Resume analysis method and system based on dynamic semantic network

Abstract

The invention discloses a resume analysis method and system based on a dynamic semantic network, and relates to the technical field of data information processing. The method comprises inputting a resume; according to the file type of the resume, multi-mode resume analysis engine processing is carried out, and information extraction is carried out; constructing a dynamic semantic network according to the extracted information; and outputting structured data based on the stored knowledge graph in combination with an industry self-adaptive mechanism to complete resume analysis. The resume analysis system is used for realizing the resume analysis method. According to the resume analysis method and system, the effects of high resume analysis accuracy, high processing efficiency and high industry adaptability can be achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of data information processing, and in particular to a resume parsing method and system based on a dynamic semantic network. Background Art

[0002] In the current talent recruitment market, the number of job seekers is huge. Companies and recruitment agencies are faced with a large amount of resume screening during recruitment. The accuracy, processing capacity, and screening cost of resume parsing are particularly important. However, existing resume parsing technology has at least the following shortcomings: 1. Insufficient multimodal processing capabilities: The parsing accuracy of resumes in formats such as PDF, Word, and images varies significantly. The character recognition error rate of traditional OCR technology for scanned documents is as high as 15%.

[0003] 2. Weak unstructured information extraction: The accuracy rate of semantic association recognition for complex descriptions such as technology stacks and work experience (such as "led the design of SpringBoot microservice architecture") is less than 70%, and it is unable to capture implicit knowledge.

[0004] 3. High industry adaptation costs: Proprietary terms in fields such as finance and healthcare (such as "anti-money laundering" and "MRI image analysis") require manual labeling of training data, and the model iteration cycle can take up to several weeks.

[0005] For example, the Chinese invention patent with publication number CN119377396A discloses a career matching and recommendation method based on natural language processing, and its technical solution is: collecting job applicant resume data and job description data; performing multi-level feature extraction on the pre-processed job applicant resume data and job description data through the BERT model; enhancing the multi-level features using weakly supervised learning technology, and deeply modeling the latent semantics of the text through a self-supervised adversarial network; constructing a semantic matching network; calculating the matching degree between the job applicant resume data and the job description data through the semantic matching network and feature set; generating and providing a corresponding job recommendation list; monitoring the status changes of job applicant resume data and job market dynamics in real time; collecting feedback information from job seekers and recruiters; and ensuring that the recommended positions are highly consistent with the career development needs of job seekers. For example, Chinese invention patent publication number CN119850159A discloses a Transformer-based person-job matching method, system, device, and medium. The technical solutions employed include: collecting job information data and job applicant resume data from multiple data sources, cleaning and organizing the collected data to obtain job descriptions and resumes; preprocessing the collected and organized job descriptions and resumes to obtain preprocessed job descriptions and resumes; integrating the Transformer with the BERT model to optimize the model; constructing a multi-label classification system; and outputting and evaluating matching results. However, the technical solutions disclosed in these two patents still suffer from the aforementioned shortcomings in resume parsing.

[0006] Therefore, it is necessary to design a resume parsing method and system based on dynamic semantic networks that can solve the problems of existing resume parsing technologies, such as insufficient multimodal processing capabilities, weak unstructured information extraction, and high industry adaptation costs. Summary of the Invention

[0007] In view of the problems existing in the prior art, the present invention provides a resume parsing method and system based on a dynamic semantic network.

[0008] To achieve the above object, the technical solution adopted by the present invention is as follows: A resume parsing method based on a dynamic semantic network includes the following steps: S1, inputting a resume; S2, extracting information according to the file type of the resume, and performing multimodal resume parsing engine optimization processing; S3, constructing a dynamic semantic network based on the extracted information; wherein the construction process is: S31, constructing an entity relationship model; S32, fusing a knowledge graph based on the entity relationship model; S33, completing the construction of the dynamic semantic network and storing the knowledge graph; S4, based on the stored knowledge graph, combined with an industry adaptive mechanism, outputting structured data to complete the resume parsing.

[0009] Based on the above technical solution, further, the logical process of S2 is: S21, preprocess the resume file and output plain text data in a unified format; S22, extract information based on the output plain text data; wherein, the logical process of information extraction is: S221, first input the plain text data output after preprocessing; S222, extract basic information, educational information, work experience and technology stack information based on the plain text data; S223, output structured information based on the extracted information.

[0010] Based on the above technical solution, further, in S222, (1) basic information: based on the BERT-BiLSTM-CRF model, the name and contact information are identified, and a structured entity list is outputted; (2) educational information: regular expression matching is used, combined with dependency syntax analysis to extract GPA and degree, supporting multiple languages, and a list of educational entries is outputted; (3) work experience: the LSTM-CRF model is used to identify the company name, position, and time range; and the attention mechanism is used to extract key events in the job description, and a list of experiences with a timeline is outputted; (4) technology stack: a term library matching and BERT semantic verification method are used to construct a set of technology points, and a list of technology stack labels is outputted.

[0011] Based on the above technical solution, further, the logical process of S2 also includes: S23, performing technical stack parsing optimization operations based on the extracted information; among them, the parsing optimization operation process of S23 includes: 1) using a hybrid expert architecture for distributed computing architecture: (1) Expert model: assigning independent BERT models to different industries and dynamically scheduling computing resources; (2) Load balancing: automatically expanding expert model instances according to resume processing volume, supporting the processing of 20 resumes per second; 2) Caching and preloading operations: (1) Glossary cache: caching high-frequency terms in memory, increasing query speed by 3 times; (2) Model preloading: loading the BERT model when the system starts, with a cold start time of <5 seconds.

[0012] Based on the above technical solution, further, the logic of S31 is: first extract the semantic relationship between entities in the resume text through dependency syntax analysis, then set dynamic weights for each type of relationship according to industry characteristics, and then complete the construction of the entity relationship model.

[0013] Based on the above technical solution, the fusion process of S32 is further as follows: first associate with the external knowledge base, supplement entity attributes, and perform multi-language alignment; then support visual query; and finally output the fused rich knowledge triples, which are the entity, relationship, and attribute data in the knowledge graph.

[0014] Based on the above technical solution, further, in S33, the process of constructing a dynamic semantic network architecture is: S331, perform entity standardization processing; S332, combine the results of S331 processing to perform attribute enhancement processing; S333, combine the results of S332 processing to generate a unique identifier; S334, perform hierarchical classification processing on the generated unique identifier.

[0015] Based on the above technical solution, further, the logical process of S4 is: input the entity, relationship, and attribute data in the knowledge graph; convert the graph data into a standard JSON format, integrate basic information, education, work experience, technology stack, and semantic network relationships; and finally output structured data.

[0016] Based on the above technical solution, an industry adaptive mechanism is further applied to the entire process of steps S2-S3, wherein the industry adaptive mechanism includes at least two learning mechanisms: transfer learning and active learning.

