An AI large model-based resume intelligent analysis and information extraction method

Abstract

The present application relates to the technical field of data processing, and more particularly to a resume intelligent analysis and information extraction method based on an AI large model, which determines main influencing factors of abnormal categories, determines main abnormal specific categories of a to-be-processed resume, analyzes abnormal significance of the to-be-processed resume for a first abnormal specific category, performs layout reconstruction and multi-modal semantic alignment on the to-be-processed resume, divides the to-be-processed resume into regions, performs regional layout determination, decouples each region into an independent semantic unit, extracts industry-related entities, establishes regional logical association, performs regional information analysis based on the logical association, disassembles work experience of the to-be-processed resume into a plurality of transferable labels for a second abnormal specific category, constructs a target post demand graph, and performs label matching analysis. The present application improves the analysis rate and accuracy of resume analysis by analyzing multiple specific abnormal categories of a to-be-processed resume and adopting different analysis and extraction methods for different categories.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for intelligent resume parsing and information extraction based on an AI large model. Background Technology

[0002] Intelligent resume parsing is a fundamental step in achieving efficient job matching. Traditional methods have evolved through stages such as rule-based template matching and deep learning sequence labeling, but they still suffer from low accuracy and weak generalization ability when dealing with complex resume formats, inconsistent semantic expressions, and differences in titles across companies. In recent years, while large language models have upgraded resume parsing from keyword matching to semantic understanding, they still require additional modules to handle complex layouts such as multi-column layouts and nested tables. Furthermore, they lack fine-grained modeling of entity relationships in the recruitment field and suffer from drawbacks such as high inference costs, slow response times, and privacy compliance risks. Therefore, there is an urgent need for an AI large-scale model parsing method that can systematically process multi-format resume layouts, adaptively resolve semantic ambiguity, and meet compliance requirements.

[0003] Chinese Patent Publication No. CN116502625A discloses a resume parsing method and system. The method includes: acquiring a PDF document and an image corresponding to the resume; determining the text line positions of one or more text lines in the resume based on the image; determining one or more first characters and their positions in the resume based on the PDF document; matching the one or more text lines and the one or more first characters based on their positions to obtain a matching result; and generating parsed text corresponding to the resume based on the matching result.

[0004] Chinese Patent Publication No. CN119227670A discloses a deep learning-based intelligent resume OCR parsing and automated information extraction system and method, including a resume acquisition module, a resume preprocessing module, a resume text acquisition module, a resume text analysis module, a comprehensive resume analysis module, a discrimination module, and a feedback module. This invention acquires a first resume text sent by a user, performs preprocessing operations to obtain the first text, and then divides it into sub-texts. Various data collected from the sub-texts are analyzed and processed, and then a mathematical model is used for comprehensive analysis to calculate the information accuracy index of the target text. This helps to present a more intuitive effect, facilitates decision-making, and ultimately improves the comprehensiveness and security of automated information extraction from intelligent resumes.

[0005] However, the following problems still exist in the existing technology. Most existing resume parsing and information extraction technologies are horizontal parsing and extraction modes that linearly extract text from documents and then perform structured processing. If a resume is formatted in two or more columns, using existing technologies for resume parsing and extraction will result in the content of different columns being incorrectly spliced ​​into the same text stream, leading to logical errors in the output, loss of key information, or misalignment of information. Summary of the Invention

[0006] To address this issue, the present invention provides a method for intelligent resume parsing and information extraction based on an AI large model. This method solves the problem that most existing technologies only focus on the storage stage, have a single quality verification mechanism before data entry, and lack dynamic evaluation of data quality. As a result, interfered data may be directly uploaded to the blockchain, which weakens the clinical credibility of on-chain data.

[0007] To achieve the above objectives, this invention provides a method for intelligent resume parsing and information extraction based on an AI large-scale model, comprising: Obtain the resume to be processed, and identify the first and second features in the resume based on the AI ​​large model; Based on the first feature and the second feature, determine whether the resume to be processed belongs to the abnormal handling category; For the anomaly handling category of the resume to be processed, determine the main influencing factors of the anomaly category, and use them to determine the main anomaly specific category of the resume to be processed; For the first anomaly specific category, the anomaly salience of the resume to be processed is analyzed, and the resume to be processed is restructured in layout and aligned with multimodal semantics. Alternatively, the resumes to be processed can be divided into regions, regional layout can be determined, each region can be decoupled into an independent semantic unit, industry-related entities of each independent semantic unit can be extracted, regional logical relationships can be established, and regional information analysis can be performed based on the logical relationships. For the second abnormal category, the work experience of the resume to be processed is decomposed into several transferable tags, a target job requirement map is constructed, and tag matching analysis is performed. The first feature is the length of the continuous blank area between adjacent text lines and the vertical alignment offset between columns; the second feature is the entity type distribution density and the continuity of the resume timeline.

