Four-dimensional-based resume and post matching analysis method and system

Abstract

The invention discloses a four-dimensional-based resume and post matching analysis method and system, and relates to the technical field of data information processing. The method comprises the steps of obtaining an original JD text and performing text preprocessing; four-dimensional information is set, and information extraction is carried out based on four dimensions; establishing a mapping relation between the JD extraction field and the resume four-dimensional degree; performing resume matching based on four dimensions; carrying out dynamic weight distribution; performing comprehensive scoring based on the dynamic weight distribution information; and result verification and output are carried out. The system comprises a preprocessing module, a four-dimensional information extraction module, a matching calculation module, an interpretability module and a verification and output module. According to the method, a four-dimensional extraction framework directly corresponding to a resume system is proposed for the first time, a four-dimensional mapping model is formed, semantic alignment of recruitment requirements and talent resumes is realized, and the analysis accuracy of complex semantics is improved in combination with a rule engine and a machine learning model.

Description

Technical Field

[0001] The present invention relates to the field of data information processing technology, and in particular to a resume and job matching analysis method and system based on four dimensions. Background Art

[0002] The current recruitment market has high requirements for talent and job matching. However, due to some shortcomings in the current recruitment market, it is unable to meet these high requirements. For example, the current recruitment market has at least the following shortcomings in matching resumes: (1) Information overload: Human Resources (HR) processes over 200 resumes per day. Manual screening is inefficient, with traditional screening efficiency being less than 10 resumes per hour.

[0003] (2) Matching bias: Traditional keyword matching ignores semantic associations, such as considering “Python development” and “Python programming” as different skills.

[0004] (3) Single dimension: Only focusing on technology stack matching, ignoring comprehensive factors such as educational level and years of work experience.

[0005] (4) Insufficient parsing accuracy: The parsing accuracy of unstructured resumes (such as PDFs and images) is <85% (Beike data).

[0006] (5) Rigid scoring model: Fixed weights cannot adapt to job differences. For example, technical positions place more emphasis on work experience, while management positions focus on academic qualifications.

[0007] (6) Lack of explainability: Matching results cannot be traced back, and HR has difficulty verifying the algorithm logic.

[0008] For example, the Chinese invention patent with publication number CN111428488A discloses a resume data information parsing and matching method, device, electronic device and medium, whose technical solution is: pre-processing the retrieved resume to obtain the resume to be parsed, and constructing a word segmentation directed acyclic graph based on a pre-constructed word segmentation dictionary to segment the resume to be parsed, so as to quickly obtain the word segmentation result of the resume to be parsed and obtain the resume text, and then construct a co-occurrence matrix based on the resume text, and determine the keywords of the resume text based on the co-occurrence matrix, and obtain the word sequence in the keyword, use the word representation model to perform word representation processing on the word sequence to obtain the word representation of the word sequence, and input the word representation into the resume label parsing model to obtain a resume label sequence, and further calculate the similarity between each label in the resume label sequence and the label of each position to determine the resume matching each position, thereby realizing fast and accurate intelligent matching of positions and resumes. For example, the Chinese invention patent with publication number CN119918859A discloses a precise talent matching and recommendation system based on big data. Its technical solution is as follows: it includes: a job requirement modeling module, which is used to convert job requirement data into multidimensional feature vectors; a candidate feature modeling module, which is used to convert candidate data into multidimensional feature vectors; a matching cost construction module, which constructs a matching cost matrix; an optimal transmission matching module, which solves the optimal transmission plan through an optimization algorithm and outputs the matching relationship between the job and the candidate; and a matching result output module, which outputs the matching results and matching priority between the candidate and the job based on the optimal transmission plan. A modular solution for job and candidate modeling is adopted to convert requirements and features into a multidimensional matrix, and a weight distribution is generated in combination with the score to accurately characterize the characteristics of the job and candidate. However, the accuracy of the above resume and job matching is still insufficient.

[0009] Therefore, it is necessary to design a resume and job matching analysis method and system based on four dimensions to solve the existing problems. Summary of the Invention

[0010] In response to the problems existing in the prior art, the present invention provides a resume and job matching analysis method and system based on four dimensions.

[0011] To achieve the above object, the technical solution adopted by the present invention is as follows: A resume and job matching analysis method based on four dimensions includes the following steps: S1. Obtain the original JD text and perform text preprocessing; S2. Establish four-dimensional information and extract information based on the four-dimensional information; the four-dimensional information includes basic information, educational information, work experience, and technology stack; S3. Establish a mapping relationship between JD extraction fields and the four dimensions of resumes; S4. Match resumes from four dimensions; S5. Dynamically assign weights based on the four-dimensional information matching results; The allocation process is: S51. Identify job types; The specific identification process is: Build a text classification model: Use a combination of the BERT model and the CRF model to classify JD job types; Set keyword weight: automatically adjust module weight according to job type; S52. Dynamically adjust the weight according to the recognition result; Specific adjustment measures adopted include: 1) Adjustment based on job characteristics: The system will automatically adapt different weight templates based on the results of job type identification; 2) Data-driven optimization: Utilizing labeled data from multiple resumes and job JDs, and using machine learning models to train weight allocation strategies; 3) Based on user customization: To meet the company's special recruitment needs, the system supports HR to customize weight templates; S6. Perform comprehensive scoring based on dynamic weight distribution information; S7. Verify and output the results.

[0012] Based on the above technical solution, further, in S1, text preprocessing includes format cleaning, sentence processing and stop word filtering.

[0013] Based on the above technical solution, further, in S3, when the information in the four dimensions of the resume is basic information, the corresponding JD extraction fields are the full name of the position, the company to which it belongs, and the work location, and the corresponding mapping rule is direct mapping; when the information in the four dimensions of the resume is educational information, the corresponding JD extraction fields are the minimum educational requirements and professional restrictions, and the corresponding mapping rule is educational field type conversion; when the information in the four dimensions of the resume is work experience, the corresponding JD extraction fields are the experience years requirement, target industry and core responsibilities, and the corresponding mapping rule is job description structuring; when the information in the four dimensions of the resume is the technology stack, the corresponding JD extraction fields are the programming language, framework tools and certificate requirements, and the corresponding mapping rule is reorganization according to the three-level structure of "technology category-specific requirements-proficiency".

[0014] Based on the above technical solution, further, in S4, the matching process of basic information is: S41, first perform key information verification; S42, after completing the key information verification, evaluate the integrity and format standardization of the basic information; S43, comprehensively consider the key information verification, field coverage and format standardization scores to generate the final score of the basic information module.

[0015] Based on the above technical solution, further, in S4, the matching process of academic qualification information is as follows: Step 1, mapping processing of the subject classification library of the Ministry of Education; Step 2, establishing an academic qualification bonus mechanism; Step 3, performing academic qualification information scoring verification operation.

[0016] Based on the above technical solution, further, in S4, the matching process of work experience is: step a, performing semantic similarity calculation; step b, performing time series verification; step c, performing work experience scoring verification operation.

[0017] Based on the above technical solution, further, in S4, the technology stack matching process is: step A, skill standardization and expansion mechanism processing; step B, calculation of matching degree; step C, technology stack scoring verification operation.

[0018] Based on the above technical solution, further, in step B, the matching logic for calculating the matching degree includes: exact matching: when the resume skills are completely consistent with the JD skills, it is counted as 100%; semantic matching: the skill vector similarity is calculated through the BERT model, and the threshold is set to 0.7; industry knowledge enhancement: combined with the skill association relationship in the knowledge graph, the matching range is expanded.

[0019] Based on the above technical solution, further, in S6, the process is: S61, using a linear weighted formula to perform formula verification and threshold control processing; wherein, the S61 process is: 1) calculating a comprehensive score; 2) setting a screening threshold; S62, performing an explainable design.

[0020] A resume and job matching analysis system based on four dimensions executes a resume and job matching analysis method based on four dimensions. The resume and job matching analysis system includes a preprocessing module, a four-dimensional information extraction module, a matching calculation module, an interpretability module, and a verification and output module; the preprocessing module transmits processed data to the four-dimensional information extraction module; the four-dimensional information extraction module transmits data to the matching calculation module; the matching calculation module outputs data to the interpretability module and the verification and output module; and the interpretability module transmits data to the verification and output module.

[0021] Compared with the prior art, the present invention has the following beneficial effects: (1) This paper proposes for the first time a four-dimensional extraction framework that directly corresponds to the resume system, forming a four-dimensional mapping model to achieve semantic alignment between recruitment needs and talent resumes, and combines the rule engine with machine learning models (such as LSTM sequence classifier) to improve the accuracy of complex semantic parsing.

