Resume analysis optimization method based on two-stage asynchronous processing in ATS system
By employing a two-stage asynchronous processing method, combined with lightweight and multimodal large-scale resume parsing, the dilemma of response speed and parsing quality in the ATS system is resolved, achieving fast and efficient resume parsing and high-quality data output.
Patent Information
- Application Number
- CN202511486867.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-17
- Publication Date
- 2026-02-06
AI Technical Summary
Existing ATS systems face a dilemma in the resume parsing process: balancing response speed and parsing quality. They cannot simultaneously meet the requirements of fast response and high parsing quality.
A two-stage asynchronous processing method is adopted. In the initial parsing stage, a lightweight module is used to quickly extract key fields. In the fine parsing stage, a multimodal large model is called through asynchronous task scheduling to perform in-depth parsing, and the results are merged to form a high-quality structured resume file.
It achieves a fast parsing experience with no delay for users, while improving parsing quality and enhancing the data foundation for subsequent job matching and screening processes. It also boasts high resource utilization and strong compatibility.
Smart Images
Figure CN121480481A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology in automated recruitment systems (ATS), and more particularly to an optimized resume parsing method based on two-stage asynchronous processing in an ATS system. Background Technology
[0002] In the ATS system workflow, resume parsing is a crucial initial step in converting unstructured resume documents into structured data, and its effectiveness directly determines the data quality and usability of subsequent processes. Existing parsing solutions mainly fall into two categories: One approach is rule-based and traditional natural language processing parsing methods, which extract data through techniques such as regular expression matching, semantic segmentation, entity recognition, and template matching. While these methods offer fast response times, they suffer from significant bottlenecks in accuracy, recall, and generalization ability, especially when dealing with resumes with diverse layouts and unique structures, where further improvement in parsing performance becomes difficult.
[0003] The second approach is a multimodal large-scale model-based parsing method. Leveraging powerful cross-modal understanding capabilities, this method converts resume documents into image input and directly outputs structured parsing results. While this method excels in accuracy and completeness, it consumes significant computational resources and suffers from high processing latency. A single parsing session often takes tens of seconds or even more than a minute, failing to meet users' demands for real-time feedback.
[0004] Therefore, existing ATS systems face a dilemma: either sacrifice parsing quality for faster response, or sacrifice user experience for better parsing results. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an optimized resume parsing method based on two-stage asynchronous processing in the ATS system.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: a resume parsing optimization method based on two-stage asynchronous processing in an ATS system, comprising the following steps: S1: Initial parsing stage: A lightweight parsing module based on regular expression rules and entity recognition technologies is used to quickly extract key fields from resumes, including name, gender, contact information, university, education and major information; this module has high real-time performance and can return parsing results within seconds; S2: User Response and Business Implementation: Returns the initial parsing results to the user interface in real time and stores them in the system, supporting users to perform subsequent operations without delay; S3: Fine-grained analysis stage: Through an asynchronous task scheduling mechanism, a multimodal large model is invoked to perform in-depth analysis on the same resume, extracting complete fields including work experience, project details, and skills and expertise; S4: Data Fusion and Update: Merge the refined analysis results into the existing resume data, cover and supplement the initial analysis results, and form the final high-quality structured resume archive.
[0007] Preferably, the user's subsequent operations in S2 include: resume screening based on key fields, initiating interview invitations through contact information, adding business tags to candidates (focusing on / postponing progress), and associating with target recruitment positions.
[0008] Preferably, the asynchronous task scheduling mechanism in S3 is implemented based on the Celery distributed task framework, with the priority ordering rule as follows: newly uploaded resumes have priority > re-parsing of historical resumes, and resumes associated with urgent positions have priority > resumes for ordinary positions; the failure retry mechanism is as follows: automatic retry 3 times when timeout / error occurs, with retry intervals of 5 seconds, 10 seconds, and 30 seconds respectively; the resource rate limiting rule is as follows: the number of large model calls in a single batch during peak periods is ≤20.