[0008] Furthermore, if the length of the continuous blank area is greater than the reference blank length or the vertical alignment offset between columns is greater than the reference offset, then there is a format specificity. If the matching degree between the entity type distribution density and the entity type required for the job is lower than the first preset benchmark value, or if the continuity of the resume timeline exceeds the second preset benchmark value interval, then there is candidate category specificity.

[0009] Further, the determination of whether the resume to be processed falls under the exception handling category, wherein, If the resume to be processed meets the preset conditions, then the category of the resume to be processed is determined to be the exception handling category; If the resume to be processed does not meet the preset conditions, it is determined that the category of the resume to be processed is not an exception handling category; The preset condition is that the resume to be processed has the format specificity or the candidate category specificity.

[0010] Furthermore, if the resume to be processed only has format specificity, then the main influencing factor of the anomaly category is determined to be format specificity, and the main anomaly category of the resume to be processed is determined to be the first anomaly category. If the resume to be processed only has candidate category specificity, then the main influencing factor of the abnormal category is determined to be candidate category specificity, and the main abnormality category of the resume to be processed is determined to be the second abnormality category.

[0011] Further, obtain the continuous blank area corresponding to the format specificity, and extract the preceding text fragment and the following text fragment of the continuous blank area; The preceding text fragment and the following text fragment are concatenated to determine the semantic fluency score of the concatenated text; If the semantic fluency score is greater than the baseline semantic fluency score, then a layout reconstruction operation is performed to remove the continuous blank area and merge the preceding text fragment and the following text fragment into a continuous text line. If the semantic fluency score is less than or equal to the benchmark semantic fluency score, the resume to be processed is divided into regions, the regional layout is determined, each region is decoupled into independent semantic units, the industry-related entities of each independent semantic unit are extracted, regional logical relationships are established, and regional information analysis is performed based on the logical relationships.

[0012] Furthermore, based on the format specificity, continuous blank areas are located, and the semantic fluency score of each continuous blank area is obtained. Continuous blank areas with semantic fluency scores greater than the benchmark semantic fluency score are used as region boundaries to divide several regions. Construct a two-dimensional coordinate system and extract the left boundary coordinates and region width of each region; When the difference between the left boundary coordinates of several adjacent regions is less than the reference offset, and the width ratio of the two regions is within the preset ratio range, they are determined to be of the parallel type. Alternatively, it can be determined as a primary or secondary type.

[0013] Furthermore, for parallel types, information columns are aligned and extracted; for primary and secondary types, the primary region is determined first, and then the primary region information is extracted by decoupling. Entities in different regions are paired up and input into the model for binary classification of relationships to determine whether there is a logical relationship. If a logical relationship exists, the relationship type is further output. The logically related entity pairs and their relationship types are confirmed as the output of the region logical association results; The association categories are time-series association, attribution association, attribute supplement association, and responsibility decomposition association.

[0014] Furthermore, based on the relationship type of the logical association, entities with attribution association and attribute supplementation association are merged, entities with responsibility decomposition association are combined into responsibility chains, and entities with temporal association are sorted by time. Detect and merge logical relationships across regions to form a complete descriptive unit.

[0015] Furthermore, the competency elements from the work experience are extracted, and these competency elements are matched with nodes in the target job requirement map; When the semantic similarity between the capability element and a node in the demand graph exceeds the baseline similarity, the capability element is labeled as a transferable tag, and the path it maps to the node is recorded.

[0016] Furthermore, the transferable labels are matched to the corresponding graph nodes; The overall matching degree between the transferable label and the target job is determined based on the weight of the matched graph nodes and the relationship strength of other nodes associated with the matched nodes. If the overall matching degree is greater than the preset overall matching degree, then the resume to be processed is determined to be a labeled resume.

[0017] Compared with existing technologies, this invention identifies the main influencing factors of anomaly categories, determines the main anomaly category of the resume to be processed, and analyzes the anomaly salience of the resume for the first anomaly category, performing layout reconstruction and multimodal semantic alignment. The resume is then divided into regions, and regional layout is determined, decoupling each region into independent semantic units, extracting industry-related entities, establishing logical relationships between regions, and performing regional information analysis based on these logical relationships. For the second anomaly category, the work experience of the resume to be processed is decomposed into several transferable tags, constructing a target job requirement map, and performing tag matching analysis. This invention improves the parsing speed and accuracy of resume parsing by performing multiple anomaly category analysis on the resume to be processed and adopting different parsing and extraction methods for different categories.