[0022] (2) The present invention realizes four-dimensional independent scoring for the first time, that is, it adopts the idea of dimensional matching, and the accuracy of key field parsing can reach 98.6% (an increase of 13.6% compared with Beisen data); and the processing time of a single resume is less than 2 seconds, supporting real-time retrieval of a resume database of millions of resumes; supporting customized weight configuration to adapt to the needs of different industries (such as financial positions pay more attention to certificates, and Internet positions focus on project experience); at the same time, compared with the traditional method of fixed weights (such as WPS Daoke resume scoring), the present invention can automatically adjust the weights based on the job type; integrating OCR, NLP, and knowledge graph technologies, supporting unstructured resume parsing (such as CV+ system), and also supporting online updating of industry terminology libraries, and automatically identifying emerging technology vocabulary through the FastText model.

[0023] (3) The present invention can be applied to: 1. The recruitment platform's intelligent job analysis system provides resume optimization suggestions; 2. Resume screening tools for corporate HR departments can process over 100,000 resumes per day, with an initial screening accuracy rate of 92% (compared to 65% for traditional methods); 3. A job demand analysis platform for headhunting companies, cross-platform resume database retrieval, and matching efficiency increased by 8 times. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flow chart of the method of the present invention; Figure 2 Flowchart showing the correlation between steps S3 and S4 in the method of the present invention; Figure 3 Schematic diagram of the system of the present invention. DETAILED DESCRIPTION

[0025] The present invention will be further described and illustrated below with reference to the accompanying drawings and specific embodiments. The technical features of each embodiment of the present invention may be combined accordingly, provided that there is no conflict between them.

[0026] In order to make the above-mentioned objects, features and advantages of the present invention more clearly understood, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in the various embodiments of the present invention can be combined accordingly without conflicting with each other.

[0027] Example Combine Figure 1As shown, this embodiment provides a resume and job matching analysis method based on four dimensions, which includes the following steps: S1. Obtain the original JD text and perform text preprocessing. It should be noted that the full English name of JD is JobDescription, and its common Chinese meaning is job description.

[0028] In this embodiment, text preprocessing includes the following steps: 1. Format cleaning: Use regular expressions to remove HTML tags and special symbols, and unify line breaks to spaces. Special symbols are symbols that are difficult to enter directly, such as mathematical symbols, unit symbols, tabs, etc. 2. Sentence segmentation: Sentence segmentation is performed based on the NLTK sentence segmenter and custom punctuation rules (such as semicolons and periods); 3. Stop word filtering: Load a custom stop word list, such as meaningless words such as "company" and "position".

[0029] S2. Establish four-dimensional information and extract information based on the four-dimensional information; the four-dimensional information includes basic information, educational information, work experience and technology stack.

[0030] In this embodiment, the process of information extraction is as follows: (1) The technical implementation process of basic information is as follows: Named Entity Recognition (NER): Use the spaCy pre-trained model to extract company names, job titles, etc.; Keyword Matching: Use regular expressions to extract work location, salary range, number of job openings, etc.; After extracting basic information, the output fields are: full job title, company, work location, salary structure, number of job openings, etc. (2) The technical implementation process of academic information is as follows: Hierarchical education model: Constructs a three-level mapping table of "education level-schooling system-major"; Dependency parsing: Identifies the education descriptions following trigger words such as "requirement" and "need"; Output fields after education information extraction include: minimum education requirements, major restrictions, schooling system requirements, GPA threshold, etc. (3) The technical implementation process of the work experience is: Sequence labeling model: Uses the BIOES labeling system to distinguish between responsibilities (R) and achievements (A); Time expression recognition: Uses DateParser to parse time descriptions such as "more than three years" and "fresh graduate"; Industry entity linking: Uses a custom industry dictionary to match keywords in fields such as "financial technology" and "intelligent manufacturing"; Output fields after work experience extraction include: experience requirements, target industry, core responsibilities list, performance indicator library, etc. (4) The technical implementation process of the technology stack is: Domain-specific vocabulary: Build a three-layer vocabulary that includes technical terms (such as "Spring Boot"), tool versions (such as "Python 3.8+"), and framework architecture (such as "microservices"); semantic similarity calculation: Use Word2Vec to calculate the cosine similarity between job descriptions and the vocabulary, setting 0.75 as the effective matching threshold; version dependency analysis: Use regular expressions to extract technical requirements after modifiers such as "≥" and "proficient"; the output fields after technology stack extraction are: programming language, framework tool, database, technical architecture, certificate requirements, etc.

[0031] S3. Establish a mapping relationship between JD extraction fields and the four dimensions of resumes; In this embodiment, when the information in the four dimensions of the resume is basic information, the corresponding JD extraction fields are the full title of the position, the company to which it belongs, and the work location, and the corresponding mapping rule is direct mapping; When the information in the four dimensions of a resume is educational background, the corresponding JD extraction fields are the minimum educational background requirements and major restrictions. The corresponding mapping rule is the educational background field type conversion, such as "Bachelor's Degree" → "Bachelor's Degree"; When the information in the four dimensions of a resume is work experience, the corresponding JD extraction fields are experience requirements, target industry, and core responsibilities. The corresponding mapping rule is structured job description, such as verb + object + achievement. When the information in the four dimensions of the resume is the technology stack, the corresponding JD extraction fields are programming language, framework tools and certificate requirements, and the corresponding mapping rules are reorganized according to the three-level structure of "technology category-specific requirements-proficiency".

[0032] In this example, the purpose of setting up step S3 is to solve the semantic alignment problem of "unstructured JD text" and "resume structured data" by establishing standardized mapping rules between JD extracted fields and the four dimensions of resumes. For example, "3 years or more of internet industry experience required" in the JD can be mapped to "Experience requirement: 3 years" and "Target industry: internet" in the work experience section. Essentially, this defines matching objects and rules to ensure that JD requirements and resume information are comparable along the same dimensions.

[0033] S4. Match resumes based on four dimensions; In this embodiment, the basic information matching process is as follows: S41. Verify key information first: First, use an exact string comparison algorithm (such as Lucene's TermQuery) and regular expression pattern matching (such as the Python re module) to accurately match core fields. Core fields such as name, contact information, and email address must be completely consistent; otherwise, the basic information module will directly score 0. For non-core fields (such as address information), use spaCy's address parsing function parse_address() to extract the city name and use a fuzzy matching algorithm (such as Levenshtein distance) for comparison. Equivalent expressions such as "Beijing" and "Beijing" are considered a match (similarity ≥ 90%).

[0034] S42. After completing the key information verification, evaluate the integrity and format standardization of the basic information: Field coverage: Refer to the Likert five-point scale method in "Resume Analysis Techniques". Missing core fields (such as name) will result in a 40-point deduction, and missing non-core fields (such as email address) will result in a 15-point deduction. Format compliance: Regular expressions are used to verify the email format (for example, ^[a-zA-Z0-9_.+-]+@[a-zA-Z0-9-]+\.[a-zA-Z0-9-.]+$) and phone number validity (for example, mobile phone numbers match ^1[3-9]\d{9}$). Fields that do not meet the standards will be scored 0.

[0035] S43. Comprehensively consider key information verification, field coverage, and format standardization scores to generate the final score for the basic information module. For example, if the resume contains a complete name, phone number, and email address and is formatted correctly, the score is 100 points. If email information is missing, 15 points will be deducted according to the rules, and the final score will be 85 points (this deduction rule must be linked to the weight of the basic information module. For example, if the module weight is 15%, the actual score included in the comprehensive score is 85×15%=12.75 points).

[0036] To summarize further, the logic of S41-S42 is: First, ensure exact match, then completeness assessment: Prioritize accuracy (exact match) for core fields (name / contact information, etc.). If core fields are missing or mismatched, the significance of the subsequent completeness score will be significantly reduced. For example, if the resume name and the JD applicant's name do not match (failed exact match), even if other information is complete, the overall basic information score should still be 0.

[0037] Then, fuzzy matching is performed: address information is a non-core field, and fuzzy matching (such as city name resolution) meets the needs of recruitment scenarios (such as accepting "Beijing" and "Beijing City" as equivalent), and should be performed after the core field is accurately matched.

[0038] To summarize further, the logic of S42-S43 is: first calculate the deductions, then summarize the results. Specifically, field coverage and format compliance are used as the basis for deductions. S43 generates the final score based on the results of the first two steps (e.g., 100 base points - 15 points for missing email address = 85 points), conforming to the "calculation process → result output" logic.

[0039] In this embodiment, the matching process of educational background information is as follows: Step 1: Map the subject classification database of the Ministry of Education: (1) Definition of professional standardization: Through the "Catalogue of Disciplines for Degree Awarding and Talent Cultivation (2022)", a "major name-discipline category" mapping table is established, and sub-majors such as "Computer Science and Technology" are uniformly mapped to "Computer Category" to solve the problem of diversity in major names.