[0009] Preferably, the preset fusion rules in S4 include: if a finely parsed field is more complete than the primary parsing, the primary result is overridden; when field values conflict, the finely parsed result takes precedence and a conflict log is recorded; newly added fields in the finely parsed analysis are directly added to the resume file.
[0010] Preferably, a configurable solution for primary parsing fields is also included: allowing enterprises to customize key fields in the ATS system administrator backend, add or delete fields that require primary parsing (such as place of origin, GitHub address), without the need for code development.
[0011] Preferably, a multi-model fallback scheme is also included: if the currently invoked multimodal large model is unavailable, the system automatically switches to the backup large model, records the model switching log, and issues an alarm.
[0012] Preferably, a visualization solution for parsing progress is also included: displaying a detailed parsing progress bar on the resume details page of the ATS system, and synchronizing the status of "Field extraction in progress / Conflict resolution in progress / Parsing completed" in real time.
[0013] Preferably, a batch re-parsing scheme for historical resumes is also included: the system automatically processes historical resumes during idle hours at night (CPU utilization <30%), sorts them in reverse order of parsing time, and gradually updates the quality of existing resume data.
[0014] Preferably, before the initial parsing stage S1, a multi-format resume preprocessing step is also included; for resume documents of different formats (PDF, Word, scanned copies), the system first performs format unification processing: For text resumes (PDF / Word): the text content is extracted using the Apache PDFBox / POI tool, redundant format symbols (such as garbled characters and page breaks) are removed, and a standardized text sequence is generated; For scanned resumes (JPG / PNG): the Tesseract OCR engine is first called to convert the image into text (supporting Chinese and English bilingual recognition), and then text standardization processing is performed; After preprocessing, the uniformly formatted text is then input into the lightweight parsing module to ensure the compatibility of the primary parsing with resumes of various formats and to avoid missing key fields due to format differences.
[0015] Preferably, the lightweight parsing module in S1 is further defined in the following technical details: the entity recognition technology adopts a compressed version of the BiLSTM-CRF model, the model parameter size is <50M (to adapt to the lightweight requirements of the ATS system), the training data covers 100,000+ resume samples from different industries and different formats, and supports entity extraction of key fields such as name, contact information, school, education, and major.
[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: 1. Significantly improves user experience: Users no longer need to wait for a long parsing process, achieving a seamless import process of "submit and complete"; 2. Significantly improved parsing quality: The multimodal large model enables accurate extraction of deep and complex fields, improving the data foundation for subsequent job matching and screening processes; 3. Reasonable resource allocation: Transfer the computing pressure to background asynchronous tasks to avoid blocking user operation threads and improve system resource utilization; 4. Strong compatibility and scalability: This invention does not rely on specific large models or traditional analysis tools, and any module can be flexibly replaced to adapt to technological iteration and development. Attached Figure Description
[0017] Figure 1 This invention presents a flowchart illustrating the steps of a resume parsing optimization method based on two-stage asynchronous processing in an ATS system. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0019] This invention provides a technical solution: a resume parsing optimization method based on two-stage asynchronous processing in an ATS system, comprising the following steps: S1: Initial parsing stage: A lightweight parsing module based on regular expression rules and entity recognition technologies is used to quickly extract key fields from resumes, including name, gender, contact information, university, education and major information; this module has high real-time performance and can return parsing results within seconds; S2: User Response and Business Implementation: Returns the initial parsing results to the user interface in real time and stores them in the system, supporting users to perform subsequent operations without delay; S3: Fine-grained analysis stage: Through an asynchronous task scheduling mechanism, a multimodal large model is invoked to perform in-depth analysis on the same resume, extracting complete fields including work experience, project details, and skills and expertise; S4: Data Fusion and Update: Merge the refined analysis results into the existing resume data, cover and supplement the initial analysis results, and form the final high-quality structured resume archive.