[0018] In particular, by analyzing the first and second characteristics of the resumes to be processed and identifying the main influencing factors of anomalies, the processing software is categorized. In practice, existing technologies often use a uniform parsing process for resumes with different layouts, especially multi-column, cross-column, and non-continuous text layouts. When encountering resumes with specific formatting characteristics, problems such as text line errors, semantic fragmentation, incomplete information extraction, and even misalignment can easily occur, leading to inaccurate parsing. Furthermore, when processing resumes lacking direct work experience in the target industry, existing technologies often directly determine them as invalid and discard them due to low matching between work experience and job requirements or obvious gaps in the resume's timeline. This results in a large number of candidates with indirect skills and transferable cross-industry abilities being mistakenly screened. Based on this, this invention considers pre-categorizing the resumes to be processed into two categories and adopting targeted analysis methods for each category to improve the efficiency and accuracy of resume analysis.

[0019] In particular, for the first anomalous category, i.e., the resumes to be processed have a first characteristic, further anomalous salience differentiation is performed on this type of resume. Specifically, it is determined whether the format particularity is caused by accidental typesetting noise, such as non-subjective extra spaces, soft returns, page margin jitter, etc., or by a true multi-column, cross-column, or complex layout structure. In practice, when existing technologies detect continuous blank areas or vertical alignment offsets, they often directly determine them as column boundaries, thereby triggering complex layout reconstruction, area division, and semantic unit decoupling processes. This leads to the over-processing of pseudo-columns caused by erroneous input or exported format distortion, incorrectly splitting text lines that should be continuous, resulting in semantic breaks. Furthermore, for primary and secondary types, such as a resume with a short personal information or skill tag in the left column and the main work experience and educational background in the right column, traditional methods often perform entity extraction, relation classification, and semantic fusion indiscriminately on all areas, resulting in the over-analysis of a large amount of low-value information from the areas, reducing the resume analysis speed. Based on this, the present invention considers dividing the first abnormal and special category of resumes into column categories, and selects different parsing methods for parallel types and primary and secondary types, thereby improving the parsing speed and parsing accuracy.

[0020] In particular, for the second anomalous category—that is, resumes exhibiting a second characteristic—transferable tags are identified, and tag matching analysis is performed. This allows for a deeper analysis of the implicit competencies within the resume that are strongly relevant to the target position. In practice, existing resume parsing and talent screening often rely on entity type distribution density, such as the matching degree between job titles, industry keywords, and professional skill terms and the target position's required entities. When the matching degree is lower than the baseline semantic fluency score, the resume is directly judged as irrelevant and filtered out. For candidates transitioning from other industries, those in emerging composite positions, or those with indirect experience, this rigid matching mechanism is highly prone to misscreening. Therefore, this invention considers analyzing the resumes to be processed to identify several transferable tags, and then performing matching analysis between the data corresponding to each tag and the target position's requirement map to improve the efficiency of resume screening. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the steps of the AI-based large model-based intelligent resume parsing and information extraction method according to an embodiment of the present invention; Figure 2 This is a logic block diagram illustrating how to determine whether the resume to be processed falls under the exception handling category in an embodiment of the present invention. Figure 3 This is a logic block diagram for analyzing the salience of anomalies in the resume to be processed according to an embodiment of the present invention; Figure 4 This is a logic block diagram for determining the regional layout in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0023] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] Please see Figure 1 The diagram illustrates the steps of the AI-based large-scale model-based intelligent resume parsing and information extraction method according to an embodiment of the present invention. The AI-based large-scale model-based intelligent resume parsing and information extraction method of the present invention includes: Step S1: Obtain the resume to be processed, and identify the first feature and the second feature in the resume to be processed based on the AI ​​big data model; Step S2: Based on the first feature and the second feature, determine whether the category of the resume to be processed is an exception handling category; Step S3: For the exception handling category of the resume to be processed, determine the main influencing factors of the exception category, so as to determine the main exception specific category of the resume to be processed. Step S4: For the first abnormal specific category, analyze the abnormal salience of the resume to be processed, and perform layout reconstruction and multimodal semantic alignment on the resume to be processed. Alternatively, the resumes to be processed can be divided into regions, regional layout can be determined, each region can be decoupled into an independent semantic unit, industry-related entities of each independent semantic unit can be extracted, regional logical relationships can be established, and regional information analysis can be performed based on the logical relationships. Step S5: For the second abnormal and specific category, decompose the work experience of the resume to be processed into several transferable tags, construct the target job requirement map, and perform tag matching analysis; The first feature is the length of the continuous blank area between adjacent text lines and the vertical alignment offset between columns; the second feature is the entity type distribution density and the continuity of the resume timeline.