[0040] (2) Logical verification of educational level: ① Basic matching rules: When the JD educational requirement is "Master's degree or above", the "Education Level" in the resume's educational field must be ≥ "Master's degree", otherwise the basic matching score will be 0 (education level definition: PhD > Master's > Bachelor's degree > Associate degree).

[0041] ② Flexible matching rules: Dynamic rules are loaded through the rule engine, such as "When the JD tag is 'management position' and the resume has a bachelor's degree, if the management experience in the work experience is ≥5 years, the basic academic matching degree will be scored 60 points."

[0042] Step 2: Establish a mechanism for adding points based on academic qualifications: (1) Bonus points for prestigious university matches: For applicants whose academic qualifications are listed in the Ministry of Education’s “Double First-Class” university list, an additional 10 points will be added to the basic matching degree (the list is synchronized in real time through the official interface). Specifically, based on the general industry practice in the recruitment field, an additional 10 points can be added for applicants whose academic qualifications are listed in the Ministry of Education’s “Double First-Class” university list (such as Tsinghua University and Peking University) or in the top 50 international higher education institutions. This bonus rule is based on the correlation analysis between the reputation of the university and the job requirements, and is configured through a dynamic rule engine to ensure the fairness and scalability of the scoring model.

[0043] (2) Certificate semantic association: Use Neo4j to build a "position type-skill-certificate" knowledge graph (such as "Internet position → project management skills → PMP certificate"). When a resume contains a certificate associated with the graph, the academic qualification matching degree can be increased by up to 15 points.

[0044] Step 3: Verify academic qualifications: Strict match example: If the JD requirement is "Master's degree or above, computer science major" and the resume's educational background is "Bachelor's degree, computer science and technology": For example, in step 1, mapping: "Computer Science and Technology" → "Computer Science" (major match), but the educational level "Bachelor's degree" is less than "Master's degree" (level mismatch), the basic match score is 0; the verification result in step 3 is: educational background match score 0.

[0045] Flexible matching example: If the JD requirement is "MBA preferred, bachelor's degree or above", and the resume's educational background is "Bachelor's degree, Marketing + 8 years of management experience": For example, in step 1, if the undergraduate degree meets the "Bachelor's degree or above" requirement, the basic matching score is 60 points; in step 2, if the management experience is ≥ 5 years, the score is increased to 80 points; then the verification result in step 3 is: the educational matching score is 80 points.

[0046] Specifically, the logic behind steps 1 and 2 is: standardize first, then differentiate and apply bonus points. Specifically, the Ministry of Education's subject catalog must first standardize the names of majors (e.g., "Computer Science and Technology" → "Computer Science") to ensure consistent academic qualifications. Bonus points rules (e.g., bonus points for prestigious universities, certificate linkage) will then be implemented based on these unified standards. For example, if major names are not standardized, "Computer Science" and "Computer Technology" might be considered different majors, leading to confusion in the application of bonus points rules.

[0047] The logic of steps 2 and 3 is: first apply the rules, then output the results. Specifically, bonus points are adjustment factors in the matching process (such as bonus points for prestigious schools and certificate matching) and must be implemented after the basic matching (step 1) is completed. The overall matching result is finally verified in step 3. In this example, the flexible rule of "bachelor's degree + 5 years of management experience" is dynamically adjusted based on the "strict matching" in step 1 through the rule library in step 2, and the verification result is output in step 3.

[0048] In this embodiment, the work experience matching process is as follows: Step a: Calculate semantic similarity: Using BERT semantic encoding, we converted the job description text (e.g., "recommender system development") into a 768-dimensional semantic vector and calculated the matching degree using cosine similarity (threshold 0.7). Incorporating a dynamic attention mechanism and employing the Transformer² model, we dynamically adjust the weight of attention given to job descriptions based on job type. For example, technical positions focus more on "algorithm optimization," while management positions focus more on "team management."

[0049] Step b: Perform time series verification: Timeline consistency detection: LSTM and CRF models are combined to analyze work history timestamps to detect full-time job conflicts during the same period, such as concurrent employment at two companies from 2022 to 2024. It should be noted that the LSTM model, as a nonlinear model, can be used as a complex nonlinear unit to construct larger deep neural networks. The CRF model is a discriminative probabilistic model, a type of random field, and is commonly used to label or analyze sequence data.

[0050] Based on experience depth assessment: Combined with the working years required for the position, the resume work experience is weighted by time, such as 5 years of experience counts as 100%, and 3 years counts as 60%.

[0051] Step c: Perform work experience scoring verification; For example, similarity calculation is performed: If the resume description is: "Responsible for recommendation system development and optimization of collaborative filtering algorithms"; JD requirements are: "Design recommendation systems and improve algorithm accuracy"; Cosine similarity: 0.92; the corresponding matching degree is 92%.

[0052] In this embodiment, the technology stack matching process is as follows: Step A: Standardize skills and implement an expansion mechanism: Synonym expansion: Build an industry terminology library based on knowledge graph technology (covering 2,000+ technical keywords). For example, "Python development" and "Python programming" are mapped to the same skill node, and fuzzy matching is achieved through semantic association.

[0053] Dynamic adjustment of skill weights: TF-IDF weighting: Calculate the TF-IDF value of skill keywords appearing in the JD, and assign higher weights to high-frequency and unique skills (such as "recommendation system" in AI positions).

[0054] Industry standard mapping: Refer to the technical capability classification standards in the "Information Technology Service Digital Transformation Maturity Model and Assessment" (GB / T43439-2023) to classify skills into levels (such as "expert level" and "proficient level").

[0055] Step B: Calculate the matching degree. The matching logic is: Exact match: 100% match when resume skills are exactly the same as JD skills (e.g. Python).

[0056] Semantic matching: The BERT model is used to calculate the similarity of skill vectors, with a threshold of 0.7 (for example, the similarity between "machine learning" and "deep learning" is 0.8, which is 80%).

[0057] Industry knowledge enhancement: Combine the skill association relationships in the knowledge graph (such as "recommendation system" includes "collaborative filtering" and "matrix decomposition") to expand the matching scope.

[0058] Step C: Perform technology stack scoring verification: For example, skill weight calculation: JD skills frequency: Python (3 times), machine learning (2 times), recommendation system (1 time); TF-IDF scores: Python (0.3), Machine Learning (0.25), Recommender Systems (0.2); Weighted average: (0.3×100%+0.25×80%+0.2×90%) / (0.3+0.25+0.2)≈83.3 points.

[0059] In this embodiment, the purpose of implementing step S4 is to perform matching calculations on specific fields of the resume and JD based on the mapping relationship defined in S3, wherein the mapping relationship is a semantic bridge connecting JD requirements and resume information, ensuring that the matching process is carried out in a unified dimensional space. For example: only by first mapping the "Python ≥ 3.8" in the JD to the "Programming Language: Python" and "Version Requirements: ≥ 3.8" of the technology stack dimension through S3, can the version compatibility check with the "Python 3.9" of the resume technology stack be performed in S4. It should be noted that the order of S3 and S4 cannot be reversed. If the mapping relationship is not established first, direct matching will lead to confusion in the dimensions of JD requirements and resume information (such as mistakenly matching the educational requirements in the JD to the work experience dimension), and accurate matching cannot be achieved; these two steps together constitute the core process from "structured parsing" to "matching execution".

[0060] In this embodiment, refer to Figure 2 As shown, the correlation analysis between steps S3 and S4 is as follows: 1. Mapping rules determine the matching strategy, as shown in Table 1 below: Table 1

[0061] The mapping rules are implemented through a predefined JSON configuration file (such as mapping_rules.json), which includes field mapping paths (such as "JD.education_requirement→Resume.Educational Information.Minimum Educational Qualification") and conversion functions (such as the academic qualification name standardization function standardize_education()).

[0062] 2. Mapping granularity affects matching accuracy: (1) Analysis from the perspective of fine-grained mapping: The S3 step splits the JD requirement of "must be proficient in TensorFlow 2.0+ and familiar with distributed training" into: ① Technology stack dimension - framework tool: TensorFlow; ② Technology stack dimension - version requirement: ≥2.0; ③ Technology stack dimension - extension requirement: distributed training.

[0063] Enable the S4 step to match "TensorFlow 2.3" (version matching) and "Participation in distributed model training projects" (extension requirement matching) in the resume, improving matching accuracy.

[0064] (2) Analysis from the perspective of coarse-grained mapping risks: If the S3 step does not separate the version and extension requirements and directly maps to "Technology Stack: TensorFlow", the S4 step may miss issues such as "version incompatibility" or "lack of distributed training experience", resulting in mismatching.