[0020] Subsequent user actions in S2 include: resume screening based on key fields, initiating interview invitations via contact information, adding business tags to candidates (focusing on / postponing progress), and associating them with target job postings; The asynchronous task scheduling mechanism in S3 is implemented based on the Celery distributed task framework. The priority ordering rule is: newly uploaded resumes > re-parsed historical resumes, and resumes associated with urgent positions > resumes for ordinary positions. The failure retry mechanism is: automatic retry 3 times in case of timeout / error, with retry intervals of 5 seconds, 10 seconds, and 30 seconds respectively. The resource rate limiting rule is: the number of large model calls in a single batch during peak hours is ≤20. The default fusion rules in S4 include: if a finely parsed field is more complete than the primary parsing, the primary result will be overridden; when field values conflict, the finely parsed result will prevail and a conflict log will be recorded; newly added fields in the finely parsed analysis will be directly added to the resume file. It also includes a configurable solution for primary parsing fields: allowing enterprises to customize key fields in the ATS system administrator backend, add or delete fields that require primary parsing (such as place of origin, GitHub address), without the need for code development; It also includes a multi-model fallback scheme: if the currently invoked multimodal large model is unavailable, the system automatically switches to the backup large model, records the model switching log, and issues an alarm; It also includes a visualization solution for parsing progress: displaying a detailed parsing progress bar on the resume details page of the ATS system, and synchronizing the "Field extraction in progress / Conflict resolution in progress / Parsing completed" status in real time; It also includes a batch re-parsing solution for historical resumes: the system automatically processes historical resumes during idle nighttime hours (CPU utilization <30%), sorts them in reverse order of parsing time, and gradually updates the quality of existing resume data.
[0021] The complete process logic of this invention is as follows: Input trigger: HR uploads resumes (single / batch) to the ATS system, and the system receives resume files (PDF / Word / scanned copy); Basic parsing: Calls a lightweight parsing module to extract key fields within 1 second, returns them to the front end, and stores them in a temporary table; Asynchronous scheduling: The system submits fine-grained parsing tasks to the Celery task queue, sorting them by priority; Fine-grained parsing: A background thread calls a multimodal large model to deeply extract complex fields, synchronizing progress during processing; Data fusion: Merges the results of two levels of parsing according to preset rules, resolves conflicts, and completes information; Result output: Updates the official resume form, notifies HR that parsing is complete, and supports subsequent job matching and screening operations.
[0022] Before the S1 initial parsing stage, there are also preprocessing steps for resumes in various formats. For resume documents of different formats (PDF, Word, scanned copies), the system first performs format unification processing: For text resumes (PDF / Word): the Apache PDFBox / POI tool is used to extract the text content, remove redundant formatting symbols (such as garbled characters and page breaks), and generate a standardized text sequence; For scanned resumes (JPG / PNG): the Tesseract OCR engine is first called to convert the image into text (supporting Chinese and English bilingual recognition), and then text standardization processing is performed; After preprocessing, the uniformly formatted text is then input into the lightweight parsing module to ensure the compatibility of the primary parsing with resumes of various formats and to avoid missing key fields due to format differences.
[0023] The lightweight parsing module in S1 is further defined in terms of technical details: the entity recognition technology adopts a compressed version of the BiLSTM-CRF model, the model parameter size is <50M (adapting to the lightweight requirements of the ATS system), the training data covers 100,000+ resume samples from different industries and different formats, and supports entity extraction of key fields such as name, contact information, school, education, and major.
[0024] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A resume parsing optimization method based on two-stage asynchronous processing in an ATS system, comprising the following steps: S1: Primary parsing stage: a lightweight parsing module based on regular rules and entity recognition is used to quickly extract key fields in the resume, including name, gender, contact information, graduation school, education and professional information; this module has high real-time performance and can return parsing results within seconds; S2: User response and business promotion: the primary parsing results are returned to the user interface in real time and stored in the system, supporting users to perform subsequent operations without delay; S3: Fine parsing stage: through an asynchronous task scheduling mechanism, a multi-modal large model is called to deeply parse the same resume, extracting complete fields including work experience, project details, and skill specialties; S4: Data fusion and update: merge the fine parsing results into the existing resume data, covering and supplementing the primary parsing results, and forming the final high-quality structured resume file.