[0017] A resume parsing system based on a dynamic semantic network is used to implement the resume parsing method based on a dynamic semantic network. The resume parsing system includes a data input module, a preprocessing module, an information extraction module, a semantic network construction module, a knowledge graph system module, an industry adaptation module, a structured output module and a performance optimization module; the data input module is used to execute S1, perform resume input operations, and transmit data to the preprocessing module; the preprocessing module is used to perform preprocessing operations on the input resume file, execute step S21; and transmit data communication to the preprocessing module; the information extraction module is used to extract information, execute step S22, and transmit the processed data to the semantic network construction module; the industry adaptation module is used to implement two learning mechanisms, and its processed data is transmitted to the information extraction module and the semantic network construction module; the performance optimization module is mainly used to execute the technology stack parsing performance optimization process, execute step S23, and its processed data is also transmitted to the information extraction module; the semantic network construction module is used to execute step S3, and its processed data is transmitted to the knowledge graph storage module; the structured output module is used to execute step S4.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention uses a dynamic semantic network architecture for resumes, specifically constructing a knowledge graph of "entity-relationship-attribute" triples, to achieve deep semantic understanding of resume information.

[0019] (2) This invention uses a multimodal collaborative parsing engine, integrating OCR, NLP, and computer vision technologies, to support intelligent parsing of resumes in all formats, effectively improving parsing accuracy and processing efficiency. Specifically, the accuracy of key information extraction can reach 98.5%, and the F1 value of technology stack recognition can be improved to 92%. The processing time for a single resume is less than 3 seconds, and it can support the concurrent parsing of tens of thousands of resumes.

[0020] (3) This invention implements an industry-adaptive mechanism, using transfer learning and active learning techniques to achieve rapid adaptation of domain knowledge. Through transfer learning, the matching accuracy in fields such as finance and healthcare can be increased by 15%-20%.

[0021] (4) The present invention solves domain-specific problems (such as qualification identification of financial resumes) by adopting an industry adaptation module; it also ensures processing efficiency in high-concurrency scenarios (such as parsing 20 resumes per second) through a performance optimization module, achieving dual optimization of industry and performance.

[0022] (5) The knowledge graph storage module provided by the present invention supports cross-industry data association analysis, laying the foundation for high-level talent analysis (such as skill trend prediction). In addition, the modular architecture provided supports independent expansion (for example, the addition of a "certificate recognition" module only requires docking with the information extraction module), which is scalable.

[0023] (6) The present invention adopts a hybrid matching strategy, performing precise matching, fuzzy matching, context-aware matching, and synonym expansion in sequence. This sequence, through the progressive logic of "from certainty to fuzziness, from form to semantics", maximizes the comprehensiveness and accuracy of technology stack recognition while ensuring efficiency, meeting the dual needs of "quick positioning + deep understanding" in resume parsing scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flowchart of the overall process of the resume parsing method of the present invention; Figure 2 This is a flowchart of the technology stack analysis in the resume parsing method of the present invention; Figure 3 Flowchart of the gap analysis 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] Combine Figure 1 and Figure 2 As shown, this embodiment provides a resume parsing method based on a dynamic semantic network, comprising the following steps: S1. Input resume. In this embodiment, it can receive resumes of various file types / formats, such as PDF / Word / image, etc., to provide raw data for subsequent steps; and finally output the raw resume file of unparsed binary or text data. S2. Extract information based on the resume's file type and optimize the multimodal resume parsing engine. In this step, the transfer learning module in the industry adaptive mechanism has pre-tuned the BERT model using the domain corpus, and the active learning module loads the latest HR annotated data to provide optimization support for the information extraction process. For example, it uses transfer learning to fine-tune the BERT model using the financial domain corpus to improve the recognition accuracy of terms such as "anti-money laundering."

[0028] In this embodiment, the logical process of S2 is: S21. Preprocess the resume file and output plain text data in a unified format; In this embodiment, the preprocessing process specifically processes the original resume file output in step S1. Specifically, for resume files in PDF / Word format, PyPDF2 / docx is used to extract text content and remove noise (such as headers and footers). For resume files in image format, Tesseract + OpenCV is used to perform OCR recognition and image enhancement, converting them into parsable text. The preprocessed resume files are output as plain text data in a unified format (UTF-8 encoding, no formatting interference).

[0029] S22. extracting information based on the output plain text data; In this embodiment, the logical process of information extraction is: S221, first input the plain text data output after preprocessing; S222. Extract basic information, educational information, work experience, and technology stack information based on plain text data; Specifically, (1) Basic information: Based on the BERT-BiLSTM-CRF model, we identify entities such as names and contact information, and finally output a structured entity list. The accuracy has been verified to be 99.2%; (2) Educational information: Regular expressions are used to match the pattern "XX University XX Major", and dependency parsing is combined to extract GPA, degree, etc., supporting multiple languages, such as "Master of Computer Science", and finally outputting a list of educational entries.

[0030] (3) Work experience: The LSTM-CRF model is used to identify company names, positions, and time frames. The attention mechanism is used to extract key events in the job description, such as "led a team of 10 people to complete project delivery," and finally outputs a list of experiences with a timeline.

[0031] (4) Technology stack: Using terminology matching and BERT semantic verification, we build a set of technology points and finally output a list of technology stack labels. For example, the Internet field includes "React" and "Kubernetes". We combine BERT semantic verification to identify implicit associations between technology points, such as the relationship between "microservices" and "Spring Cloud".

[0032] S223. Output structured information based on the extracted information; specifically, output key-value pair data in JSON format based on the four modules of basic information, educational information, work experience, and technology stack.

[0033] S23. Perform technology stack parsing and optimization based on the extracted information. It should be noted that this step is optional and can be added to the process based on actual circumstances. This can improve parsing speed and accuracy and facilitate subsequent semantic network construction in S3 (e.g., more precise extraction of the "technology stack - work experience" relationship). The extracted information here specifically refers to the output technology stack module data.

[0034] In this embodiment, the parsing and optimization process of the technology stack is initiated, and the input data is the technology stack label list data, that is, the technology stack module data output in step S222; and the specific parsing and optimization operation processing process includes: 1. Adopting Hybrid Experts (MoE) architecture for distributed computing architecture: (1) Expert model: Assign independent BERT models to different industries (such as the Internet and finance) and dynamically schedule computing resources.

[0035] (2) Load balancing: Automatically expand expert model instances based on resume processing volume, supporting processing of 20 resumes per second.

[0036] 2. Cache and preloading operations: (1) Term library cache: Cache high-frequency terms (such as "Java" and "Spring Boot") in memory, increasing query speed by 3 times.

[0037] (2) Model preloading: BERT models for common industries are loaded when the system starts, with a cold start time of <5 seconds.

[0038] It should be noted that the analysis and optimization process of the technology stack ultimately outputs an optimized technology stack label with industry-related attributes, such as "Java-Internet Development".

[0039] It should be further explained that the goal of technology stack analysis is to accurately identify technology stack keywords from the resume text and understand their contextual associations (such as the relationship between "microservices" and "Spring Cloud").

[0040] The implementation methods of technology stack analysis include: (1) Industry terminology matching: identifying technical points through predefined industry terminology (such as "React" and "Kubernetes" in the Internet field). (2) BERT semantic verification: using pre-trained models to verify the rationality of technical points in the context (such as determining whether "blockchain" is relevant to the "financial transaction" scenario).