[0065] In this embodiment, the correlation between steps S3 and S4 is logically verified as follows: The S3 mapping process is as follows: JD technical requirements "Python, machine learning, recommendation systems" → mapped to "Programming language: Python" and "Technical direction: Machine learning, recommendation systems" in the technology stack dimension. JD educational requirements "Master's degree or above" → mapped to "Minimum educational requirement: Master's degree" in the educational dimension.

[0066] The S4 matching process is as follows: Technology Stack Matching: The resume's "Python, TensorFlow, Java" and JD's "Python, Machine Learning, Recommender Systems" were compared through semantic similarity calculation (Python 100%, Machine Learning 80%, Recommender Systems 90%), resulting in an average score of 83.3. Education Matching: The resume's "Bachelor's Degree" does not meet the JD's "Master's" requirement and is therefore scored 0 (since S3 clearly defines the mapping rule for the education dimension as "minimum education requirement," S4 only requires rule verification).

[0067] Conclusion: If the mapping rule of "educational requirements → minimum education" is not established in advance in step S3, step S4 cannot determine whether "bachelor's degree" meets the "master's degree" requirement, and the matching logic will not be established.

[0068] S5. Dynamically assign weights based on the four-dimensional information matching results; In this embodiment, the allocation process is: S51. Identify job types: Building a text classification model: A combined BERT and CRF models were used to classify JD positions (e.g., technical, management, and finance). The model achieved 95% accuracy on a test set containing 50,000 annotated samples. The test set covered 10 industries, including internet, finance, and manufacturing, with a balanced distribution of positions (technical positions accounted for 32%, management positions accounted for 25%, and finance positions accounted for 18%), validating the model's classification effectiveness across multiple domains. It should be noted that the BERT model is a pre-trained language model based on the Transformer architecture. It significantly improves the performance of natural language processing (NLP) tasks through innovative technologies such as bidirectional contextual understanding and the Masked Language Model (MLM).

[0069] Set keyword weights: Automatically adjust module weights based on job type, such as 40% for technical work experience and 30% for technology stack.

[0070] S52. Dynamically adjust the weights based on the recognition results: In this embodiment, the adjustment means adopted in S52 include: 1. Adjustment based on job characteristics: The system automatically applies different weighting templates based on the results of job type identification. For technical positions, work experience and technical stack are crucial for determining candidate suitability, so work experience is weighted at 40% and technical stack at 30%. For management positions, which place greater emphasis on personnel management and overall coordination, work experience may be weighted at 35%, basic information (including soft skills such as communication skills and team management experience) at 30%, and technical stack weighted at 10%. For finance positions, which require high levels of professional knowledge and industry experience, academic qualifications are weighted at 30%, work experience at 25%, and technical stack at 15%. These templates are derived from a large amount of industry recruitment data statistics to meet the actual demand for information in various dimensions for different positions.

[0071] 2. Data-driven optimization: Leveraging annotated data from over 100,000 resumes and job JDs, machine learning models (such as LightGBM) are used to train a weighting strategy. The model analyzes various textual features in the JDs, such as keyword frequency (for example, in a technical position JD, the high frequency of words like "algorithm optimization" indicates the importance of this skill) and the proportion of industry jargon. It also considers job profile characteristics, such as industry (internet, finance, manufacturing, etc.), company size (startups may prioritize practical skills, while larger companies may prioritize academic qualifications and stability), and salary level. Based on these characteristics, the model intelligently fine-tunes weighting. For example, if a startup internet company is hiring, given its need for rapid results, the model may increase the default weighting of work experience by 3%-5%, while correspondingly reducing the weighting of academic qualifications.

[0072] 3. User-defined: To meet a company's specific recruitment needs, the system supports HR customizing weighting templates. For example, if a company is recruiting high-end R&D talent and considers a technical stack fit crucial, they can increase the weighting of the technical stack to 45% or even higher through the system interface. The system utilizes a microservices architecture, allowing HR adjustments to be instantly synchronized with the algorithm, ensuring that weightings are flexibly adapted to the company's current recruitment priorities.

[0073] It should be noted that industry adaptation supports user-defined weight templates, such as a 20% weight for financial position certificates, and dynamic loading is achieved through the microservice architecture.

[0074] Data-driven: Based on over 100,000 annotated resumes / JDs, a machine learning model optimized weight distribution (F1 score of 0.92) was used. It should be noted that the F1 score, also known as the Chinese F1 score, is a core metric used in machine learning to evaluate model classification performance. It is a weighted harmonic mean of precision and recall. Precision is the ratio of samples correctly predicted as positive to the total number of predicted positive samples. For example, in a job classification task, precision indicates the proportion of samples predicted as "technical position" that are actually technical positions. Recall is the ratio of samples correctly predicted as positive to the total number of actual positive samples. For example, recall indicates the proportion of samples correctly identified as "technical position" by the model as technical positions. The F1 score balances precision and recall and ranges from 0 to 1, with higher values indicating better model performance. The "F1 value of 0.92" in this solution indicates that the weight allocation model, trained with over 100,000 resume / JD annotated data, achieved a high combined precision and recall of 92% in the job type classification task. This demonstrates that the model accurately judges job types and can reliably provide a basis for dynamic weight adjustment.

[0075] In this embodiment, the specific dynamic adjustment logic is: 1. The triggering conditions for dynamic adjustment are as follows: The core basis for dynamic weight adjustment is the position type identification result (such as the "technical position" label of AI engineers). The system achieves "dynamicity" through the following methods: 1. Predefined template matching: The "Technical Position Weighting Configuration: Work Experience 40%, Technology Stack 30%" in the file is a default template. This template already has implicit dynamic adjustment logic—that is, different default weightings are assigned to different position types (for example, management positions and finance positions have different weighting ratios).

[0076] Example: When the position type is identified as "AI Engineer" (technical position), the system automatically calls the default template for the technical position without manual intervention, reflecting the "dynamic loading" feature.

[0077] 2. Data-driven fine-tuning: The document mentions "optimizing weight distribution through machine learning models based on over 100,000 resume / JD annotated data." This indicates that the weight ratio of the default template is not fixed, but is adaptively optimized based on industry / job characteristics through data training.

[0078] 3. Implicit logic: For example, if data statistics show that a certain type of technical position (such as "recommendation algorithm engineer") places more emphasis on "technology stack" matching, the model may fine-tune the technology stack weight from the default 30% to 35%, and adjust the work experience weight accordingly to 35%, forming a dynamic weighting scheme for segmented positions.

[0079] 2. Specific implementation of dynamic adjustment: 1. Mapping between job types and weight templates: The document's "Job Type Identification" section mentions using the BERT+CRF model to classify JDs, outputting labels such as "Technical Position, Management Position." Combined with the description of "Dynamic Weight Assignment," we can infer that each job type label corresponds to a predefined weight template, and the weight ratio in the template is derived based on historical recruitment data for similar positions.

[0080] Example of a technical position template: work experience 40%, technology stack 30%, education 15%, basic information 15%.

[0081] Example of a management position template: work experience 35%, basic information (such as communication skills) 30%, education 25%, and technology stack 10%.

[0082] 2. Dynamically adjusted "data-driven" mechanism: The "technical solution" section of the document mentions "dynamic weight algorithm" and "job profile correction score". It can be understood that the system analyzes the keyword density in the JD (such as the number of occurrences of "algorithm optimization"), industry terms (such as "financial technology") and other features to make local fine-tuning of the weight of the default template.

[0083] For example, if the AI engineer's JD frequently mentions "PhD preferred" (a keyword related to academic qualifications), the weighting of academic qualifications might be increased from the default 15% to 20%, and the weighting of the technical stack might be reduced to 25%. If the JD emphasizes "more than 5 years of project experience" (related to work experience), the weighting of work experience might be increased from 40% to 45%, and the weighting of academic qualifications might be reduced to 10%.

[0084] 3. Dynamically adjusted result output: The document "Example 1" demonstrates the technical position matching process, where "Education Match: Bachelor's Degree (Does not meet the master's degree requirement) → 0 points" indirectly reflects the dynamic impact of weight distribution. Since the default education weight for technical positions is 15%, education mismatch directly leads to a lower comprehensive score (for example, the comprehensive score in this case = 63.4 points).

[0085] Logical chain: Position type → trigger technical position weight template → educational qualification accounts for 15% → when educational qualification does not match, points are deducted according to weight, reflecting the mechanism of "dynamic weight affecting the final score".

