A resume feature word intelligent extraction method based on semantic enhancement

Abstract

The application discloses a resume feature word intelligent extraction method based on semantic enhancement and belongs to the technical field of natural language processing and information extraction. The method comprises seven processes, namely, resume source file analysis, text preprocessing, semantic enhancement prompt word construction, LLM phased feature extraction, structured output generation, result verification and post-processing, quality evaluation and feedback optimization. The method is compatible with multiple formats of resumes, relies on layout information to assist text blocking, strengthens semantic understanding through multi-level prompt words, extracts and fuses cross-module features in three stages, and combines rules and industry knowledge base to correct model output deviation. The application quantifies the extraction effect by using two sets of confidence formulas, iteratively optimizes the system after manually correcting low-quality samples, solves the problems of poor generalization, weak semantic recognition and high maintenance cost of traditional extraction methods, outputs unified structured data, and is suitable for human resource scenes such as resume analysis, talent pool construction and person-job matching.

Description

Technical Field

[0001] This invention relates to the field of natural language processing and information extraction technology, and in particular to a method for intelligent extraction of resume feature words based on semantic enhancement. Background Technology

[0002] Resumes are the core data carrier for corporate recruitment, talent pool building, and intelligent job matching. However, resumes collected online and offline suffer from pain points such as diverse sources, inconsistent file formats, and significant differences in text expression. Resumes include PDFs, Word documents, TXT files, HTML files, and scanned images. Different job seekers have vastly different writing habits, resulting in numerous issues with similar information, including paraphrasing, omissions, misplaced fields, and colloquial abbreviations. Existing resume feature extraction solutions fall into two categories: Rule-based extraction solutions rely on regular expressions, keyword dictionaries, and manually written matching rules to extract fields. This approach is highly dependent on resume formats and fixed sentence structures; new wording and resume formats require continuous rule iteration. It cannot understand deep semantics, and synonym recognition fails. The manual maintenance of the rule base and feature word list is extremely costly; strict rules are prone to missing information, while lenient rules are prone to incorrect sampling, making it impossible to simultaneously achieve high accuracy and recall.

[0003] Traditional NER entity recognition models rely on labeled datasets for training, can only recognize fixed entity categories, have weak generalization ability, perform poorly in extracting complex fields such as skills, projects, and work experience that are scattered across paragraphs and modules in resumes, and cannot handle free text without standardized sentence structures.

[0004] Current Large Language Models (LLMs) possess strong semantic understanding and structured generation capabilities, providing new ideas for resume extraction. However, directly calling the native inference of LLMs has obvious defects: the output JSON format is disordered, key fields are missing, semantic illusions occur, the same feature is repeatedly extracted across modules without normalization, temporal logic is not consistent, and there are problems such as inconsistent abbreviations of company / school names. It lacks a complete closed loop of pre-text processing, hierarchical extraction constraints, post-validation correction and continuous iterative optimization, and cannot be applied to industrial-grade resume parsing scenarios. Summary of the Invention

[0005] The purpose of this invention is to propose an intelligent resume feature word extraction method based on semantic enhancement. Through semantic enhancement structured prompt word constraints, phased and layered extraction, multi-dimensional post-verification normalization, and sample feedback iterative optimization of the whole process architecture, it can achieve high-precision and high-recall structured extraction of feature words from multi-format non-standard resumes, output unified standardized talent data, and provide reliable underlying data support for talent profiling, job matching, and HR big data analysis.

[0006] To achieve the above objectives, this invention provides a method for intelligent extraction of resume feature words based on semantic enhancement, specifically including the following steps: Step S1: Resume Source File Parsing: Receive the resume source file, identify the file format, and use the corresponding parsing engine to convert it into plain text carrying layout metadata; Step S2: Text preprocessing: Clean the plain text, normalize the format, divide the resume module into blocks, and process excessively long text segments to output standardized segmented text; Step S3: Semantic Enhancement Prompt Construction: Based on the preset extracted field schema and standardized segmented text, assemble structured semantic enhancement prompts containing seven fixed components; Step S4: LLM phased feature extraction: After concatenating the structured semantic enhancement prompts with the standardized segmented text, the large language model is called to perform three levels of extraction in sequence: module recognition, sub-module field extraction, and cross-module feature association and fusion. Step S5: Structured Output Generation: Parse the output of the large language model, map it to a predefined unified data model, and generate a standardized JSON structured resume dataset; Step S6: Post-processing for result verification and completion: Perform format verification, business logic verification, external knowledge base normalization verification, missing field completion, and conflict information resolution on the standardized JSON structured resume dataset; Step S7: Quality assessment and feedback optimization: The quality of extraction is quantified by the confidence calculation formula. Samples that are lower than the threshold for manual review are manually corrected and then fed back to iteratively optimize the prompt words and knowledge base, forming a self-optimizing closed loop.