[0025] In practice, resumes are often formatted in columns. Two-column or three-column resumes are easily misread by AI models, so special analysis is needed for resumes with specific formats. Furthermore, candidates have specific categories. In reality, candidates may be recent graduates or candidates changing industries. These candidates do not have relevant experience in the current industry. If existing analysis methods for resumes of candidates in the current industry are used, these candidates may be directly eliminated. Therefore, special analysis is needed for candidates with specific categories.

[0026] Specifically, the source of the resume to be processed is not limited. For example, in this embodiment of the invention, it can be an electronic document uploaded by the user through a web page or mobile device. Of course, it can also be from other channels, as long as the user authorizes it and the required data is obtained. This will not be elaborated further.

[0027] Specifically, the AI ​​large model, a current technology, is used to identify the length of consecutive blank areas and the vertical alignment offset between columns, as well as the distribution density of entity types and the continuity of the resume's timeline. In this embodiment of the invention, the AI ​​large model is a multimodal large model, such as LayoutLM, GPT-4V, Qwen-VL, etc., and simultaneously receives image information or layout coordinate information of the resume. The multimodal large model directly outputs the length of consecutive blank areas between adjacent text lines and the vertical alignment offset between columns in an end-to-end manner. Simultaneously, it extracts entity types from the resume and calculates their distribution density through built-in named entity recognition capabilities, and analyzes the resume's timeline and determines its continuity through time information extraction capabilities.

[0028] Specifically, if both format specificity and candidate category specificity exist, the candidate category is selected according to a preset priority. In this embodiment of the invention, when both format specificity and candidate category specificity exist, the candidate is preferentially identified as the first anomalous category, because format issues will cause all subsequent parsing to fail.

[0029] Specifically, if the length of the continuous blank area is greater than the reference blank length or the vertical alignment offset between columns is greater than the reference offset, then there is a format specificity. If the matching degree between the entity type distribution density and the entity type required for the job is lower than the first preset benchmark value, or if the continuity of the resume timeline exceeds the second preset benchmark value interval, then there is candidate category specificity.

[0030] Specifically, the horizontal spacing between adjacent text lines is defined as the length of the continuous blank area. The length of the continuous blank area characterizes the degree of semantic segmentation between adjacent text lines in the resume being processed in the horizontal direction. The longer the continuous blank area, the greater the possibility that the text on the left and right sides is intentionally separated on the page, and the more likely it is to have a multi-column layout structure.

[0031] Specifically, the baseline blank length is predetermined. The historical blank area lengths of several different resume formats are obtained in advance, and the average of the historical blank area lengths is determined as the baseline blank length.

[0032] Specifically, the vertical offset difference between the starting positions of adjacent text lines is determined as the vertical alignment offset between columns. The vertical alignment offset between columns characterizes the vertical alignment accuracy of the potential column structure in the resume to be processed. In a multi-column layout, the text on the left and right sides of the same line should be aligned vertically. When the offset is too large, it indicates that the probability of inter-column misalignment is greater, and the possibility of a multi-column layout is greater.

[0033] Specifically, the baseline offset is predetermined by obtaining the historical offsets of several different resume formats in advance, and determining the average of each historical offset as the baseline offset.

[0034] Specifically, the ratio of the frequency of each entity type appearing in the resume to the total text length is determined as the entity type distribution density. The entity type distribution density characterizes the degree of relevance of the resume to the target job domain in terms of content. The higher the distribution density, the more focused the resume is on a specific domain; the lower the distribution density, the greater the difference between the entity types and the requirements of the target job.

[0035] Specifically, the matching degree is determined by the ratio of the number of target job requirement entity types in the resume to the total number of requirement entity types. The matching degree characterizes the content relevance between the resume and the target job at the entity type level. A higher matching degree indicates that the resume contains more key entity types relevant to the target job, and that the candidate's domain aligns better with the target job. Specifically, the first preset benchmark value is predetermined, and a number of qualified resume sample sets for the target position are obtained in advance, and the average matching degree of each sample is determined to be the first preset benchmark value.

[0036] Specifically, the calculation process for the timeline coherence of a resume is as follows. Extract the start and end dates of all work experiences from the resumes to be processed and arrange them in chronological order; Calculate the number of months between the end time of the previous experience and the start time of the next experience; If any interval exceeds the second baseline semantic fluency score, the continuity of the resume timeline is deemed abnormal. Specifically, the second preset benchmark value is predetermined. Several qualified resume sample sets for the target position are obtained in advance, and the average number of months between each sample is determined as the second preset benchmark value.

[0037] Please see Figure 2 The diagram shown is a logic block diagram for determining whether the resume to be processed belongs to the exception handling category according to an embodiment of the present invention. Specifically, determining whether the resume to be processed belongs to the exception handling category, wherein... If the resume to be processed meets the preset conditions, then the category of the resume to be processed is determined to be the exception handling category; If the resume to be processed does not meet the preset conditions, it is determined that the category of the resume to be processed is not an exception handling category; The preset condition is that the resume to be processed has the format specificity or the candidate category specificity.