[0086] S6. Comprehensive scoring based on dynamic weight distribution information. The idea is to calculate a comprehensive score by combining the resume's four-dimensional matching results with the JD's (basic information, education, work experience, and technical stack), combined with the dynamically adjusted weights of the position type, and using explainable design to help HR understand the matching logic. The specific process is as follows: S61. Use a linear weighted formula to perform formula verification and threshold control processing: The S61 process is as follows: 1. Calculate the comprehensive score: The system automatically obtains the matching score for each dimension (e.g., 85 points for technology stack matching and 92 points for work experience matching) and the corresponding weight (e.g., 30% for technology stack matching and 40% for work experience matching). Multiply the matching score for each dimension by its weight and add them together to obtain the comprehensive score. Example: 85 points for technology stack × 30% + 92 points for work experience × 40% + 0 points for education × 15% + 100 points for basic information × 15% = 63.4 points (consistent with the results of Example 1 in the document). It should be noted that the formula verification can refer to the addition formula in "Resume Analysis Technology". The comprehensive score S = Σ(Wi × Mi), which is applicable to most job types.

[0087] 2. Set screening thresholds: The system defaults to a minimum match threshold (e.g., 60 points). Resumes with a comprehensive score that meets or exceeds this threshold enter the initial screening process. HR can adjust the threshold based on job requirements (e.g., lowering it to 50 points for scarce positions and raising it to 70 points for popular positions). Threshold control: Set a minimum match threshold (e.g., a comprehensive score of ≥60 points for initial screening) to prevent low-quality resumes from being rejected.

[0088] S62. Design for Interpretability: Specifically, the S62 process is to generate a matching detailed report: clearly displaying the matching scores and key basis of each dimension, for example: technology stack matching degree 85%, matching keywords: Python, TensorFlow; work experience matching degree 92%, matching responsibilities: "Recommendation system development" is highly correlated with "algorithm optimization" in JD.

[0089] It's important to note that when designing for explainability, the matching degree and key matching points of each module must be output, such as "Technology stack matching degree 85%, main matching keywords: Python, TensorFlow." Attribution analysis is also required, using a knowledge graph to demonstrate matching paths, such as linking "recommender system development" to "algorithm optimization" in JD. This process not only addresses existing technical shortcomings but also supports the rationality of the overall scoring.

[0090] In this embodiment, the explainability design and comprehensive scoring for step S62 together constitute a closed-loop logic of "score calculation → result verification → business application". The correlation between the two is reflected in the following aspects: 1. Interpretable Design is a “Verification Tool” for Comprehensive Scoring 1. Verify the rationality of the scoring results: The overall score (calculated by S61) is a quantitative result of the four-dimensional match and dynamic weighting. Interpretability design verifies whether the score aligns with the job requirements by breaking down the contribution of each dimension. For example, if the overall score is 63.4 points (for a technical position), and the interpretability report shows "Education Match 0" (because the bachelor's degree does not meet the master's requirement), the low score is primarily due to the academic qualifications, not the lack of a good match in the technology stack or work experience, thus verifying the rationality of the scoring results.

[0091] 2. Identify unusual matching scenarios: By displaying matching details for each dimension, you can identify anomalies in the score (e.g., a dimension with a significantly lower matching score than others), helping to troubleshoot data or algorithmic issues. For example, if the technology stack match is only 30% but the overall score is 75, you need to check whether the technology stack weight has been incorrectly increased, or whether the matching scores of other dimensions are inflated.

[0092] 2. Explainable Design Provides “Business Semantics” for Comprehensive Scores 1. Translate quantitative results into business language: A comprehensive score (e.g., 63.4) is an abstract numerical value. Interpretability design translates this numerical value into understandable hiring logic for HR by describing key matching points (e.g., "Technology stack matching keywords: Python, TensorFlow") and skill association paths (e.g., "Recommendation system development → Collaborative filtering algorithm → JD requirements"). This helps HR avoid doubts about the fairness of the algorithm due to the fact that the score is just a number, and enhances trust in the automated screening results.

[0093] 2. Prioritization Support for Hiring Decisions: Explainable reports reveal the actual impact of each dimension on the overall score, helping HR adjust their screening strategies based on the specifics of the position. For example, if the JD emphasizes "educational qualifications first," but the weight of "educational match" in the overall score is low (e.g., 15%), HR can use the report to identify any discrepancies between the weighting settings and the job requirements, manually adjust the weights, and recalculate the score. If multiple resumes have similar overall scores, HR can prioritize candidates who perform better in key dimensions based on detailed dimensions such as "technology stack match" and "work experience relevance."

[0094] 3. Explainable Design Strengthens the Compliance of Comprehensive Scores 1. Meet data transparency requirements: In the context of privacy protection and algorithm compliance (such as GDPR), explainable design uses attribution analysis (such as the "matching path knowledge graph") to prove that the scoring results are not randomly generated, but are based on the semantic association between resumes and JDs, meeting the "traceable and verifiable" compliance requirements.

[0095] 2. Reduce the risk of discrimination in recruitment: If the comprehensive score involves sensitive dimensions (such as education level and age), the explainable report must clearly demonstrate the matching logic of this dimension (for example, "education level requirement is a master's degree or above" is a clear JD requirement and not algorithmic bias) to avoid legal disputes caused by "black box scoring."

[0096] 4. Verification of association relationships in files 1. Implicit associations in the embodiment: In the file "Example 1", the calculation result of the comprehensive score of 63.4 points can be further broken down through explainable design into: technology stack matching degree 83.3 points × 30% weight = 25 points; work experience matching degree 92 points × 40% weight = 36.8 points; educational qualification matching degree 0 points × 15% weight = 0 points; basic information matching degree 100 points × 15% weight = 15 points; these detailed data are directly related to the composition of the comprehensive score, reflecting the support of explainable design for the scoring results.

[0097] 2. Clear direction in technical advantages: The document's "Technical Advantages" section mentions "Interpretability Module: Providing a detailed report on matching," which complements the comprehensive score in terms of functionality. The former addresses the question of "whether the score is accurate," while the latter addresses the question of "why the score is so."

[0098] S7. Verify and output the results to ensure the reliability of the conclusions.

[0099] In this embodiment, a logical verification method is used to check the rationality of the years of experience and educational requirements (for example, if the years of experience corresponding to "PhD" is ≤ 2 years, a warning will be issued). The core goal of this logical verification is to verify the rationality of the comprehensive scoring results and the matching data of each dimension through preset business rules, ensure that the system output conforms to the basic logic of the recruitment scenario, and avoid abnormal results such as "high scores but logical contradictions".

[0100] Furthermore, the specific logic verification rules and implementation process are as follows: Verification 1. Verification of the correlation between academic qualifications and years of experience: Rule definition 1: When the JD specifies educational requirements, the resume's educational requirements and years of experience must meet general industry logic (such as the document example "PhD corresponding experience ≤ 2 years").

[0101] Other typical rule definitions 2: (1) Bachelor's / Master's degree: Fresh graduates (no experience): if the academic qualifications match and the years of experience = 0 years, they will be allowed to pass; non-fresh graduates: the years of experience must be ≥ the average starting years in the industry (e.g. technical positions ≥ 1 year, management positions ≥ 3 years).

[0102] (2) For higher vocational colleges / junior colleges: The years of experience must be ≥ the JD requirement (because the industry places more emphasis on practical skills).

[0103] Correspondingly, the technical implementation process of verification one is: 1. Rule engine: Store logical rules through a custom rule library (such as JSON format). Example: json { "Education Verification Rules": [ { "Education": "PhD", "Experience Years Threshold": 2, "Verification logic": "if years of experience > threshold then warn" }, { "Education": "Master", "Experience Years Threshold": 1, "Verification logic": "if years of experience < threshold then prompt 'years of experience is too low'" } ] } 2. Data trigger: The system extracts "education level" (such as PhD) and "years of experience" (such as 3 years) from the resume parsing results and matches them with the rule library for verification.

[0104] 3. Output response: If the "PhD + 3 years of experience" rule is triggered, the system will generate a warning message "There is a logical contradiction between academic qualifications and years of experience. Manual review is recommended."

[0105] Verification 2: Verification of the correlation between skills and job types Rule definition: Resumes for technical positions (such as AI engineers) must contain at least one core technology stack keyword (such as Python, TensorFlow), otherwise the matching degree will be forcibly reduced.

[0106] Resumes for management positions (such as marketing director) must include soft skill keywords such as "team management" and "project coordination", otherwise the basic information matching degree will be ≤50 points.

[0107] Correspondingly, the technical implementation process of verification 2 is: Keyword mapping: associate "job type-core skills" through knowledge graph (such as technical position → algorithm development, model training).

[0108] The verification logic code is implemented as follows: def validate_skill_relevance(job_type, skill_list): Core skills = Knowledge Graph Query (Job Type, "Core Skills") if not set(core skills)&set(skill list): return "Warning: Lack of core skills, matching may be inflated."

[0109] Verification 3: Timeline consistency check (work experience conflict detection) Rule definition: 1. In the same resume, the time periods of two full-time jobs must not overlap (e.g., working for two companies simultaneously from 2022 to 2024).