[0007] Preferably, in step S1, the resume source file format includes PDF, Word, TXT, HTML, and image; for image format resumes, OCR optical character recognition is performed to convert them into text; the layout metadata includes paragraph number, module title, list level, and text block coordinates, and the layout metadata is used to assist in the module segmentation in step S2 and the context embedding in step S3.

[0008] Preferably, in step S2, the text preprocessing execution flow is as follows: remove garbled characters, headers and footers, and redundant blank symbols; unify Chinese and English punctuation, date, and number formats; divide personal basic information, education background, work experience, project experience, professional skills, and self-evaluation modules according to the page metadata; when the text length exceeds the threshold of the large language model context window, use a sliding window within the module to segment the text, and retain a single education, work, or project record completely in a single window.

[0009] Preferably, in step S3, the seven fixed components are, in order, the role definition segment, the task description segment, the field schema definition segment, the semantic enhancement sample segment, the output format constraint segment, the extraction constraint rule segment, and the local context embedding segment; the field schema definition segment uses a JSON template to annotate the field name, semantic interpretation, data type, value constraint, required field identifier, and standard example; the semantic enhancement sample segment stores synonym expressions, time omissions, field misalignments, and irregular resume extraction samples with abbreviated date types.

[0010] Preferably, in step S4, the three-level extraction execution process is as follows: Phase 1: Module recognition, identifying all modules of the resume and their corresponding text character ranges; Phase Two: Module-based field extraction. Feature fields are extracted independently according to the divided module units, and several item structures are output as nested arrays of sub-objects. Phase 3: Cross-module feature association and fusion, locating similar features within multiple modules, and performing deduplication, synonym unification, and summary merging operations.

[0011] Preferably, in step S5, the standardized JSON structured resume dataset contains five core objects: basic personal information, education experience array, work experience array, project experience array, and one-dimensional skills array. The time, education, company, and skills fields are all stored in a unified and standardized manner.

[0012] Preferably, in step S6, three independent knowledge bases are configured: a standard college database, a business database, and a skills dictionary. The knowledge base unification verification performs the following checks: unification of college abbreviations, unification of enterprise aliases, and conversion of skills abbreviations and terms. The business logic verification includes timeline consistency verification, enumeration field value verification, and contact information regular expression format verification.

[0013] Preferably, in step S6, conflict information resolution adopts a priority rule: when multiple modules have conflicting values ​​for the same field, the value from the independent skills column is adopted, followed by the value from the project description, then the value from the work experience, and finally the value from the self-evaluation; the missing field completion process involves carrying existing structured data and calling the large language model for targeted extraction.

[0014] Preferably, in step S7, two sets of confidence calculation formulas are configured, namely, a single-field confidence calculation formula and a document overall confidence calculation formula. Formula for calculating single-field confidence: ; In the formula: For single-field confidence, the range of values ​​is... ; Output the original probability scores of the fields for the large language model; This is a verification pass indicator; The score is determined by the normalized matching of the knowledge base. These are the weighting coefficients, and the sum of the weighting coefficients equals 1. Formula for calculating overall document confidence: ; In the formula: The overall confidence level of a single resume document; This represents the total number of feature fields to be extracted from the resume. For the first Confidence levels for each field; In step S7, configure the manual review threshold. , The value is less than Samples are pushed to the manual annotation platform for correction in real time.

[0015] Preferably, in step S7, the samples after manual correction undergo four types of backflow operations: expanding the semantic enhancement sample library, iteratively optimizing the prompt word template, updating the knowledge base matching words, and using them as training material for fine-tuning the large language model or optimizing the prompt words. The four types of backflow operations are executed simultaneously to optimize the accuracy of resume feature extraction.