[0038] Specifically, if the resume to be processed only has format specificity, then the main influencing factor of the anomaly category is determined to be format specificity, and the main anomaly category of the resume to be processed is determined to be the first anomaly category. If the resume to be processed only has candidate category specificity, then the main influencing factor of the abnormal category is determined to be candidate category specificity, and the main abnormality category of the resume to be processed is determined to be the second abnormality category.

[0039] Specifically, by analyzing the first and second characteristics of the resumes to be processed and determining the main influencing factors of anomalies, the processing software is categorized. In practice, existing technologies often use a uniform parsing process for resumes with different layouts, especially multi-column, cross-column, and non-continuous text layouts. When encountering resumes with specific formatting characteristics, problems such as text line errors, semantic fragmentation, incomplete information extraction, and even misalignment can easily occur, leading to inaccurate parsing. Furthermore, when processing resumes lacking direct work experience in the target industry, existing technologies often directly determine them as invalid and discard them due to low matching between work experience and job requirements or obvious gaps in the resume's timeline. This results in a large number of candidates with indirect skills and cross-industry transferable abilities being mistakenly screened. Based on this, the embodiments of this invention consider pre-categorizing the resumes to be processed into two categories, and adopting targeted analysis methods for each category of resumes to improve the efficiency and accuracy of resume analysis.

[0040] Please see Figure 3 The diagram shown is a logical flowchart for analyzing the salience of anomalies in the resume to be processed according to an embodiment of the present invention. Specifically, it involves obtaining the continuous blank areas corresponding to the format specificity, and extracting the preceding text fragments and following text fragments of the continuous blank areas. The preceding text fragment and the following text fragment are concatenated to determine the semantic fluency score of the concatenated text; If the semantic fluency score is greater than the baseline semantic fluency score, then a layout reconstruction operation is performed to remove the continuous blank area and merge the preceding text fragment and the following text fragment into a continuous text line. If the semantic fluency score is less than or equal to the benchmark semantic fluency score, the resume to be processed is divided into regions, the regional layout is determined, each region is decoupled into independent semantic units, the industry-related entities of each independent semantic unit are extracted, regional logical relationships are established, and regional information analysis is performed based on the logical relationships.

[0041] Specifically, the semantic coherence of the concatenated text is determined by the semantic fluency score. The semantic fluency score characterizes the natural language fluency and logical coherence of the new text formed after directly merging two paragraphs of text with consecutive blank areas. The higher the semantic fluency score, the higher the probability that the two paragraphs belong to the same semantic unit, and the greater the probability that the consecutive blank areas are accidental formatting noise. The lower the semantic fluency score, the higher the probability that there is a significant semantic break between the two paragraphs, and the greater the probability that the consecutive blank areas are real column or section boundaries.

[0042] Specifically, the baseline semantic fluency score is predetermined. Several semantic fluency samples of randomly spliced ​​text in normal single-column resumes are obtained in advance to determine the first historical semantic fluency score. Several semantic fluency samples of cross-column spliced ​​text in real multi-column resumes are obtained to determine the second historical semantic fluency score. The average of the sum of the first historical semantic fluency scores and the second historical semantic fluency scores is determined as the baseline semantic fluency score.

[0043] Please see Figure 4 This is a logical block diagram for determining the regional layout in an embodiment of the present invention. Specifically, based on the format specificity, continuous blank areas are located, the semantic fluency score of each continuous blank area is obtained, and continuous blank areas with semantic fluency scores greater than the benchmark semantic fluency score are used as regional boundaries to divide several regions; Construct a two-dimensional coordinate system and extract the left boundary coordinates and region width of each region; When the difference between the left boundary coordinates of several adjacent regions is less than the reference offset, and the width ratio of the two regions is within the preset ratio range, they are determined to be of the parallel type. Alternatively, it can be determined as a primary or secondary type.

[0044] Specifically, the horizontal axis of the two-dimensional coordinate system represents the horizontal position on the page, and the vertical axis represents the vertical position on the page. In this embodiment, the lower left corner of the page is taken as the origin. The horizontal and vertical axes can uniquely determine the position, width, and height of each area on the page for subsequent layout determination.

[0045] Specifically, the area boundary is the position where the blank area first appears and the position where it ends, that is, the range defined by the start and end coordinates of the continuous blank area on the page. This boundary divides the resume layout into different areas that are adjacent to each other on the left and right or top and bottom, and the text within each area remains relatively continuous in the horizontal or vertical direction.