[0041] It should be further explained that the goal of optimizing the technology stack performance is to improve the speed and stability of technology stack parsing and support high-concurrency scenarios (such as processing 20 resumes per second).

[0042] The implementation methods for technology stack performance optimization include: (1) Mixed-of-Experts (MoE) architecture: assigning independent BERT models to different industries to avoid wasting resources when general models process industry-specific terms. (2) Caching and preloading: caching high-frequency terms in memory to reduce the time spent on repeated queries; preloading commonly used industry models to shorten service startup time.

[0043] In general, "technical stack parsing" is the core function of step S222, and "performance optimization" is an extended optimization of step S223. Together, the two constitute the complete chain of information extraction from the technical stack. Through the combination of "precise parsing + performance optimization," steps S222 and S223 not only improve the accuracy of technical stack identification (F1 value 92%), but also meet the efficiency requirements of engineering implementation (high concurrency processing). In other words, "technical stack parsing and performance optimization" is a specialized processing flow for the technical stack module in step S2 (information extraction). It is both the implementation link for the core function and the key optimization point for improving the overall performance of the system, connecting document parsing and semantic network construction.

[0044] S3. Construct a dynamic semantic network based on the extracted information. In this step, the relationship weights and knowledge fusion strategies of the semantic network can be dynamically adjusted based on the domain-optimized model obtained through transfer learning in S2 and the terminology library updated through active learning. For example, the relationship weight of "education-certificate" can be dynamically adjusted based on the HR annotation data fed back by active learning.

[0045] In this embodiment, the construction process is: S31. Build entity relationship model; In this embodiment, the model is constructed based on the structured information in step S2. The construction process is as follows: first, input the name (entity) in the basic information; the school and major (entity) in the academic information; the company, position, and responsibility keywords (entity + event) in the work experience; and the technology stack label (entity). Stanford NLP is then used to analyze text dependencies and extract inter-entity relationships, such as "Zhang San - graduated from - Tsinghua University" and "Python - used for - data mining projects," which include intra-module relationships and inter-module cross-relationships. Relationship weights are dynamically assigned based on industry type (e.g., technical position / finance position) (e.g., the weight associated with technology stack and work experience is increased to 60%). Finally, a weighted set of entity-relationship triples, such as (subject, predicate, object, weight), is output, completing the model construction. This set of entity-relationship triples constitutes the constructed entity relationship model.

[0046] Furthermore, the logic of step S31 is as follows: first, semantic relationships between entities in the resume text are extracted through dependency parsing (involving cross-correlations between the four major information modules), then dynamic weights are assigned to each type of relationship based on industry characteristics, completing the construction of the entity relationship model. It should be noted that the essence of relationship extraction is semantic analysis, not direct module associations; relationship types are not limited to cross-module interactions but also include relationships within modules; and weights are assigned based on industry characteristics, not just arbitrary "per-relationship pair."

[0047] It should be further explained that in the resume parsing scenario, in addition to the "technology stack - work experience" relationship, there are also rich entity relationships within and between the four modules of basic information, education information, work experience, and technology stack. The following are specific categories and examples: 1. Internal Relationships within Modules: ① Basic Information: For example, name - contact information / email: logical association (e.g., "Zhang San" and "zhangsan@example.com" belong to the same person). Name - address: residential association (e.g., "Zhang San - resides in Chaoyang District, Beijing"). ② Education Information: For example, university - major: affiliation (e.g., "Tsinghua University - major - Computer Science and Technology"). Education - degree / GPA: attribute association (e.g., "Master's - degree level - second-class degree," "GPA - 3.8 / 4.0 - grade attribute"). ③ Work Experience: Company name - position: employment relationship (e.g., "XX Technology Company - position - Java Development Engineer"). Time Range - Job Description: time-event association (e.g., "2020-01 to 2023-01 - responsible for - microservice architecture design"). ④ Technology Stack: Technology Point - Category: hierarchical relationship (e.g., "Spring Boot - belongs to - microservice framework"). Technology Point - Version: attribute association (e.g., "Java - version - 11.0.20").

[0048] II. Inter-module cross-relationships: ① Basic Information - Educational Information: Name - Graduate School: Educational experience association (e.g., "Zhang San - Graduated from - Peking University"). Name - Degree: Qualification association (e.g., "Li Si - Obtained - Doctorate in Engineering"). ② Basic Information - Work Experience: Name - Company Name: Professional resume association (e.g., "Wang Wu - Previously worked at - Alibaba"). Name - Position: Position association (e.g., "Zhao Liu - Currently - Senior Algorithm Engineer"). ③ Education Information - Work Experience: School - Company Name: Campus recruitment association (e.g., "Tsinghua University - Campus recruitment partner - Tencent"). Major - Position: Major matching association (e.g., "Financial Engineering major - Suitable position - Quantitative Analyst"). ④ Education Information - Technology Stack: Major - Technical points: Course association (e.g., "Computer Science major - Courses studied - Data Structures and Algorithms"). Degree - Technical Ability: Research direction association (e.g., "PhD - Research field - Artificial Intelligence model training"). ⑤ Work Experience - Technology Stack: Job Description - Technical points: Skill application association (e.g., "Responsible for building a big data platform - Using technologies - Hadoop / Spark"). Position - Technology Stack: Job requirements (e.g., "Blockchain Development Engineer - Essential Skills - Solidity Smart Contracts"). ⑥ Technology Stack - Basic Information: Technical Points - Name: Skill Mastery (e.g., "Zhang San - Mastery - Python Programming").

[0049] 3. Introducing Extended Relationships with External Knowledge: By integrating external knowledge bases (such as DBpedia and industry terminology libraries), we can further build: Technology Stack - Industry Trends: For example, "Kubernetes - Industry Application Trends - Mainstream Tools for Cloud Native Architecture" (from a technical white paper). Company Name - Industry: For example, "Huawei - Industry - Communications Equipment / Cloud Computing" (from an enterprise information database). Position - Talent Demand: For example, "AI Trainer - Market Demand - Growth Rate of 30% by 2023" (from a recruitment report).

[0050] Specifically, this step enriches relationships: by mining the internal and cross-relationships within the four modules, a complete resume semantic network is constructed, enhancing the depth of analysis (for example, inferring a candidate's professional match from the "name-school-major-position" chain). It also demonstrates industry adaptability: a dynamic weighting mechanism, integrated with external knowledge, enables the system to focus on key relationships tailored to the needs of different industries, demonstrating the solution's innovativeness and engineering practicality.

[0051] S32. Perform knowledge graph fusion based on the entity relationship model; In this example, the fusion process involves first linking external knowledge bases (such as DBpedia and industry terminology libraries), supplementing entity attributes (such as version information and application scenarios for "Java"), and performing multilingual alignment (e.g., "Master of Computer Science" corresponds to "Master of Computer Science"). Furthermore, visual queries are supported (e.g., "Find all Java developers who master Kubernetes"). The final output is the fused, rich knowledge triples (including attributes and cross-domain associations), which represent the entity, relationship, and attribute data in the knowledge graph.

[0052] S33. Store the knowledge graph and complete the construction of the dynamic semantic network.