[0110] 2. Internship / part-time experience can be included in parallel with full-time experience, but must be marked with the "internship" or "part-time" label.

[0111] Correspondingly, the technical implementation process of verification three is: Use LSTM+CRF model to parse work experience timestamps and generate timeline vectors; Detect conflicts through interval overlap algorithms (such as calculating the intersection length of two time intervals). Python running example: def check_time_conflict(experience1, experience2): if experience1.endtime > experience2.starttime and experience2.endtime > experience1.starttime: return "Conflict: Two experiences overlap."

[0112] Verification 4: Threshold and weight rationality verification Rule definition: 1. The overall score threshold (e.g., 60 points) must be lower than the weighted sum of the full scores for each dimension (e.g., if the total weight of each dimension for a technical position is 100% and the full score is 100 points, the 60-point threshold must be less than 100 points).

[0113] 2. The weight of a single dimension should not exceed 60% (to avoid excessive influence of a single dimension on the score).

[0114] Correspondingly, the technical implementation process of verification four is: Add verification logic to the dynamic weight distribution module (S5), such as running in Python: def validate_weight(weight list): if max(weight list)>60%: return "Warning: The weight of a single dimension is too high, adjustment is recommended" if sum(weightlist) != 100%: return "Error: weights do not sum to 100%". Specifically, in this embodiment, the execution process of the aforementioned logic check is: Data preprocessing: Extract data such as education level, years of experience, skill list, timeline, etc. from the comprehensive scoring results.

[0115] Rule matching: Call the corresponding rule subset based on the position type (such as technical position, management position) (for example, technical positions focus on skill verification, and management positions focus on the relationship between experience and academic qualifications).

[0116] Conflict detection: Execute checks in the order of rules (e.g., check timeline conflicts first, then check the relationship between education and experience), and record all triggered warnings or error messages.

[0117] Result output: No conflict: Generate normal JSON format output data; Conflict: Attach a verification report to the output data, such as running the code: json { Overall rating: 63.4, "Verification result": [ { "Type": "Education-Experience Conflict", "Details": "Doctoral degree corresponds to 3 years of experience (threshold 2 years), manual review is recommended" } ] }.

[0118] In this embodiment, an integrity check method is used to ensure that the required fields of the four dimensions (job title, years of experience, and main technology stack) are not empty; after the check, JSON format data that complies with the HR system standard is generated, including a reserved interface for extended fields.

[0119] Specifically, the core goal of completeness verification is to ensure that all required fields in the four dimensions of information between the resume and JD are complete, preventing matching failures or inaccurate scoring due to missing key data. The document explicitly mentions required fields including: position title (basic information dimension), years of experience (work experience dimension), and primary technology stack (technology stack dimension).

[0120] Furthermore, the specific integrity verification rules and implementation process are as follows: 1. Basic information dimension: job title verification Verification rules: The resume must contain a "Position Title" field (e.g., "AI Engineer," "Marketing Director"), and this field must be extracted from the resume text using a named entity recognition (NER) algorithm (mentioned in the "Four-Dimensional Information Extraction" section of the document).

[0121] The "Position Name" must be clearly marked in the JD (e.g. the title line should include "Recruitment Position: XX Engineer"), otherwise the JD information will be deemed incomplete and the applicant will be refused entry into the matching process.

[0122] Technical implementation: ① Resume parsing stage: Use the named entity recognition function of spaCy or BERT model (mentioned in step S41 of the file) to extract the job title. If no valid name is detected (such as the field is empty or the value is "None"), it will be marked as "Basic information missing".

[0123] ②JD preprocessing stage: Use regular expressions to match the title line (such as "^Recruitment position:.*") or keywords (such as "Position:" "Position:"). If the position name is not matched, the system prompts "Please supplement the JD position name."

[0124] 2. Work experience dimension: experience years verification Verification rules: The "Years of Experience" field in the resume must be parsed using a time expression recognition algorithm (such as DateParser, mentioned in step b of the document), and the value must be ≥ 0 years (fresh graduates can be considered to have 0 years).

[0125] If the JD explicitly requires “years of experience” (e.g. “3 years or more”), a description of the effective time must be included; otherwise, the “default minimum requirement” will be applied (e.g., “years of experience ≥ 0 years”).

[0126] Technical implementation: During the Yi resume parsing phase, the work experience text is analyzed over time (as mentioned in step b of the document). Keywords such as "more than X years" and "fresh graduate" are extracted and converted to numeric fields (e.g., "fresh graduate" → 0 years, "more than 3 years" → 3 years). If parsing fails (e.g., a field is empty), it is marked as "Missing years of work experience."

[0127] ②JD parsing stage: Use regular expressions to match time units (such as "years" and "months"). If the JD does not mention the experience requirement, the system automatically fills in the default value (such as "experience ≥ 1 year" for technical positions and "≥ 3 years" for management positions).

[0128] 3. Technology stack dimension: main technology stack verification Verification rules: The "Main Technology Stack" field in the resume must contain at least one valid technical keyword (such as "Python" or "TensorFlow") and be verified as a valid skill node through the industry terminology library (mentioned in step A of the document).

[0129] The "Technology Stack Requirements" field in the JD must contain at least one technical keyword, otherwise it will be considered that the job requirements are unclear and need to be supplemented by HR.

[0130] Technical implementation: ① Resume parsing stage: The cosine similarity between the text and the technical vocabulary is calculated through the Word2Vec model (mentioned in the document technology stack extraction module). If the highest similarity is less than 0.75 (the threshold of step B in the document) or there is no matching keyword, it is marked as "technology stack missing".

[0131] ②JD parsing stage: Use the TF-IDF algorithm (mentioned in step A of the document) to extract high-frequency technical keywords. If no valid technical terms are identified (such as "no special technical requirements"), the system prompts "Please supplement the JD technology stack requirements."

[0132] Furthermore, the integrity check execution process includes: 1. Field extraction and standardization: Extract four-dimensional information from the resume and JD, and standardize it according to the file's "four-dimensional information extraction" rules (for example, unify academic qualifications into subject categories and map technology stack keywords to terminology nodes).

[0133] 2. Required field null value detection: Check whether the job title, years of experience, and main technology stack fields are null or empty strings. If so, a missing warning will be triggered.

[0134] 3. Validity Verification: Position Title: Must conform to industry-standard naming conventions (e.g., "XX Engineer," "XX Director"), and no meaningless characters (e.g., "Position: Negotiable") are allowed. Years of Experience: Must be a non-negative integer and logically valid (e.g., "-1 year" is invalid). Primary Technology Stack: Must fall within the scope of the terminology library (e.g., "Python" is valid, but "Programming" is invalid due to semantic ambiguity).

[0135] 4. Result Handling: Resume completeness failure: Rejected from the match, with a prompt stating "Please provide additional job title / years of experience / main technology stack." JD completeness failure: The match process is suspended, and HR is notified to complete the information and resubmit.

[0136] In some other embodiments, referring to Figure 3 As shown, a four-dimensional resume and job matching analysis system is also provided, which is used to implement the above resume and job scoring method. Specifically, the system includes a preprocessing module, a four-dimensional information extraction module, a matching calculation module, an interpretability module, and a verification and output module.

[0137] Among them, the preprocessing module transmits the processed data to the four-dimensional information extraction module. Specifically, the preprocessed standardized text (such as the JD content after sentence segmentation) is simultaneously input into the four extraction modules of basic information, education, work experience, and technology stack to achieve parallel parsing.

[0138] The four-dimensional information extraction module transmits data to the matching calculation module. Specifically, the structured data output by each extraction module (such as basic information fields and academic qualification standardization results) are aggregated into the matching calculation module: (1) The basic information extraction module transmits fields such as "job title" and "company" for basic information matching (such as fuzzy matching of work location); (2) The academic qualification information extraction module transmits "minimum academic qualification" and "major category" for academic qualification dimension matching calculation (such as whether "bachelor's degree" meets the "master's degree and above" requirement); (3) The work experience extraction module transmits "years of experience" and "core responsibilities" for semantic similarity calculation and timeline verification; (4) The technology stack extraction module transmits "technical keywords" and "version requirements" for technology stack matching calculation (such as whether "Python 3.9" meets "≥3.8").

[0139] The matching calculation module outputs data to the explainability module and the verification and output module. The explainability module then transmits data to the verification and output module. Specifically, the matching calculation module outputs data such as the matching degree of each dimension, dynamic weight, and comprehensive score, which are then passed to the explainability module (to generate a matching report) and the verification and output module (to verify logical consistency). The matching details generated by the explainability module (e.g., "Technology stack match 85%") are combined with the verification results from the verification and output module (e.g., "No timeline conflicts") before being output to the HR system or front-end interface.