[0046] Specifically, the width ratio represents the proportional relationship between the widths of two adjacent areas in the horizontal direction. For two areas side by side, the closer the width ratio is to 1, the more similar the widths of the two areas are, and the more likely they are to be arranged symmetrically. The smaller the width ratio is, the more likely one area is wider than the other, and the more likely it is to be arranged in a primary-secondary manner.

[0047] Specifically, the average of the first preset ratio and the second preset ratio is determined as the preset ratio. To determine this in advance, several resumes of parallel types are obtained in advance, and the historical width ratio of each is determined as the first preset ratio. Several resumes of primary and secondary types are obtained in advance, and the historical width ratio of each is determined as the second preset ratio.

[0048] Specifically, for the first abnormal category, i.e., the resume to be processed has a first characteristic, further distinguish the abnormality of this type of resume. Specifically, determine whether the format particularity is caused by accidental typesetting noise, such as non-subjective extra spaces, soft returns, page margin jitter, etc., or by the formation of a real multi-column, cross-column or complex layout structure. In practice, when existing technology detects continuous blank areas or vertical alignment offsets, it often directly judges them as column boundaries, thereby triggering a complex layout reconstruction, area division and semantic unit decoupling process. This leads to the over-processing of pseudo-columns caused by erroneous input or exported format distortion, which incorrectly splits the text lines that should be continuous, causing semantic breaks. Furthermore, for primary and secondary types, such as the left column of the resume containing short personal information or skill tags, and the right column containing the main work experience and educational background, traditional methods often perform entity extraction, relation classification and semantic fusion indiscriminately on all areas, resulting in the over-analysis of a large amount of low-value information from the areas, reducing the resume analysis speed. Based on this, embodiments of the present invention consider dividing the first abnormal and special category of resumes into column categories, and selecting different parsing methods for parallel types and primary and secondary types, thereby improving the parsing speed and parsing accuracy.

[0049] Specifically, for parallel types, information columns are aligned and extracted; for primary and secondary types, the primary region is determined first, and then the primary region information is decoupled and extracted. Entities in different regions are paired up and input into the model for binary classification of relationships to determine whether there is a logical relationship. If a logical relationship exists, the relationship type is further output. The logically related entity pairs and their relationship types are confirmed as the output of the region logical association results; The association categories are time-series association, attribution association, attribute supplement association, and responsibility decomposition association.

[0050] Specifically, information column alignment extraction refers to aligning semantically related items in different sections of a parallel structure with similar layout and status. For example, in a two-column resume, the left column describes work experience and the right column describes project achievements. This alignment, based on row numbers or vertical position, allows content on the same logical line to be correctly linked. For instance, if the third line of the left column describes a job responsibility and the third line of the right column describes the corresponding technical achievements, alignment extraction can merge them into a single complete descriptive unit.

[0051] Specifically, the widths of two adjacent regions are compared; the region with the larger width is identified as the primary region, and the region with the smaller width is identified as the secondary region. When extracting information, the primary region is given priority for complete semantic parsing, while the secondary regions undergo lightweight extraction. This differentiated processing reduces unnecessary complex parsing of secondary regions within the primary-secondary type framework, thus improving analysis efficiency.

[0052] Specifically, the entity pairs and their respective context texts are concatenated, with an additional prompt asking the user to determine whether the following two entities have a logical relationship. The output is either yes or no, and the classification result can be output based on a large AI model.

[0053] Specifically, temporal correlation indicates that there is a chronological order or simultaneous occurrence between two entities. For example, there is a temporal correlation between the product manager position held from 2020 to 2022 in the left column and the product iteration project led in 2021 in the right column.

[0054] Affiliation means that one entity belongs to another entity. For example, there is an affiliation between XX Technology Co., Ltd. in the left column and the entity responsible for the research and development of the company's core products in the right column.

[0055] Attribute supplementary associations indicate that one entity provides supplementary information about another entity in terms of attribute dimensions. For example, there is an attribute supplementary association between "Java Developer" in the primary region and "Proficient in Spring and MyBatis frameworks" in the secondary region.

[0056] A responsibility breakdown relationship indicates that one entity is a detailed breakdown of a specific responsibility within another entity. For example, there is a responsibility breakdown relationship between those responsible for e-commerce platform construction and those responsible for order module development and payment interface integration.

[0057] Specifically, when there is a hierarchical relationship or a complementary attribute relationship between two entities, the two entities are merged into a complete entity description with attribute information. For example, during resume parsing, the entity "Backend Development Engineer" is extracted from the main area, and the entities "Java," "Spring Boot," and "MySQL" are extracted from the sub-areas. If a complementary attribute relationship is determined between the two during parsing, the skill entities are merged as attributes of the job entity to form a structured description: Job: Backend Development Engineer, Skills: Java, Spring Boot, MySQL.

[0058] Specifically, when there is a responsibility decomposition relationship between multiple entities, these entities are combined into a responsibility chain according to the hierarchical relationship from general to specific, forming a structured responsibility description.