[0016] Therefore, the present invention employs the above-mentioned intelligent extraction method for resume feature words based on semantic enhancement, which has the following advantages: (1) Significantly improved generalization ability: It does not rely on fixed resume format or fixed sentence rules, and can be compatible with various custom expressions and multi-format resumes, without the need to continuously add matching rules for new resumes; (2) Outstanding deep semantic understanding ability: Relying on LLM semantic enhancement prompts, it automatically identifies synonym replacements, elliptical sentences, misplaced fields, and unconventional expressions that mix Chinese and English; (3) Simultaneous optimization of extraction accuracy and recall: hierarchical extraction ensures full feature recall, multi-dimensional post-verification and knowledge base normalization eliminate false sampling and illusion problems, taking into account both accuracy and recall. (4) Standardized and unified output data: Forced JSON structured output, unified fields, can be directly connected to talent pool, job matching, HR data analysis downstream systems, without the need for secondary data cleaning; (5) Low operation and maintenance costs: No need for large-scale manual maintenance of the rule dictionary. It relies on manual correction of samples to automatically iterate and optimize prompt words and knowledge base, which significantly reduces the long-term operation and maintenance manpower input.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1This is an overall flowchart of a semantically enhanced resume feature word intelligent extraction method in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0021] Example like Figure 1 As shown in the figure, this embodiment proposes an intelligent method for extracting resume feature words based on semantic enhancement. The specific steps are as follows: S1. Resume Source File Parsing: The system receives resume source files uploaded by candidates, identifies file extensions and encodings to distinguish file types, and adapts its parsing engine to PDF, Word, TXT, HTML, and image resumes respectively, uniformly converting them into plain text containing layout metadata, while simultaneously retaining the original layout structure for subsequent semantic enhancement. (1) PDF file: Use a PDF text parsing tool to extract full text, paragraph pagination, heading levels, and table cell layout information; (2) Word (DOC / DOCX) files: parsing paragraphs, built-in tables, heading levels, and ordered / unordered list hierarchy markers; (3) Plain text TXT file: Read the entire text directly and divide the basic paragraphs according to newline characters and blank lines; (4) HTML resume: Filter JS scripts, CSS styles, and page tags, and keep only the main text and module titles; (5) Image resumes (scanned copies, photo resumes): First, use OCR optical character recognition to output the text, and simultaneously record the position of the image text block.

[0022] The parsed output is plain text carrying page layout metadata, which includes paragraph numbers, module titles, list levels, and text block coordinates, serving as auxiliary context for subsequent module division and semantic judgment.

[0023] S2. Text preprocessing: Perform cleaning, format normalization, resume module segmentation, and long text fragmentation on the plain text output by S1 to output standardized segmented text: (1) Data cleaning: Remove headers and footers, watermarks, garbled text, meaningless special symbols, continuous blank lines, and duplicate or redundant text; (2) Standardize format: unify Chinese and English punctuation, and convert dates to the same format. Standard format, unified half-width characters for numbers; (3) Resume module segmentation: The text is divided into independent text blocks based on the page metadata and title keywords. The module types are personal basic information, education background, work experience, project experience, professional skills, self-evaluation, and honors and certificates. (4) Long text fragmentation: If the length of a single text character exceeds the threshold of the LLM context window, a sliding window within the module is used for fragmentation to ensure that a single education / work / project record is completely retained within a single window, thus avoiding field fragmentation.

[0024] S3. Construction of semantically enhanced prompt words: Based on a pre-defined extracted field schema and standardized segmented text, structured semantic enhancement prompts are assembled. These prompts are composed of seven fixed components arranged in an orderly manner, used to constrain LLM inference logic and unify output specifications. (1) Role definition section: The model's identity is limited to a professional resume information extraction expert; (2) Task description section: Clearly inform the model of all resume feature fields to be extracted and the extraction target; (3) Field Schema Definition Section: Use JSON templates to annotate the field name, semantic meaning, data type, value constraints, required field identifier, and standard example value for each field; (4) Semantic Enhancement Sample Segment: Input multiple sets of irregular original resume texts and corresponding standardized extraction result samples, covering synonymous expressions, time omissions, field misalignments, abbreviated dates, and abnormal sentence structures with mixed Chinese and English; (5) Output format constraint section: Forces the output to return only standard JSON objects, and specifies the rules for filling arrays, nested sub-objects, and null values; (6) Extract constraint rule segment: Define field boundary judgment, ambiguous information rejection, multiple instance merging, conflict information priority adjudication rules; (7) Local context embedding segment: embed prompt words into the module title and the metadata of the adjacent paragraphs corresponding to the current text block to be extracted, to help the model distinguish the module to which the field belongs.