[0140] The preprocessing module further converts the raw JD text and resume into structured text by performing operations such as formatting, sentence segmentation, and stop word filtering. This process involves using regular expressions to remove HTML tags, performing sentence segmentation using NLTK or spaCy, and loading a custom stop word list to filter out invalid terms (such as "company" and "position"). The final output is a list of standardized text paragraphs (e.g., a collection of JD sentences after sentence segmentation).

[0141] The four-dimensional information extraction module includes a basic information extraction module, an educational information extraction module, a work experience extraction module, and a technology stack extraction module. The basic information extraction module extracts basic information such as name, contact information, job title, and company name from text. Its technical implementation involves extracting entities such as organization names (ORG) and person names (PER) through named entity recognition (NER), such as BERT or spaCy models, and matching phone and email address formats using regular expressions. The module ultimately outputs structured basic information fields (such as full job title, company affiliation, and work location). The educational information extraction module parses information such as educational level (bachelor's / master's, etc.), major, and duration of study, mapping and standardizing it with the Ministry of Education's subject classification database. Its technical implementation involves constructing a three-level model: "education level-duration-major." It then uses dependency parsing to identify educational descriptions following trigger words like "required" and "needed" (e.g., "Master's, computer science major"). The module ultimately outputs standardized educational information (such as minimum educational requirements and major category). The work experience extraction module extracts information such as company name, position, job description, and timeline, verifying temporal consistency. Its technical implementation involves using the BIOES sequence annotation system to distinguish responsibilities and achievements, parsing time expressions with DateParser, and detecting time conflicts (e.g., two full-time jobs held during the same period) using an LSTM+CRF model. It ultimately outputs structured work experience (e.g., years of experience, target industry, and a list of core responsibilities). The technology stack extraction module identifies technical keywords, framework tools, and certification requirements, supporting synonym expansion and version parsing. Its technical implementation involves building an industry terminology library (over 2,000 technical keywords) based on a knowledge graph, calculating semantic similarity using Word2Vec or BERT, and extracting version requirements (e.g., "Python ≥ 3.8") using regular expressions. It ultimately outputs a standardized technology stack list (e.g., programming languages, framework tools, and certifications).

[0142] The matching calculation module calculates the degree of match in each of the four dimensions and generates a comprehensive score. It includes a semantic similarity calculation module, a dynamic weight assignment module, and a comprehensive scoring module. The semantic similarity calculation module uses the BERT model to generate semantic vectors and matches text using cosine similarity (e.g., the similarity between job descriptions and JD responsibilities). The dynamic weight assignment module automatically adjusts the weights of each dimension based on the position type (e.g., technical / management positions) (e.g., work experience is weighted 40% for technical positions) and supports user-defined templates. The comprehensive scoring module calculates the total score using the linear weighting formula S = ∑(Wi × Mi), setting a threshold to filter out low-matching resumes.

[0143] The explainability module is used to generate detailed matching reports, displaying scores for each dimension and key matching points, and tracing matching paths through the knowledge graph. Its technical implementation is as follows: outputting a JSON-formatted report containing technology stack matching keywords, work experience semantic similarity values, etc., and using the Neo4j graph to display skill associations (such as "recommendation system" → "collaborative filtering algorithm").

[0144] The verification and output module is used to verify data integrity and logical consistency, and output results that meet HR system standards. Its verification content includes: (1) Completeness check: Ensure that required fields such as job title, years of experience, and main technology stack are not empty.

[0145] (2) Logical verification: Check the rationality of academic qualifications and years of experience (e.g., the corresponding years of experience for a "PhD" is ≤ 2 years), conflicts in work timelines, etc.

[0146] The final output of the verification and output module is in JSON data format, including comprehensive scores, matching details of each dimension, verification results, etc., supporting seamless integration with the HR system.

[0147] Specifically, in this embodiment, the implementation process of the preprocessing module is: import re import spacy nlp = spacy.load("en_core_web_sm") stopwords = set(open("stopwords.txt", "r").read().split()) def preprocess_text(jd_text): # Format cleaning cleaned_text = re.sub(r'<.*?>', '', jd_text) # Remove HTML tags cleaned_text = re.sub(r'[^\w\s\u4e00-\u9fa5]', '', cleaned_text)# Keep Chinese, English and numbers # Clause processing sentences = [sent.text for sent in nlp(cleaned_text).sents] # Stop word filtering filtered_sentences = [[token.text for token in nlp(sent) iftoken.text not in stopwords] for sent in sentences] return filtered_sentences. The implementation process of the basic information extraction module is as follows: from spacy.matcher import Matcher matcher = Matcher(nlp.vocab) # Define company name matching rules company_pattern = [{"ENT_TYPE": "ORG"}] matcher.add("COMPANY", [company_pattern]) def extract_basic_info(sentences): doc = nlp(" ".join([" ".join(sent) for sent in sentences])) matches = matcher(doc) company_names = [doc[start:end].text for match_id, start, end inmatches] # Extract job title (assuming it is in the title row) job_title = next((sent for sent in sentences if "position" in sent),[""])[0] return {"job_title": job_title, "company": company_names[0] ifcompany_names else None}. The implementation process of the technology stack extraction module is as follows: import numpy as np from gensim.models import Word2Vec model = Word2Vec.load("tech_embedding.model") def extract_tech_stack(sentences, threshold=0.75): tech_stack = empty list for sent in sentences: for word in sent: if word in model.wv: # Calculate the similarity with the technical vocabulary sim_scores=[np.dot(model.wv[word], model.wv[tech]) for tech inmodel.wv.index_to_key] if max(sim_scores)>= threshold: tech_stack.append(word) return list(set(tech_stack))# Deduplication. In this embodiment, the hardware environment includes a server and a development platform. The server preferably has a 4-core CPU, 16GB of memory, and a 256GB SSD. The development platform can use Python 3.9 as the programming language, TensorFlow 2.12 as the machine learning platform, and spaCy 3.7 as the natural language processing library (NLP).

[0148] In this embodiment, the technical implementation architecture includes at least: Data layer: Supports resume parsing in 21 formats, including PDF, Word, and images (integrated with Tesseract OCR). Specifically, it supports multimodal parsing, integrating Tesseract OCR to parse PDF / image resumes with 98.6% accuracy. It also builds knowledge graphs, integrating industry terminology libraries and the Ministry of Education's subject classification database to support semantic associations between skills and academic qualifications.

[0149] Algorithm layer: Utilizing the Transformer architecture, combined with LSTM and CRF models for sequence tagging (achieving 95.3% accuracy in work experience parsing). Specifically, BERT semantic encoding is implemented, and the BAAI / bge-m3 model is used to generate 768-dimensional semantic vectors, supporting real-time retrieval of millions of resumes. Furthermore, the Transformer² architecture is used to implement job type classification and dynamic weighting adjustments.

[0150] Application layer: Provides APIs to support seamless integration with HR systems (response time < 2 seconds per resume). Specifically, this layer includes APIs that provide HR system integration with a response time < 2 seconds per resume; it also includes visualization tools: using an explainability module to generate match heat maps to assist HR decision-making.

[0151] It should be noted that the technical implementation complies with industry standards and is compliant. For example, in terms of data security, AES-256 encrypted transmission is used to comply with GDPR requirements. For standard compliance, please refer to the technical capability assessment framework in the "Information Technology Service Digital Transformation Maturity Model and Assessment" (GB / T43439-2023).

[0152] Example 1 Based on the solution records of the above methods and systems, taking the technical post as an example, the entire process is as follows: S1. Obtain the original JD text and preprocess it: Specifically, input the JD text "Recruiting AI engineers, responsible for recommendation system development, must master Python, machine learning, master's degree or above, computer-related majors", and use the preprocessing module to clean HTML tags and sentences (such as splitting into "Technical requirements: Python, machine learning, recommendation system" and "Education requirements: Master's degree or above, computer-related majors"), and filter out stop words "recruitment" and "need".

[0153] S2. Establish four-dimensional information and extract information based on the four-dimensional information: For example, the job title in the basic information is extracted: AI Engineer (identified through NER); Company name: not explicitly mentioned (left blank by default or supplemented by resume later).

[0154] The minimum educational requirement in the educational information is: Master's degree or above; professional restrictions: Computer major (combined with the standardized "computer-related majors" in the subject classification database of the Ministry of Education).

[0155] Work experience extraction: No JD work experience requirement (default experience ≥ 1 year).

[0156] Extract technical keywords from the technology stack: Python, machine learning, and recommendation systems (matching the industry terminology library, mapping "recommendation system" to "technology stack - application field").

[0157] S3. Establish a mapping relationship between JD extraction fields and the four dimensions of resumes: Technology stack: JD "Python" → resume "Python" (exact match); "Machine Learning" → resume has no direct keywords, but is associated with "TensorFlow" through the knowledge graph (semantic similarity 80%); "Recommendation System" → resume "Recommendation System Development" (exact match).