[0053] In this embodiment, the knowledge-rich triples are stored in a graph database (Neo4j / JanusGraph), a dynamic semantic network architecture of nodes (entities) and edges (relationships) is constructed, and finally a queryable knowledge graph is output (supporting Cypher / SPARQL statement retrieval).

[0054] Furthermore, the process of constructing a dynamic semantic network architecture of nodes (entities) and edges (relationships) is as follows: S331. Perform entity standardization: For example, standardize "Spring Boot" / "SpringBoot" to "SpringBoot" and use an alias mapping table to solve the problem of different expressions of the same entity.

[0055] S332. Combined with the results of S331, perform attribute enhancement processing; specifically, associate with external knowledge bases such as DBpedia to supplement entity attributes, such as adding "Release Date 1995" and "Application Field Enterprise Development" to "Java", and align standard entities through entity linking technology.

[0056] S333. Combine the results of S332 and generate a unique identifier. Specifically, use the UUID algorithm to combine the entity name and type to generate a globally unique ID (such as "TechnicalStack+Java" to generate a unique hash value), ensuring that the IDs of multi-language entities (such as "Java" and "Java Language") are consistent.

[0057] S334. Perform hierarchical classification on the generated unique identifier; specifically, classify entities according to a predefined label system (such as "programming language" and "framework"), and build hierarchical relationships (such as "Java→programming language") to support subsequent semantic retrieval.

[0058] In some embodiments, an industry-adaptive mechanism is applied throughout the entire process from S2 to S3. This mechanism includes at least two learning mechanisms: transfer learning and active learning. Specifically, during the information extraction phase of S2, the transfer learning mechanism uses domain corpora (financial / medical resumes) to fine-tune the BERT model, improving its ability to recognize domain entities (e.g., "anti-money laundering" and "MRI images"). In S3, the relationship weight assignment process adjusts weights based on domain characteristics (e.g., financial positions prioritize the "education-certificate" relationship). During the information extraction phase of S2, the active learning mechanism automatically annotates difficult cases (e.g., ambiguous technical terms) using HR feedback data to optimize model parameters. Furthermore, the knowledge graph in S3 dynamically updates the domain terminology library (e.g., adding a "Large Language Model Training" tag). For example, transfer learning fine-tunes the BERT model during the information extraction phase of S2 to improve entity recognition and is a technical implementation detail of that step. Dynamic weight adjustment occurs during the relationship construction process in S3 and is a strategic configuration for that step. Furthermore, S2-S3 optimizes information extraction accuracy and semantic network dynamics through transfer learning and active learning via industry-adaptive mechanisms, ultimately outputting structured data tailored to industry needs. It's important to note that transfer learning can influence model training in S2, and dynamic weighting can influence relationship building in S3.

[0059] S4. Based on the stored knowledge graph, output structured data and complete resume parsing.

[0060] In this embodiment, the entity, relationship, and attribute data in the knowledge graph are input; the graph data is converted into a standard JSON format, integrating basic information, education, work experience, technology stack, and semantic network relationships; and finally, structured data is output for downstream system calls, such as recruitment platform search and talent pool management.

[0061] In this embodiment, the system architecture and technology selection are shown in Table 1 below: Table 1

[0062] The technical effect verification is shown in Table 2 below: Table 2

[0063] Through the above-mentioned refinement, the technology stack parsing module can accurately capture the implicit knowledge associations in resumes, and the dynamic semantic network can adaptively adjust the relationship strength according to industry characteristics, significantly improving the accuracy of resume parsing and industry adaptation efficiency.

[0064] In other embodiments, combined Figure 3 As shown, a resume parsing system based on a dynamic semantic network can also be provided. The resume parsing system is used to implement the aforementioned resume parsing method based on a dynamic semantic network. The resume parsing system includes a data input module, a preprocessing module, an information extraction module, a semantic network construction module, a knowledge graph system module, an industry adaptation module, a structured output module and a performance optimization module; each module works together through a data interface to form a complete resume intelligent parsing link.

[0065] Specifically, the data input module is used to execute S1, which can perform resume input operations and transmit data to the preprocessing module; the core functions of the data input module are: supporting resume uploads in multiple formats such as PDF / Word / picture / HTML; and providing an API interface to connect to the recruitment platform and enterprise OA system.

[0066] The preprocessing module is used to perform preprocessing operations on the input resume file, that is, to execute step S21; wherein, the data input module transmits data to the preprocessing module; its core functions are: it can use PyPDF2 / docx to extract text, remove noise, and parse PDF / Word format documents; it can also use Tesseract+OpenCV to perform image enhancement operations on images.

[0067] The information extraction module is used to extract information, that is, to execute step S22; the data it processes is transmitted to the semantic network construction module; its core function is to parse the four modules of basic information, educational information, work experience and technology stack.

[0068] The industry adaptation module is mainly used to implement two learning mechanisms. The data it processes is transmitted to the information extraction module and the semantic network construction module. Among them, the industry adaptation module can also perform the following operations: (1) Dynamic configuration center: storing industry-specific terminology libraries (such as "anti-money laundering" and "credit risk control" in the financial field) and model parameters (such as the BERT fine-tuned version of medical resumes). (2) Support for hot updates: loading new industry configurations in real time through the API interface without restarting the system.

[0069] This performance optimization module is mainly used to perform the technology stack analysis performance optimization process, which is also the S23 step; the data it processes is also transmitted to the information extraction module, and the performance optimization module adopts a hybrid expert architecture and cache mechanism to improve the efficiency of technology stack analysis. Among them, the performance optimization module can also perform the following operations: (1) Hybrid expert routing mechanism: Based on the industry keywords in the resume text (such as "blockchain" and "clinical trials"), the BERT expert model of the corresponding industry is automatically scheduled to reduce the redundant calculation of the general model. (2) Multi-level cache architecture: Level 1 cache: Memory stores high-frequency term query results (such as matching results for "Java" and "Python"). Level 2 cache: Distributed cache (such as Redis) stores model inference results and supports cross-node sharing.

[0070] The semantic network construction module is used to execute the S3 step, and the processed data is transmitted to the knowledge graph storage module; its core functions include entity relationship modeling, knowledge graph fusion, dynamic weight allocation, etc.; among them, the semantic network construction module can also perform the following operations: (1) Relationship extraction engine: based on Stanford NLP, it implements dependency syntax analysis to extract relationships such as "technology stack-work experience" and "education-position"; (2) Weight allocator: loads predefined weight strategies according to industry type (such as technical position / financial position) (such as increasing the weight of technology stack related relationships to 60%).

[0071] The structured output module is used to execute the S4 step. Its core functions include data format conversion and structure services.

[0072] Example 1 The process of parsing a resume for a technical position using the above parsing method is as follows: S1. Enter resume: Enter the resume of a Java development engineer, where the resume is a scanned PDF (including image text); S2. Based on the resume file type, the multimodal resume parsing engine processes the resume and extracts information. The specific processing steps are as follows: First, perform document preprocessing: extract the scanned PDF text through PyPDF2, combine it with OpenCV to enhance the image clarity, and output plain text content, such as "5 years of Java development experience, proficient in Spring Boot and MyBatis, and familiar with Docker container deployment."