[0059] For example, the resume contains the following entity descriptions in different areas or within the same area. Entity A: Responsible for the overall construction of the e-commerce platform. Entity B: Order module development, Entity C: Payment interface integration. Entity D: User access control During the parsing process, it is determined that entities B, C, and D are all related to entity A through responsibility decomposition. These entities are then combined into a responsibility chain, and the output is: Main responsibility: responsible for the overall construction of the e-commerce platform; Specific tasks: order module development, payment interface integration, and user permission management.

[0060] Specifically, when multiple entities are temporally related, these entities are sorted according to their chronological order to form a timeline-structured description. For example, the following entities with time information can be extracted from a resume: Entity A: From 2018 to 2020, served as a junior engineer. Entity B: Served as an intermediate engineer from 2020 to 2022. Entity C: Since 2022, has served as a senior engineer. During the parsing process, it was determined that there was a temporal relationship among the three. Therefore, the output after sorting by time is as follows: Time period: 2018-2020, Position: Junior Engineer; Time period: 2020-2022, Position: Intermediate Engineer; Time period: 2022 to present, Position: Senior Engineer.

[0061] Specifically, after completing the merging, combination, and sorting of the above-mentioned associations, the logical associations that cross different regional boundaries are further detected, and entities that are scattered in multiple regions but belong to the same semantic unit are merged to form a complete description unit.

[0062] Specifically, the competency elements in the work experience are extracted and matched with nodes in the target job requirements map; When the semantic similarity between the capability element and a node in the demand graph exceeds the baseline similarity, the capability element is labeled as a transferable tag, and the path it maps to the node is recorded.

[0063] Specifically, competency elements refer to the underlying competencies that reflect a candidate's ability to be transferred to other positions or industries, including but not limited to: management skills, technical skills, and business skills.

[0064] Specifically, the method for constructing the target job requirement map is as follows: Obtain text data such as job descriptions, job responsibilities, and job requirements for the target position; Perform entity and relation extraction to generate an initial graph; Based on the calibration of node weights by experts or recruiters, a target job demand map is obtained.

[0065] Specifically, there are no restrictions on the method for calculating semantic similarity. For example, in practice, it can be calculated using the cosine similarity method. Of course, those skilled in the art can also use other methods, as long as the required data can be obtained. This will not be elaborated further.

[0066] Specifically, the baseline similarity is predetermined and used to determine whether the semantic matching between the capability element and the capability node in the target job requirement map is effective. In this embodiment of the invention, several sets of historical similarities of known matching relationships are obtained in advance, and the average of each historical similarity is determined as the baseline similarity.

[0067] Specifically, the transferable label is matched to the corresponding graph node; The overall matching degree between the transferable label and the target job is determined based on the weight of the matched graph nodes and the relationship strength of other nodes associated with the matched nodes. If the overall matching degree is greater than the preset overall matching degree, then the resume to be processed is determined to be a labeled resume.

[0068] Specifically, for the second anomalous category—that is, resumes exhibiting a second characteristic—transferable tags are identified, and tag matching analysis is performed. This allows for in-depth mining of implicit competencies strongly relevant to the target position. In practice, existing resume parsing and talent screening often rely on entity type distribution density, such as the matching degree between job titles, industry keywords, and professional skill terms and the target position's required entities. When the matching degree is lower than the baseline semantic fluency score, the resume is directly judged as irrelevant and filtered out. For candidates transitioning from cross-industry roles, emerging composite positions, or those with indirect experience, this rigid matching mechanism is highly prone to misscreening. Therefore, this invention considers analyzing resumes to identify several transferable tags and performing matching analysis between the data corresponding to each tag and the target position's requirement map to improve resume screening efficiency.

[0069] Understandably, existing technologies rely solely on literal matching of explicit keywords in resumes with job requirements, ignoring the inherent connections between different skills, which leads to incorrect filtering of resumes.

[0070] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent resume parsing and information extraction based on an AI large-scale model, characterized in that, include: Obtain the resume to be processed, and identify the first and second features in the resume based on the AI ​​large model; Based on the first feature and the second feature, determine whether the resume to be processed belongs to the abnormal handling category; For the anomaly handling category of the resume to be processed, determine the main influencing factors of the anomaly category, and use them to determine the main anomaly specific category of the resume to be processed; For the first anomaly specific category, the anomaly salience of the resume to be processed is analyzed, and the resume to be processed is restructured in layout and aligned with multimodal semantics. Alternatively, the resumes to be processed can be divided into regions, regional layout can be determined, each region can be decoupled into an independent semantic unit, industry-related entities of each independent semantic unit can be extracted, regional logical relationships can be established, and regional information analysis can be performed based on the logical relationships. For the second abnormal and specific category, the work experience of the resume to be processed is decomposed into several transferable tags, a target job requirement map is constructed, and tag matching analysis is performed; The first feature is the length of the continuous blank area between adjacent text lines and the vertical alignment offset between columns; the second feature is the entity type distribution density and the continuity of the resume timeline.

