights. This indicates the weight corresponding to the core level of JD. Indicates gap risk The corresponding weights, The importance value of this skill in the job description map.

6. The intelligent interview scheduling method based on bilateral knowledge graph alignment and RAG technology according to claim 1, characterized in that, The multi-dimensional evaluation of the quality of the responses includes: The current question, candidate answers, and relevant job requirements are incorporated into prompts to assess accuracy, depth, and relevance, and candidate scores are obtained.

7. The intelligent interview scheduling method based on bilateral knowledge graph alignment and RAG technology according to claim 1, characterized in that, Updating the initial weights based on the evaluation results includes: ； in, For the final weight, These are the initial weights. Indicates job gain. Indicates the historical coverage balance factor. For feedback decision factors.

8. An intelligent interview scheduling system based on bilateral knowledge graph alignment and RAG technology, implemented according to any one of claims 1-7, characterized in that, include: The data preparation module is used to obtain candidates' text resumes and job requirement documents; The bilateral knowledge graph construction module is used to extract logic and perform depth segmentation on the job requirement document and the text resume respectively, and construct the job-side graph and the candidate-side graph. The talent matching and alignment module is used to perform in-depth comparison of the job-side map and the candidate-side map, generate gap signals, generate problem nodes of different granularities for each skill topic based on the gap signals, and assign multi-level depth labels to obtain problem nodes; The problem graph construction module is used to establish jump edges using the jump relationships between the problem nodes and to assign initial weights to the jump edges in order to get into the interview. The weight scheduling module is used to evaluate the quality of answers from multiple dimensions during the interview process, update the initial weights based on the evaluation results, and select the node with the highest updated initial weight as the next question.

Images