[0073] Then perform information extraction operation: (1) Technology stack information: First, the term base matching module is used to identify keywords such as "Java", "Spring Boot", and "Docker" from the text (without relying on continuous sentences); then the BERT model is used to perform semantic verification on "Docker" and "containerized deployment", and the vector cosine similarity is calculated to be 0.89, confirming that the two are in a "tool-function" relationship, and generating a technology stack tag list [Java, Spring Boot, Docker].

[0074] (2) Work experience information: The LSTM-CRF model is used to identify "Company Name: XX Technology Co., Ltd." and "Position: Java Development Engineer," and extract the job keywords "Microservice Architecture Design" and "Docker Containerized Deployment." Specifically, the in-depth implementation process of the technology stack information is as follows: 1. Termbase matching and synonym expansion: (1) Term library structure design: A hierarchical classification system is used to store industry terms. For example, the Internet field is divided into subcategories such as "programming language", "framework", and "tool". Each term is associated with its synonym set (such as "microservice" is associated with "Microservices" and "distributed architecture").

[0075] (2) A hybrid matching strategy is adopted, whose design goal is to gradually relax the matching conditions from "exact matching" to "fuzzy expansion" while combining semantic analysis to improve accuracy. The specific process is: The first step is to perform exact matching. The logic behind this is that the terminology database stores industry-standard terms (such as "Java" and "Spring Boot"). Exact matching quickly locates specific technical points, maximizing efficiency and minimizing false positives. For example, directly matching keywords like "Java" and "Docker" eliminates the need for complex calculations to identify the technology stack label. The final output is a set of explicitly matched terms (e.g., ["Java," "Docker"]).

[0076] The second step involves fuzzy matching. This logic addresses terminology variations (such as hyphens and mixed case) and non-standard expressions due to differences in resume formatting. For example, "SpringBoot" and "Spring-boot" are matched uniformly to "Spring Boot" using regular expressions; "K8s" is matched to "Kubernetes" (abbreviation mappings are predefined in the terminology library). The final output is a standardized set of fuzzy-matched terms (e.g., ["Spring Boot"]).

[0077] The third step is context-aware matching. This logic is based on the following: Based on the matched terms, dependency syntactic analysis is used to identify the modification relationship between the terms and the context, resolving polysemy and compound terminology issues. For example, in the sentence "Proficient in Spring Boot microservice architecture design," "Spring Boot" is first identified through exact matching. Then, syntactic analysis determines that "microservice architecture" is its modifier and adds the tag "microservice architecture." This distinguishes "Python development" (programming language) from "Python crawler" (application scenario) to avoid mixing irrelevant tags. The final output is an expanded set of terms with contextual modifications (e.g., ["Spring Boot," "microservice architecture"]).

[0078] The fourth step is to perform a global catch-all synonym expansion match. The logic behind this is: for text segments that miss the first three levels of matching, potential terms are expanded using industry dictionaries and BERT semantic computation to cover less-frequent or emerging technologies. For example, if there is no exact match for "containerized deployment," "Docker containerization" is associated with it through a synonym library. BERT then verifies its semantic similarity with "Docker" (e.g., a cosine value of 0.85), ultimately adding the "Docker" label. For the emerging technology "Large Language Model Training," BERT semantic computation is used to associate it with standard terms such as "LLM development" and "NLP model training." The final output is a semantically verified synonym term set (e.g., ["Docker"]).

[0079] In this embodiment, the matching strategies of the first to fourth steps are executed sequentially. The technical advantages thereof are as follows: 1. Efficiency first: First, quickly capture high-confidence terms through exact matching and fuzzy matching to reduce the computational complexity of subsequent semantic analysis.

[0080] 2. Increasing semantic reliability: Gradually transition from "formal matching" (exact / fuzzy) to "semantic matching" (context / synonymous) to ensure that the expanded terms are logically consistent with the resume content.

[0081] 3. Comprehensive coverage of scenarios: Exact matching can handle standard terminology; fuzzy matching is compatible with format variations; context-aware matching can parse complex expressions; and synonym expansion matching can uncover hidden technical points.

[0082] In some other embodiments, there are exceptions and dynamic adjustments, such as industry-specific adjustments: in the financial industry, "Anti-Money Laundering" may have variants such as "AML" and "Anti-Money Laundering", which may prioritize triggering fuzzy matching (the second step) rather than exact matching.

[0083] Avoiding frequent errors: For easily confused terms (such as "React" and "React Native"), we add an ambiguity resolution step after exact matching (for example, determining whether it is mobile development based on context) to avoid the introduction of incorrect labels through synonym expansion.

[0084] 2. Enhanced implementation of BERT semantic verification: (1) Context feature extraction: The resume text is segmented into sentence-level segments (such as "familiar with Docker container deployment") and input into the BERT model to generate contextual embedding vectors.

[0085] Extract the token embedding (such as the vector of "Docker") and sentence-level CLS embedding of the term location as feature input for semantic association.

[0086] (2) Calculate the correlation strength. The process is as follows: The cosine similarity method is used to calculate the vector similarity of term pairs (such as “Docker” and “containerized deployment”), and the threshold is set to 0.7.

[0087] Using the relationship classification task, the fine-tuned BERT model is used to classify term pairs (such as the "tool-function" relationship) with an accuracy of 91.5%.

[0088] (3) Conduct cross-domain verification: For complex descriptions (such as "Java development of financial risk control system"), verify the relevance of the technology stack and business scenarios by combining it with industry terminology (such as "credit model" in the financial field).

[0089] S3. Construct a dynamic semantic network based on the extracted information. The specific construction process is as follows: S31. Constructing entity relationship model: The specific process is as follows: (1) First, perform step decomposition: Use StanfordNLP to perform dependency syntax analysis on the technology stack keywords (Java, SpringBoot, Docker) and work experience and job descriptions extracted by S2 to identify the relationship between entities.

[0090] (2) Apply the relationship type: Hierarchical relationship: Classify "Java" as "Programming language" and establish the "Java → Programming language" relationship. Dependency relationship: Analyze "Building a microservice architecture using Spring Boot" and determine the "Spring Boot → Microservice architecture" dependency relationship. Tool-function relationship: Extract the "Docker → Containerized deployment" relationship from "Implementing containerized deployment through Docker".

[0091] (3) Dynamic weighting: For technical positions, the weight of the "technology stack-work experience" relationship is set to 60% according to the preset rules. For example, the relationship "Java-used for-e-commerce platform development" is given a high weight. Combined with HR feedback data, if the position JD frequently mentions "Docker", the weight of the "Docker-containerized deployment" relationship is dynamically increased (for example, from 0.6 to 0.8).

[0092] S32. Fusion of knowledge graphs based on entity relationship models: The specific process is as follows: Linking external knowledge bases: Connecting to DBpedia, supplementing "Java" attribute information (such as release time and application scenarios), and improving the "Java-Technology Stack" node; introducing industry terminology libraries, linking the English expression "MicroservicesArchitecture" to "MicroservicesArchitecture" to enhance internationalization support; Multilingual relationship expansion: Use the mBERT model to achieve cross-language alignment and establish a bilingual relationship chain from "Spring Boot (Chinese) - Framework" to "Spring Boot (English) - Framework".