[0025] S4, LLM staged feature extraction: The structured semantic enhancement prompts constructed by S3 are concatenated with the standardized segmented text preprocessed by S2. A three-level hierarchical extraction is then performed using the large language model API, reducing the complexity of a single inference iteration and improving the completeness of multi-module related feature extraction. (1) Phase 1: Resume module recognition: Input complete preprocessed text, drive LLM to recognize all resume modules in the text, output the start and end character range of the text corresponding to each module, and divide into independent extraction units; (2) Phase 2: Precise extraction of fields by module: For each independent module unit divided in Phase 1, LLM is called separately in combination with semantic enhancement prompts to extract all feature fields in the module; For the work experience, project experience, education and other item structures, the array nested sub-object format is forced to be output. (3) Phase 3: Cross-module feature association and fusion: Identify similar features scattered across multiple modules, perform cross-module merging, duplicate feature deduplication, and synonymous skill normalization, and summarize them into a unified skill field array.

[0026] S5. Structured Output Generation: It receives the raw string returned by LLM, validates and parses the JSON structure, maps it to a predefined unified data model, and generates a standardized JSON structured resume dataset. The standard output structure contains four core objects: basic personal information (basic_info), education (education), work experience (work_experience), project experience (projects), and one-dimensional skills (skills). All time, education, company, and skill fields are stored in a unified standardized format.

[0027] S6. Post-processing for result verification and completion: Perform multi-dimensional validation, conflict resolution, and missing field completion on the standardized JSON structured resume dataset output by S5 to eliminate LLM illusions, format errors, and logical contradictions. (1) Format validation: Validate the integrity of the JSON, the data type matching of each field, and whether the required fields are empty; (2) Business logic validation: Timeline logic validation (education end time is earlier than the start time of the first work segment, and there is no overlap in the time intervals of each segment), enumeration field value validation (education level values ​​are doctoral / master / bachelor / associate / high school), and contact information regular expression validation (mobile phone number and email format compliance). (3) External knowledge base unification verification: Connect to the pre-set standard knowledge base, unify the abbreviations / aliases of schools in the school database, unify the full names of companies in the industrial and commercial enterprise database, and complete the conversion of abbreviations to standard terms in the skills dictionary; (4) Intelligent completion of missing fields: For key fields with confidence scores below the preset threshold or missing verification, LLM is called again to extract them with the existing structured data. (5) Conflict information resolution: When there are conflicting values ​​in multiple modules for the same field, the information with higher credibility is adopted according to the priority of independent skills column > project description > work experience > self-evaluation.

[0028] S7. Quality assessment and feedback optimization closed loop: Establish a full-sample extraction quality scoring mechanism, determine the threshold for manual review based on field confidence and overall document confidence, and collect low-quality samples to form a continuously iterative dataset: (1) Field confidence calculation: ; In the formula: For single-field confidence, the value range is \([0,1]\). The original probability scores for the LLM output fields; The checksum is set to 1 if all checks pass, and 0 if there are logical / format errors. The score is determined by the knowledge base normalization matching score: 1 for a perfect match of the standard library, 0.5 for a partial match of abbreviation, and 0 for no match. These are the weighting coefficients, and the sum of the weighting coefficients equals 1.

[0029] (2) Calculation of overall document confidence: ; In the formula: The overall confidence level of a single resume document; This represents the total number of all feature fields to be extracted from this resume. For the first Confidence level of each field.

[0030] (3) Feedback Iteration Process: Set the threshold for manual review ,like Below The samples are then pushed to a manual annotation platform for correction. The complete samples after manual correction are stored in the iterative sample library to expand the semantic enhancement prompt word sample library, iteratively optimize the prompt word template, update the knowledge base matching words, and serve as training material for LLM fine-tuning / prompt word optimization, thereby continuously improving the subsequent extraction accuracy.