2. The method for intelligent resume parsing and information extraction based on an AI large model according to claim 1, characterized in that, If the length of the continuous blank area is greater than the reference blank length or the vertical alignment offset between columns is greater than the reference offset, then there is a format specificity. If the matching degree between the entity type distribution density and the entity type required for the job is lower than the first preset benchmark value, or if the continuity of the resume timeline exceeds the second preset benchmark value interval, then there is candidate category specificity.

3. The method for intelligent resume parsing and information extraction based on an AI large model according to claim 1, characterized in that, The determination of whether the resume to be processed falls under the exception handling category, wherein, If the resume to be processed meets the preset conditions, then the category of the resume to be processed is determined to be the exception handling category; If the resume to be processed does not meet the preset conditions, it is determined that the category of the resume to be processed is not an exception handling category; The preset condition is that the resume to be processed has the format specificity or the candidate category specificity.

4. The method for intelligent resume parsing and information extraction based on an AI large model according to claim 1, characterized in that, If the resume to be processed only has format specificity, then the main influencing factor of the anomaly category is determined to be format specificity, and the main anomaly category of the resume to be processed is determined to be the first anomaly category. If the resume to be processed only has candidate category specificity, then the main influencing factor of the abnormal category is determined to be candidate category specificity, and the main abnormality category of the resume to be processed is determined to be the second abnormality category.

5. The method for intelligent resume parsing and information extraction based on an AI large model according to claim 1, characterized in that, Obtain the continuous blank area corresponding to the format specificity, and extract the preceding text fragment and the following text fragment of the continuous blank area. The preceding text fragment and the following text fragment are concatenated to determine the semantic fluency score of the concatenated text; If the semantic fluency score is greater than the baseline semantic fluency score, then a layout reconstruction operation is performed to remove the continuous blank area and merge the preceding text fragment and the following text fragment into a continuous text line. If the semantic fluency score is less than or equal to the benchmark semantic fluency score, the resume to be processed is divided into regions, the regional layout is determined, each region is decoupled into independent semantic units, the industry-related entities of each independent semantic unit are extracted, regional logical relationships are established, and regional information analysis is performed based on the logical relationships.

6. The method for intelligent resume parsing and information extraction based on an AI large model according to claim 5, characterized in that, Based on the format specificity, locate continuous blank areas, obtain the semantic fluency score of each continuous blank area, and use the continuous blank areas with semantic fluency scores greater than the benchmark semantic fluency score as the area boundary to divide several areas. Construct a two-dimensional coordinate system and extract the left boundary coordinates and region width of each region; When the difference between the left boundary coordinates of several adjacent regions is less than the reference offset, and the width ratio of the two regions is within the preset ratio range, they are determined to be of the parallel type. Alternatively, it can be determined as a primary or secondary type.

7. The method for intelligent resume parsing and information extraction based on an AI large model according to claim 6, characterized in that, For parallel types, information columns are aligned and extracted; for primary and secondary types, the primary region is determined first, and then the primary region information is extracted by decoupling. Entities in different regions are paired up and input into the model for binary classification of relationships to determine whether there is a logical relationship. If a logical relationship exists, the relationship type is further output. The logically related entity pairs and their relationship types are confirmed as the output of the region logical association results; The association categories are time-series association, attribution association, attribute supplement association, and responsibility decomposition association.

8. The method for intelligent resume parsing and information extraction based on an AI large model according to claim 7, characterized in that, Based on the relationship type of the logical association, entities with attribution association and attribute supplementation association are merged, entities with responsibility decomposition association are combined into responsibility chain, and entities with temporal association are sorted by time. Detect and merge logical relationships across regions to form a complete descriptive unit.

9. The method for intelligent resume parsing and information extraction based on an AI large model according to claim 1, characterized in that, Extract the competency elements from the work experience and match the competency elements with nodes in the target job requirements map; When the semantic similarity between the capability element and a node in the demand graph exceeds the baseline similarity, the capability element is labeled as a transferable tag, and the path it maps to the node is recorded.

10. The method for intelligent resume parsing and information extraction based on an AI large model according to claim 1, characterized in that, Match the transferable label to the corresponding graph node; The overall matching degree between the transferable label and the target job is determined based on the weight of the matched graph nodes and the relationship strength of other nodes associated with the matched nodes. If the overall matching degree is greater than the preset overall matching degree, then the resume to be processed is determined to be a labeled resume.

Images