[0093] S33. Store the knowledge graph and complete the construction of the dynamic semantic network. This process involves storing the entity relationships (e.g., "Java - Programming Language," "Spring Boot - Microservice Architecture," "Docker - Containerized Deployment") and their weights as triples in a knowledge graph (e.g., Neo4j). A visual interface allows users to intuitively view the complete relationship chain from "Java - Technology Stack" → "Spring Boot - Framework" → "Docker - Tool," and supports querying related work experience (e.g., "Developing an e-commerce platform using Java and Spring Boot, deploying via Docker").

[0094] Specifically, the in-depth implementation process of dynamic semantic network construction is as follows: 1. Optimize entity relationship modeling: (1) Definition of relationship type: Hierarchical relationship: such as "Java → programming language".

[0095] Dependency: For example, "Spring Boot → Microservice Architecture".

[0096] Tool-function relationship: such as "Docker → containerized deployment".

[0097] (2) Setting of dynamic weight mechanism: Industry characteristic weight: The weight of the relationship between "technology stack-work experience" in technical positions accounts for 60%, and the weight of the relationship between "certificate-job requirements" in financial positions accounts for 50%.

[0098] Data-driven adjustments: Dynamically adjust relationship weights based on HR feedback data (such as the frequency of "Kubernetes" in job JDs). For example, the relationship weight of "Kubernetes-Containerization" was increased from 0.6 to 0.8.

[0099] In this embodiment, the frequency of dynamic adjustment of relationship weights is positively correlated with the dynamic adjustment of weights. The specific logic is as follows: 1. Positive Association (Increased Frequency → Increased Weight): When a keyword (such as "Kubernetes") appears significantly more frequently in job postings or HR feedback, it indicates a growing demand for this technology in the target position. The system will automatically increase the weight of the relationship between this keyword and related terms to reinforce their importance in the semantic network. For example, if the frequency of "Kubernetes" in recent JDs increases from 10 to 30 times per month, indicating a growing demand for "containerization" skills, the system will increase the weight of the "Kubernetes-Containerization" relationship from 0.6 to 0.8 (a larger weight indicates a stronger association).

[0100] 2. Negative correlation (decreasing frequency → decreasing weight): If the frequency of a keyword continues to decrease, the system will deem the demand for that technology point to be decreasing and automatically reduce the weight of its relationship with related terms. For example, if the frequency of "Docker Swarm" in JD drops from 20 to 5 times per month, it indicates that the demand for this tool in the position is decreasing. The system may reduce the weight of the relationship "Docker Swarm - container orchestration" from 0.7 to 0.5.

[0101] Further details on the implementation of the dynamic weight adjustment mechanism: Its data sources primarily rely on HR feedback (such as job JD documents and recruitment requirements change records) and historical resume analysis results. Keyword frequency is extracted through natural language processing (NLP). The trigger condition is a set frequency change threshold (such as a ±20% week-over-week / month change). When the actual frequency exceeds the threshold, the weight adjustment process is triggered. Adjustments are limited to industry-wide terms or high-frequency terms for positions (such as technology stack and business scenario keywords), avoiding over-adjustment of low-frequency or outdated terms.

[0102] Furthermore, the weight dynamic adjustment mechanism realizes the dynamic optimization of the semantic network through "data-driven", making the resume parsing more in line with the real-time recruitment needs. For example, if a large number of JDs in a certain industry recently emphasize "Kubernetes", the system can give it a higher weight, so that relevant skill descriptions related to "Kubernetes" can be identified more preferentially during resume parsing, avoiding missing key information. On the contrary, if the frequency of a certain technology (such as "Struts framework") decreases, after the weight is reduced, the system will reduce the priority of its associated relationships and focus on more mainstream technical points.

[0103] 2. Fusion and expansion of the knowledge graph: (1)External knowledge base integration: DBpedia: Supplement the version information (such as Java8, Java11) and application scenarios of "Java".

[0104] Industry terminology library: Integrate terms in fields such as "fintech" and "anti-money laundering", and associate their English equivalents (such as "anti-money laundering" → "Anti-Money Laundering").

[0105] (2)Multi-language support: Cross-language entity alignment: Use the mBERT model to establish associations between Chinese terms (such as "microservices") and English terms ("Microservices").

[0106] Bilingual relationship chain: Establish a cross-language relationship of "Java (Chinese) - technology stack" → "Java (English) - Programming Language".

[0107] 3. Semantic parsing of complex technology stacks: (1)Dependency syntactic analysis: Conduct syntactic analysis on "leading the design of Spring Boot microservice architecture", extract "Spring Boot" as the tool for "design", and "microservice architecture" as the target of design, and establish the relationship of "Spring Boot - tool → microservice architecture - target".

[0108] (2)Implicit knowledge mining: Temporal relationship: Analyze the usage order of the technology stack through timestamps (such as "using Spring Boot in 2018 and switching to Kubernetes in 2020").

[0109] Project association: Associate the technology stack with project results (such as "using Docker to achieve continuous integration and the delivery efficiency is increased by 30%").

[0110] S4: Based on the stored knowledge graph and industry-adaptive mechanisms, the system outputs structured data to complete resume parsing. Specifically, the structured data output includes technology stack tags (Java, Spring Boot, Docker) and work experience (company name, position, project responsibilities).

[0111] Example 2 The technical background for adapting resume parsing for the financial industry is that resumes in the financial industry have the characteristics of strong compliance, qualification orientation, and specialized terminology (such as "anti-money laundering" and "credit risk control"), and the parsing logic needs to be optimized in a targeted manner.

[0112] Specific adaptation process: S1. Enter your resume: This is a Word document for a financial analyst (including a tabular list of educational background and certifications). Key information includes: Education: Undergraduate degree in Finance from Peking University, CFA charterholder (passed in 2017); Work experience: 8 years of credit risk control experience in the Risk Management Department of XX Bank, responsible for anti-money laundering compliance reviews; Technology stack: Python (financial modeling), SQL, anti-money laundering regulations, and credit risk control systems. Format characteristics include: Certification information is often presented in a tabular format (e.g., "Certificate Name - Date Obtained"); and job descriptions include industry compliance terms (e.g., "FATF Standards" and "Customer Due Diligence").

[0113] S2. Based on the resume file type, perform multimodal resume parsing engine processing and extract information; For example, the customized processing process for the financial industry is as follows: S21. Preprocess the resume file and output plain text data in a unified format. Specifically, use python-docx to parse the table and extract structured data such as "CFA charterholder (2017-06)" and "Peking University Finance (2007-09 to 2011-06)". Segment processing is performed on text paragraphs such as "Anti-money laundering compliance review" to facilitate subsequent semantic analysis.

[0114] S22. extracting information based on the output plain text data; Specifically, the extracted basic information is: the BERT-BiLSTM-CRF model recognizes "Name: Li Si" and "Email: lisi@finance.com", and newly adds "Financial Practitioner Qualification Certificate Number" entity recognition.