2. The resume parsing optimization method based on two-stage asynchronous processing in the ATS system according to claim 1, characterized in that: The user's subsequent operations in S2 include: resume screening based on key fields, initiating an interview invitation through contact information, adding business tags (focus / temporarily promote) to candidates, and associating target recruitment positions. 3.The resume parsing optimization method based on two-stage asynchronous processing in an ATS system according to claim 1, characterized in that: The asynchronous task scheduling mechanism in S3 is implemented based on the Celery distributed task framework, with priority sorting rules as follows: new resume priority > historical resume re-parsing, urgent position associated resume priority > ordinary position resume; the failure retry mechanism is: automatically retry 3 times when timeout / error, with retry intervals of 5 seconds, 10 seconds, and 30 seconds; the resource throttling rule is: the number of large model calls per batch ≤20 during peak hours.
4. The resume parsing optimization method based on two-stage asynchronous processing in an ATS system according to claim 1, characterized in that: The preset fusion rules in S4 include: if the fine parsing fields are more complete than the primary parsing, the primary results are overwritten; in case of field value conflict, the fine parsing result is used as the standard, and conflict logs are recorded; fine parsing new fields are directly supplemented to the resume file.
5. The resume parsing optimization method based on two-stage asynchronous processing in an ATS system according to claim 1, characterized in that: It also includes a primary parsing field configurable scheme: supports enterprises to customize key fields in the ATS system administrator background, add or delete fields that need primary parsing (such as hometown, GitHub address), without code development.
6. The resume parsing optimization method based on two-stage asynchronous processing in an ATS system according to claim 1, characterized in that: It also includes a multi-model fallback scheme: if the currently called multi-modal large model is unavailable, the system automatically switches to the backup large model, records model switching logs and alarms.
7. The resume parsing optimization method based on two-stage asynchronous processing in an ATS system according to claim 1, characterized in that: It also includes a parsing progress visualization scheme: a fine parsing progress bar is displayed on the ATS system resume detail page, with real-time synchronization of "field extraction in progress / conflict processing in progress / parsing completed" status.
8. The resume parsing optimization method based on two-stage asynchronous processing in an ATS system according to claim 1, characterized in that: It also includes a historical resume batch re-parsing scheme: the system automatically processes historical resumes during the night idle period (CPU utilization <30%), sorted in reverse order of parsing time, gradually updating the quality of inventory resume data.
9. The resume parsing optimization method based on two-stage asynchronous processing in an ATS system according to claim 1, characterized in that: Before the S1 primary analysis stage, a multi-format resume preprocessing step is included; for different formats of resume documents (PDF, Word, scan), the system first performs format uniform processing: for text type resume (PDF / Word): extract the text content through Apache PDFBox / POI tools, remove redundant format symbols (such as garbled code, page break), and generate a standardized text sequence; for scanned resume (JPG / PNG): first call the Tesseract OCR engine to convert the image to text (supporting bilingual recognition of Chinese and English), and then perform text standardization processing; After preprocessing, the uniformly formatted text is input into the lightweight parsing module to ensure the compatibility of the primary parsing for multi-format resumes and avoid missing key fields due to format differences.
10. The resume parsing optimization method based on two-stage asynchronous processing in an ATS system according to claim 1, characterized in that: The lightweight parsing module in S1 is further limited in technical details: the entity recognition technology uses a compressed BiLSTM-CRF model, the model parameter size is <50M (adapted to the lightweight requirements of the ATS system), the training data covers 100,000+ resume samples from different industries and different formats, and supports entity extraction of key fields such as name, contact information, college, education, major, etc.