[0031] One specific implementation method is as follows: This embodiment uses PDF resumes uploaded by job seekers on enterprise recruitment platforms as the processing object, and fully executes all steps of the present invention to completely reproduce the entire process logic: S1. Resume Source File Parsing: Receives PD resumes uploaded by job seekers, calls the PDF parsing engine to extract full text, paragraph headings, and table information, and records layout metadata: paragraphs 1-3 are personal information modules, paragraphs 4-6 are educational background, paragraphs 7-15 are work experience, paragraphs 16-22 are project experience, and paragraph 23 is professional skills; outputs complete text with layout marks.

[0032] S2. Text Preprocessing: Clean and remove company advertisements and consecutive spaces from PDF headers; unify "2020 / 3", "2021.05", and "to date" to... Standard format; divided into six independent text blocks according to page metadata; the character length of each text block does not exceed the LLM context window threshold, and no sliding segmentation is required.

[0033] S3. Semantic Enhancement Prompt Construction: Seven types of components are concatenated sequentially, and the field schema defines twenty-two extracted fields such as education, company, position, skills, and project name; the semantic enhancement sample input includes fifteen sets of irregular resume texts such as "worked at XX company for three years" and "served as a backend developer from 2018 to present" and standardized extraction results; the output is forced to a nested JSON structure, embedding the module title context corresponding to each text block.

[0034] S4, LLM staged feature extraction: Phase 1: Input complete preprocessed text, LLM outputs six types of resume modules and corresponding text character ranges; Phase Two: LLM extraction is called separately for each module, and work experience and project experience are output as array sub-objects. Two jobs and three projects are extracted from a single resume. Phase 3: Cross-module integration, identifying duplicate Redis and MySQL entries in the project description and skills list, deduplicating them, and summarizing the skills array.

[0035] S5. Structured Output Generation: Parse the JSON string returned by the LLM, map it to the standard data model, and generate a structured dataset containing name, phone number, school, two work experiences, three projects, and eight standardized skills.

[0036] S6. Result Validation and Completion: Format validation: The JSON has no missing fields, and all data types match. Logical validation: Education period from September 2016 to June 2020, first job started in July 2020, no time overlap, education field is compliant as undergraduate; regular expression validation of mobile phone number and email passed; Knowledge base unification: The matching of "PKU" in the resume to the university database has been corrected to "Peking University", and "js" has been unified to "JavaScript"; There are no missing fields or conflicting information, so no secondary completion extraction is required.

[0037] S7. Quality Assessment and Feedback Optimization: Substitute each field into the field confidence formula to calculate The sum of the weight coefficients equals 1, all fields pass validation and are a perfect match to the knowledge base, and the confidence level of each individual field is 1; the total number of fields in the document. Substitute into the overall document confidence formula to calculate , Higher than the threshold for manual review No manual annotation is required; samples do not need to be returned to the iteration library and can be directly stored in the standardized talent database for job matching.

[0038] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent extraction of resume feature words based on semantic enhancement, characterized in that, Includes the following steps: Step S1: Resume Source File Parsing: Receive the resume source file, identify the file format, and use the corresponding parsing engine to convert it into plain text carrying layout metadata; Step S2: Text preprocessing: Clean the plain text, normalize the format, divide the resume module into blocks, and process excessively long text segments to output standardized segmented text; Step S3: Semantic Enhancement Prompt Construction: Based on the preset extracted field schema and standardized segmented text, assemble structured semantic enhancement prompts containing seven fixed components; Step S4: LLM phased feature extraction: After concatenating the structured semantic enhancement prompts with the standardized segmented text, the large language model is called to perform three levels of extraction in sequence: module recognition, sub-module field extraction, and cross-module feature association and fusion. Step S5: Structured Output Generation: Parse the output of the large language model, map it to a predefined unified data model, and generate a standardized JSON structured resume dataset; Step S6: Post-processing for result verification and completion: Perform format verification, business logic verification, external knowledge base normalization verification, missing field completion, and conflict information resolution on the standardized JSON structured resume dataset; Step S7: Quality assessment and feedback optimization: The quality of extraction is quantified by the confidence calculation formula. Samples that are lower than the threshold for manual review are manually corrected and then fed back to iteratively optimize the prompt words and knowledge base, forming a self-optimizing closed loop.