[0115] The extracted educational information is as follows: a regular expression matches the "CFA charterholder" pattern, and dependency parsing is used to extract "CFA-Level III-passed-in-2017." Multi-language support is also provided: "Master of Finance" is identified and translated into "Master of Finance."

[0116] The extracted work experience is: LSTM-CRF model recognition of "Company Name: XX Bank", "Position: Financial Analyst", "Time Range: 2015-06 to Present"; attention mechanism extraction of responsibility keywords: "Credit risk control model development", "Anti-money laundering regulations compliance review", "Customer risk rating".

[0117] The extracted technology stack information is as follows: (1) Terminology matching: triggering the financial field terminology database to identify “anti-money laundering”, “credit risk control”, and “Python (financial modeling)”; (2) BERT semantic verification: verifying the association between “Python” and “financial modeling” (cosine similarity 0.88), and confirming the compliance association between “anti-money laundering” and “FATF standards” (verified by the industry knowledge base).

[0118] S3, building a dynamic semantic network based on the extracted information; Specifically, for financial scenarios, S31, build an entity relationship model; the process is: Dependency syntax analysis: Extract the relationships from "Responsible for anti-money laundering regulations compliance review": "Li Si-Execution-Anti-Money Laundering Review" and "Anti-Money Laundering Regulations-Based on FATF Standards"; extract from "Using Python to develop credit risk control models": "Python-Tools-Credit Risk Control Model Development".

[0119] Dynamic weighting mechanism: Industry characteristic weight: The weight of the "certificate-job requirements" relationship in financial positions is increased to 50% (such as "CFA-matching-financial analyst"); Compliance weight: The weight of the "anti-money laundering-responsibilities-risk management department" relationship is set to 0.9 (higher than the general responsibility relationship).

[0120] S32. Perform knowledge graph fusion based on the entity relationship model; Specifically, link to external knowledge bases: access the "Financial Industry Compliance Terminology Library" to supplement "Anti-Money Laundering - Regulatory Agency - People's Bank of China" and "CFA - Examination Subjects - Ethics and Professional Standards"; link to Wikidata to obtain corporate information such as "XX Bank - Establishment Date - 1995 - Headquarters - Beijing".

[0121] Multilingual integration: Establish a cross-language relationship chain of "Anti-Money Laundering-Chinese" → "Anti-Money Laundering-English" → "AML-abbreviation".

[0122] S33. Store the knowledge graph and complete the construction of the dynamic semantic network; for example, for the storage triple examples: (Li Si, holds, CFA certificate, weight = 0.95); (credit risk control model, application tool, Python, weight = 0.85); (anti-money laundering regulations, belongs to, financial compliance, weight = 1.0).

[0123] It should be noted that the two learning mechanisms used in financial scenarios are: Transfer learning application: Using a corpus of resumes in the financial sector (including over 100,000 annotated data points for "credit risk control" and "financial technology"), we fine-tuned the BERT model. The F1 value of the technology stack recognition increased from 85% to 92%, and the recognition accuracy of educational information (such as "CFA charter") reached 98.5%.

[0124] Active learning optimization: HR annotates difficult cases such as "green finance bond issuance experience," automatically expands the terminology library, and adjusts the weight of the "green finance-bond-issuance" relationship; the double-buffer architecture enables hot model updates, completing the iteration of parsing capabilities for emerging terms such as "ESG investment" within 24 hours.

[0125] Specifically, the optimization process of transfer learning is: 1. Build corpora in different fields: (1) Financial field: Collect resume texts containing "CFA" and "credit risk control" to construct a corpus of 200,000 annotated records.

[0126] (2) Medical field: Integrate terms such as "MRI image analysis" and "electronic medical record system" to construct a corpus of 150,000 annotated items.

[0127] 2. Fine-tune the BERT model: (1) Layer freezing: Freeze the first 6 layers of BERT and only fine-tune the last 4 layers to adapt to domain characteristics.

[0128] (2) Mixed training: The domain corpus and the general resume corpus are mixed for training in a ratio of 1:3, and the F1 value of the technology stack recognition is increased to 92%.

[0129] It should be noted that layer freezing and hybrid training are complementary technical approaches: the former optimizes training efficiency through network structure design, while the latter improves model robustness through data strategies, both of which serve the core goal of domain adaptation. The following is a detailed analysis: 1. Logical relationship analysis 1. The role of layer freezing (1): balancing generalization and domain adaptation. Principle: The bottom layer of the BERT model (the first few layers) primarily learns general language features (e.g., parts of speech, basic semantics), while the upper layers (the last few layers) are better at capturing domain-specific semantics (e.g., technical terms and job descriptions in resumes). Operation: Freeze the first six layers (to retain general semantic capabilities) and only fine-tune the last four layers (to adapt to resume domain features). Goal: Avoid "catastrophic forgetting" (loss of general semantics) caused by directly fine-tuning the entire model, while allowing the model to focus on key information in the resume domain (e.g., technology stack, action words in work experience).

[0130] 2. The role of hybrid training (2): Enhance the representativeness of domain data. Principle: Pure domain corpus (e.g., a small number of resumes from the financial industry) may have data sparsity issues, and direct training is prone to overfitting; general resume corpus (e.g., a large number of resumes from the internet industry) contains richer basic semantic patterns. Operation: Hybrid training is performed with a domain corpus: general corpus ratio of 1:3, leveraging the diversity of general data to enhance model robustness while injecting industry characteristics through domain data. Goal: Address the issue of "uneven distribution of technology stacks" in resume parsing (e.g., a small number of samples of certain niche technical terms) and improve the model's ability to identify rare technical points (e.g., industry-specific tools).

[0131] 3. Synergistic effect: (1) Layer freezing provides a structural foundation for hybrid training: After freezing the underlying network, the data for hybrid training does not need to bear the task of "reconstructing general semantics". It only needs to learn the mapping relationship between domain features and general features through the high-level network (such as associating "development" in the general corpus with "smart contract development" in the domain corpus). (2) Hybrid training provides data support for layer freezing: A reasonable proportion of hybrid data avoids the risk of overfitting when fine-tuning with only domain data, ensuring that the activated last four layers of the network can learn stable domain patterns (such as the contextual association of the technology stack).

[0132] 2. Logical verification of technical effects The causal chain for improving F1 scores: Layer freezing (preserving general semantics) + hybrid training (supplementing domain semantics) → The model possesses both generalization capabilities (recognizing common entities) and domain sensitivity (capturing industry-specific keywords) → Improved accuracy in identifying technology stacks. For example, in a finance resume, the model recognizes the general semantics of "system development" through a frozen underlying network. Then, through a fine-tuned higher-level network combined with hybrid training data, it associates "blockchain + smart contract development" with a specific technology stack, avoiding misclassification as a general development skill.

[0133] This paper preserves general language comprehension capabilities by freezing the underlying BERT network, while adopting a domain-general corpus hybrid training strategy to enable the model's high-level network to efficiently learn the special semantic patterns in the resume domain (such as industry-specific expressions of the technology stack). The combination of the two solves the overfitting problem caused by insufficient data in a single domain and the semantic drift problem caused by full model fine-tuning.