[0158] Education: JD "Master's degree or above" → resume "Bachelor's degree" (no match, 0 points).

[0159] S4. Match resumes based on four dimensions: Technology stack matching: Python (100%) + Machine Learning (80%) + Recommender System (90%) → Average score: 83.3 (arithmetic mean).

[0160] Work experience matching: The semantic similarity between the resume "recommendation system development" and the JD "recommendation system" is calculated using BERT, with scores of 0.92 → 92.

[0161] Educational Match: Bachelor’s degree does not meet the master’s degree requirements → 0 points.

[0162] Basic information match: The resume includes name and contact information (assuming it is complete) → 100 points.

[0163] S5. Dynamic weight allocation is performed based on the four-dimensional information matching results: Position type: Technical position → Weight template: Technology stack 30%, work experience 40%, education 15%, basic information 15% (reference document).

[0164] S6. Comprehensive scoring based on dynamic weight distribution information: Formula: S = 30% × 83.3 + 40% × 92 + 15% × 0 + 15% × 100 = 63.4 points.

[0165] S7. Verify and output the results to ensure the reliability of the conclusions: Logical verification: There is no conflict between the academic qualification of "Bachelor's degree" and the experience of "2 years" (technical positions allow bachelor's degree + experience), but the academic qualification does not meet the requirements, triggering the warning "Academic qualification does not meet JD requirements".

[0166] Completeness check: All four required fields (job title, technology stack, and years of experience) are not empty → passed.

[0167] Output: Generate a JSON report containing matching details and warning information for each dimension for HR review.

[0168] Example 2: Based on the above methods and system solutions, taking the management position as an example, the specific process is as follows: S1. Obtain the original JD text and preprocess it: Specifically: Enter the JD text "Recruiting Marketing Director, requiring a bachelor's degree or above, MBA preferred, and more than 5 years of team management experience", and after pre-processing, it will be split into "Education requirements: bachelor's degree or above, MBA preferred" and "Management experience: more than 5 years of team management experience".

[0169] S2. Establish four-dimensional information and extract information based on the four dimensions: For example, basic information extraction: Job title: Marketing Director.

[0170] Educational information extraction: Minimum educational requirement: Bachelor's degree; MBA preferred (additional 10 points).

[0171] Experience requirement: 5 years or more; Core responsibilities: Team management (LSTM+CRF identifies "managing a team of 10 people" as valid experience).

[0172] Technology stack extraction: No clear technical requirements → Default match score is 0.

[0173] S3. Establish a mapping relationship between JD extraction fields and the four dimensions of resumes: Mapping relationship: Education: JD "Bachelor's degree or above" → resume "MBA" (satisfied, and triggers MBA priority points); Work experience: JD “more than 5 years of team management” → resume “5 years of management experience” (exact match).

[0174] S4. Match resumes based on four dimensions: Educational qualification matching: MBA meets "bachelor's degree or above" + priority conditions → basic score 90 points + 10 points → 100 points.

[0175] Work experience matching: 5 years of management experience → 100 points.

[0176] Technology stack matching: No relevant requirements → 0 points.

[0177] Basic information matching: The resume includes name and contact information → 100 points.

[0178] S5. Dynamic weight allocation is performed based on the four-dimensional information matching results: Position type: Management position → Weight template: Education 35%, Work experience 30%, Technology stack 20%, Basic information 15% (reference document).

[0179] S6. Comprehensive scoring based on dynamic weight distribution information: Formula: \(S = 35\% \times 100 + 30\% \times 100 + 20\% \times 0 + 15\% \times 100 = 70\) points.

[0180] S7. Verify and output the results to ensure the reliability of the conclusions: Logical verification: There is no conflict between the MBA degree and the 5 years of experience, and the applicant meets the dual requirements of "degree + experience" for management positions → passed.

[0181] Completeness check: The required fields of job title, education background, and years of experience are complete → passed.

[0182] Output: Generate a matching report, mark it as "obvious educational advantage, complete work experience match", and recommend an interview.

[0183] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions of the technical solution of the present invention by ordinary technicians in this field do not deviate from the essence and scope of the technical solution of the present invention.

Claims

1. A resume and job matching analysis method based on four dimensions, characterized in that: The following steps are involved: S1. Obtain the original JD text and perform text preprocessing; S2. Establish four-dimensional information and extract information based on the four-dimensional information; the four-dimensional information includes basic information, educational information, work experience, and technology stack; S3. Establish a mapping relationship between JD extraction fields and the four dimensions of resumes; S4. Match resumes from four dimensions; S5. Perform dynamic weight allocation based on the four-dimensional information matching results. The allocation process is as follows: S51. Identify the job type. The specific identification process is as follows: Build a text classification model: Use a combination of the BERT model and the CRF model to classify JD job types; Set keyword weight: automatically adjust module weight according to job type; S52. Dynamically adjust the weights based on the recognition results. Specific adjustment methods include: 1) Adjustment based on job characteristics: The system will automatically adapt different weight templates based on the results of job type identification; 2) Data-driven optimization: Utilizing labeled data from multiple resumes and job JDs, and using machine learning models to train weight allocation strategies; 3) Based on user customization: To meet the recruitment needs of enterprises, the system supports HR to customize weight templates; S6. Perform comprehensive scoring based on dynamic weight distribution information; S7. Verify and output the results.

2. A resume and job matching analysis method based on four dimensions according to claim 1, characterized in that: In S1, text preprocessing includes format cleaning, sentence segmentation, and stop word filtering.

3. The resume and job matching analysis method based on four dimensions according to claim 1 is characterized in that: In S3, when the information in the four dimensions of a resume is basic information, the corresponding JD extraction fields are position full name, company, and work location, and the corresponding mapping rule is direct mapping; When the information in the four dimensions of a resume is educational background information, the corresponding JD extraction fields are the minimum educational background requirements and professional restrictions, and the corresponding mapping rules are educational background field type conversion; When the information in the four dimensions of a resume is work experience, the corresponding JD extraction fields are experience requirements, target industry, and core responsibilities, and the corresponding mapping rule is structured job description; When the information in the four dimensions of the resume is the technology stack, the corresponding JD extraction fields are programming language, framework tool, and certificate requirements. The corresponding mapping rule is to reorganize according to the three-level structure of "technology category-specific requirements-proficiency".

4. The resume and job matching analysis method based on four dimensions according to claim 1 is characterized in that: In S4, the matching process of basic information is as follows: S41, first perform key information verification; S42. After completing key information verification, evaluate the integrity and format compliance of basic information; S43. Comprehensively calculate the key information verification, field coverage and format standardization scores to generate the final score of the basic information module.

5. The resume and job matching analysis method based on four dimensions according to claim 1 is characterized in that: In S4, the matching process of educational information is as follows: Step 1: Map the subject classification database of the Ministry of Education; Step 2: Establish a mechanism to increase points based on academic qualifications; Step 3: Perform academic qualification information scoring verification.

6. The resume and job matching analysis method based on four dimensions according to claim 1 is characterized in that: In S4, the matching process of work experience is as follows: Step a, performing semantic similarity calculation; Step b, perform time series verification; Step c: Perform work experience score verification.

7. The resume and job matching analysis method based on four dimensions according to claim 1 is characterized in that: In S4, the technology stack matching process is as follows: Step A: Standardize skills and expand the mechanism; Step B, calculating the matching degree; Step C: Perform technology stack scoring verification.

8. The resume and job matching analysis method based on four dimensions according to claim 7 is characterized in that: In step B, the matching logic for calculating the matching degree includes: Exact match: When the resume skills are exactly the same as the JD skills, it counts as 100%; Semantic matching: Calculate skill vector similarity using the BERT model, with a threshold of 0.7; Industry knowledge enhancement: Combine the skill association relationships in the knowledge graph to expand the matching scope.

9. The resume and job matching analysis method based on four dimensions according to claim 1 is characterized in that: In S6, the process is: S61: Use a linear weighted formula to perform formula verification and threshold control processing; wherein, the process of S61 is: 1) Calculate the comprehensive score; 2) Set the screening threshold; S62. Conduct interpretable design.

10. A resume and job matching analysis system based on four dimensions, characterized by: Execute a resume and job matching analysis method based on four dimensions as described in any one of claims 1 to 9, wherein the resume and job matching analysis system includes a preprocessing module, a four-dimensional information extraction module, a matching calculation module, an interpretability module, and a verification and output module; The preprocessing module transmits the processed data to the four-dimensional information extraction module; the four-dimensional information extraction module transmits the data to the matching calculation module; the matching calculation module outputs the data to the interpretability module and the verification and output module; And the interpretability module transmits the data to the verification and output module.

Images