2. The method for intelligent extraction of resume feature words based on semantic enhancement according to claim 1, characterized in that: In step S1, the resume source file formats include PDF, Word, TXT, HTML, and image; for image format resumes, OCR optical character recognition is performed to convert them into text; the layout metadata includes paragraph number, module title, list level, and text block coordinates. The layout metadata is used to assist in the module segmentation in step S2 and the context embedding in step S3.

3. The method for intelligent extraction of resume feature words based on semantic enhancement according to claim 1, characterized in that: In step S2, the text preprocessing execution flow is as follows: remove garbled characters, headers and footers, and redundant blank symbols; unify Chinese and English punctuation, date, and number formats; divide personal basic information, education background, work experience, project experience, professional skills, and self-evaluation modules according to the page metadata; when the text length exceeds the threshold of the large language model context window, use a sliding window within the module to segment the text, and retain a single education, work, or project record completely in a single window.

4. The method for intelligent extraction of resume feature words based on semantic enhancement according to claim 1, characterized in that: In step S3, the seven fixed components are, in order: role definition section, task description section, field schema definition section, semantic enhancement sample section, output format constraint section, extraction constraint rule section, and local context embedding section. The field schema definition section uses a JSON template to annotate field names, semantic interpretations, data types, value constraints, required field identifiers, and standard examples. The semantic enhancement sample section stores synonym expressions, time omissions, field misalignments, and irregular resume extraction samples with abbreviated date types.

5. The method for intelligent extraction of resume feature words based on semantic enhancement according to claim 1, characterized in that: In step S4, the three-level extraction execution process is as follows: Phase 1: Module recognition, identifying all modules of the resume and their corresponding text character ranges; Phase Two: Module-based field extraction. Feature fields are extracted independently according to the divided module units, and several item structures are output as nested arrays of sub-objects. Phase 3: Cross-module feature association and fusion, locating similar features within multiple modules, and performing deduplication, synonym unification, and summary merging operations.

6. The method for intelligent extraction of resume feature words based on semantic enhancement according to claim 1, characterized in that: In step S5, the standardized JSON structured resume dataset contains five core objects: basic personal information, education experience array, work experience array, project experience array, and one-dimensional skills array. The time, education, company, and skills fields are all stored in a unified and standardized manner.

7. The method for intelligent extraction of resume feature words based on semantic enhancement according to claim 1, characterized in that: In step S6, three independent knowledge bases are configured: a standard college database, a business database, and a skills dictionary. The knowledge base unification verification performs the following checks: unification of college abbreviations, unification of enterprise aliases, and conversion of skills abbreviations and terms. The business logic verification includes timeline consistency verification, enumeration field value verification, and contact information regular expression format verification.

8. The method for intelligent extraction of resume feature words based on semantic enhancement according to claim 1, characterized in that: In step S6, conflict information resolution adopts a priority rule: when multiple modules have conflicting values ​​for the same field, the value from the independent skills column is adopted, followed by the value from the project description, then the value from the work experience, and finally the value from the self-evaluation; the missing field completion process involves carrying existing structured data and calling the large language model for targeted extraction.

9. The method for intelligent extraction of resume feature words based on semantic enhancement according to claim 1, characterized in that: In step S7, two sets of confidence calculation formulas are configured: a single-field confidence calculation formula and a document overall confidence calculation formula. Formula for calculating single-field confidence: ； In the formula: For single-field confidence, the range of values ​​is... ; Output the original probability scores of the fields for the large language model; This is a verification pass indicator; The score is determined by the normalized matching of the knowledge base. These are the weighting coefficients, and the sum of the weighting coefficients equals 1. Formula for calculating overall document confidence: ； In the formula: The overall confidence level of a single resume document; This represents the total number of feature fields to be extracted from the resume. For the first Confidence levels for each field; In step S7, configure the manual review threshold. , The value is less than Samples are pushed to the manual annotation platform for correction in real time.

10. The method for intelligent extraction of resume feature words based on semantic enhancement according to claim 1, characterized in that: In step S7, the samples after manual correction undergo four types of backflow operations: expanding the semantic enhancement sample library, iteratively optimizing the prompt word template, updating the knowledge base matching words, and using them as training material for fine-tuning the large language model or optimizing the prompt words. The four types of backflow operations are executed simultaneously to optimize the accuracy of resume feature extraction.

Images