[0134] In this example, the active learning process involves HR identifying difficult cases such as "credit model development experience," and the model automatically optimizing the weighting of the "credit-risk control" relationship. A hot update mechanism dynamically adjusts model parameters, boosting the F1 score for the technology stack to 92%.

[0135] Specifically, the enhancement process of active learning is: 1. Screen difficult cases first: (1) Confidence threshold: annotate cases where the term association probability output by the BERT model is lower than 0.6 (e.g., “Large language model training”).

[0136] (2) Conflict detection: When multiple HRs have inconsistent annotations for the same case, secondary annotation is automatically triggered.

[0137] 2. Then perform hot update of the model: (1) Double buffer architecture: Maintain two model instances (A and B). After new data is annotated, model B is updated first. After verification, it is switched to the production model. The service interruption time is <1ms.

[0138] (2) Incremental training: Only newly labeled data is trained, and the model iteration cycle is shortened to 24 hours.

[0139] S4. Based on the stored knowledge graph, output structured data and complete resume parsing.

[0140] Specifically, the structured data output (fields specific to the financial industry) is: { "Basic Information": { "Name": "Li Si", "Financial Qualification": "CFA Charterholder (2017)", "Compliance Period": "8 Years" }, "Education Information": [ { "School": "Peking University", "Major": "Finance", "Certificate": ["CFA Level III"] } ], "Work Experience": [ { "Company": "XX Bank", "Position": "Financial Analyst", "Compliance responsibilities": ["Anti-money laundering review", "Credit model development"], "Technology Stack": ["Python (Financial Modeling)", "Anti-Money Laundering Regulations"] } ], "Industry Adaptation Features": { "Compliance Matching Degree": 0.92, "Certificate Priority": 50%, "Regulatory Link": ["People's Bank of China", "FATF"]} } It should be noted that the core differences in financial industry adaptation include but are not limited to: (1) Information extraction focuses on: adding entities such as "financial practitioner qualifications" and "years of compliance" to strengthen the identification of certificates and compliance responsibilities; (2) Semantic network characteristics: highlighting industry-specific relationships such as "certificate-position" and "compliance-regulatory agency", with weight distribution tilted towards qualifications and compliance; (3) External knowledge integration: Deeply integrate financial regulatory policies and industry standards (such as FATF and Basel Accord) to enhance the compliance verification capability of analytical results.

[0141] 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 parsing method based on dynamic semantic network, characterized in that: The following steps are involved: S1. Enter your resume; S2. Extract information based on the resume file type and optimize the multimodal resume parsing engine; S3, building a dynamic semantic network based on the extracted information; The construction process is: S31. Build entity relationship model; S32. Perform knowledge graph fusion based on the entity relationship model; S33. Complete the construction of the dynamic semantic network and store the knowledge graph; S4. Based on the stored knowledge graph and combined with the industry adaptive mechanism, it outputs structured data and completes resume parsing.

2. A resume parsing method based on dynamic semantic network according to claim 1, characterized in that: The logical process of S2 is: S21. Preprocess the resume file and output plain text data in a unified format; S22. extracting information based on the output plain text data; Among them, the logical process of information extraction is: S221, first input the plain text data output after preprocessing; S222. Extract basic information, educational information, work experience, and technology stack information based on plain text data; S223. Output structured information based on the extracted information.

3. A resume parsing method based on dynamic semantic network according to claim 2, characterized in that: In S222, (1) Basic information: Identify name and contact information based on the BERT-BiLSTM-CRF model, and finally output a structured entity list; (2) Educational information: Regular expression matching is used, combined with dependency syntax analysis to extract GPA and degree information, supporting multiple languages, and ultimately outputting a list of educational entries; (3) Work experience: Use the LSTM-CRF model to identify company name, position, and time range; and use the attention mechanism to extract key events in the job description, and finally output a list of experiences with a timeline; (4) Technology stack: Use terminology matching and BERT semantic verification to build a set of technology points and finally output a list of technology stack labels.

4. A resume parsing method based on dynamic semantic network according to claim 2, characterized in that: The logical process of S2 also includes: S23, performing technology stack analysis and optimization operations based on the extracted information; The analytical optimization operation process of S23 includes: 1) Using hybrid expert architecture for distributed computing architecture: (1) Expert model: allocate independent BERT models to different industries and dynamically schedule computing resources; (2) Load balancing: Automatically scale expert model instances based on resume processing volume, supporting processing of 20 resumes per second; 2) Cache and preloading operations: (1) Term library cache: Cache high-frequency terms in memory, increasing query speed by 3 times; (2) Model preloading: The BERT model is loaded when the system starts, and the cold start time is less than 5 seconds.

5. The resume parsing method based on dynamic semantic network according to claim 1, characterized in that: The logic of S31 is: first extract the semantic relationship between entities in the resume text through dependency syntax analysis, then set dynamic weights for each type of relationship based on industry characteristics, and then complete the construction of the entity relationship model.

6. The resume parsing method based on dynamic semantic network according to claim 1, characterized in that: The fusion process of S32 is as follows: first, associate with the external knowledge base, supplement entity attributes, and perform multilingual alignment; It also supports visual query; the final output is the fused rich knowledge triples, which are the entities, relationships, and attribute data in the knowledge graph.

7. The resume parsing method based on dynamic semantic network according to claim 1, characterized in that: In S33, the process of building a dynamic semantic network architecture is as follows: S331, perform entity standardization processing; S332, combining the result of S331, performing attribute enhancement processing; S333. Generate a unique identifier based on the result of S332. S334. Perform hierarchical classification processing on the generated unique identifier.

8. The resume parsing method based on dynamic semantic network according to claim 1, characterized in that: The logical process of S4 is: input the entity, relationship, and attribute data in the knowledge graph; convert the graph data into standard JSON format, integrate basic information, education, work experience, technology stack and semantic network relationships; and finally output structured data.

9. The resume parsing method based on dynamic semantic network according to claim 1, characterized in that: An industry adaptive mechanism is also used in the entire process of steps S2-S3, wherein the industry adaptive mechanism includes at least two learning mechanisms: transfer learning and active learning.

10. A resume parsing system based on dynamic semantic network, characterized in that: Implementing a resume parsing method based on a dynamic semantic network as described in any one of claims 1 to 9, the resume parsing system includes a data input module, a preprocessing module, an information extraction module, a semantic network construction module, a knowledge graph system module, an industry adaptation module, a structured output module, and a performance optimization module; The data input module is used to execute S1, perform resume input operation, and transmit data to the preprocessing module; The preprocessing module is used to perform preprocessing operations on the input resume file, execute step S21; and transmit the data communication to the preprocessing module; The information extraction module is used to extract information and execute step S22; the processed data is transmitted to the semantic network construction module; The industry adaptation module is used to implement two learning mechanisms, and the data it processes is transmitted to the information extraction module and the semantic network construction module; The performance optimization module is mainly used to execute the technology stack analysis performance optimization process and execute step S23; the data processed by it is also transmitted to the information extraction module; The semantic network construction module is used to execute the S3 step, and the processed data is transmitted to the knowledge graph storage module; The structured output module is used for the execution of step S4.

Images