A resume intelligent retrieval method and system based on a large language model

Abstract

The application provides a resume intelligent retrieval method and system based on a large language model, which comprises: performing format analysis on an unstructured resume document to obtain a resume text; performing structured information extraction on the resume text and an unstructured post demand description text based on a large language model to obtain structured personnel portrait labels and structured demand portrait labels; establishing label alignment mapping between the structured personnel portrait labels and the structured demand portrait labels, and performing matching based on the label alignment mapping. The application performs multi-path recall through inverted index matching and vector similarity matching, and performs comprehensive sorting in combination with statistical information, matching rules and a matching model to obtain a recommended resume list, thereby reducing matching omissions caused by different word expressions, improving the accuracy and interpretability of the person-post matching result; meanwhile, the retrieval strategy is dynamically adjusted according to user natural language feedback, improving the adaptability and efficiency of resume retrieval.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and information retrieval technology, and in particular to a resume intelligent retrieval method and system based on a large language model. Background Technology

[0002] With the increasing digitalization of the human resources service industry, resume retrieval and job matching systems are widely used in corporate recruitment, talent recommendation, and headhunting services. Existing systems typically parse resumes and job descriptions to extract personnel information and job requirements, and then retrieve and sort candidates using keywords, tags, filtering criteria, or text similarity. However, existing technologies still have the following shortcomings.

[0003] (1) Insufficient ability to structure resumes and job descriptions and match tags.

[0004] Resumes and job descriptions are typically written in natural language, and different documents vary in format, layout, field names, and information expression. For complex layouts, image-based documents, or non-standard format information, existing parsing methods are prone to issues such as missing information, incomplete field recognition, or inconsistent expression of similar information. Furthermore, the resume side and the job requirements side may use different tags or classification methods. For synonymous or similar information, different tag expressions can easily lead to incorrect matching, resulting in missed opportunities to find suitable candidates.

[0005] (2) The existing job matching method is difficult to form a detailed sort when taking into account different types of matching information.

[0006] Existing systems typically employ single or relatively fixed matching methods such as keywords, tags, or text similarity. Relying solely on rule or field matching makes it difficult to identify semantic connections between different textual expressions; relying solely on text or vector similarity fails to adequately reflect hard constraints such as education level, years of work experience, and skill requirements. Therefore, when multiple matching factors are involved, existing methods struggle to comprehensively process different matching information and generate candidate ranking results that reflect all matching factors.

[0007] (3) Existing search methods are difficult to dynamically adjust based on search results and user feedback.

[0008] Current search processes are typically based on preset search strategies. When search results do not meet user needs, or when multiple conditions combine to result in empty or few search results, users usually need to modify their search strategy and search again. The current search results and user feedback are difficult to use directly in subsequent search processes, leading to repeated manual adjustments to search strategies.

[0009] In summary, existing intelligent resume retrieval technologies lack the ability to process unstructured resumes and structure job requirements using appropriate tags, and current job-person matching and retrieval methods struggle to dynamically adjust to various matching information and user feedback. Therefore, it is necessary to provide an intelligent resume retrieval method and system based on a large language model to address the aforementioned problems in existing technologies. Summary of the Invention

[0010] In view of the shortcomings of the prior art described above, the purpose of this application is to provide a resume intelligent retrieval method and system based on a large language model, which can solve the problems of insufficient structured processing capability of unstructured resumes and job requirement descriptions, difficulty in effectively corresponding between personnel profile tags and requirement profile tags, single retrieval interaction method, and lack of interpretability of matching results in the prior art.

[0011] To achieve the above and other related objectives, the first aspect of this application provides an intelligent resume retrieval method based on a large language model, comprising: parsing the format of an unstructured resume document to obtain resume text; extracting structured information from the resume text and the unstructured job requirement description text based on a large language model to obtain structured personnel profile tags and structured requirement profile tags; establishing a tag alignment mapping between the structured personnel profile tags and the structured requirement profile tags based on a preset matching rule; matching the structured personnel profile tags and the structured requirement profile tags based on the tag alignment mapping to obtain a matching result; determining a set of candidates based on the matching result, calculating the total matching score for each candidate in the set, sorting the candidates according to the total matching score, and outputting a recommended resume list.

[0012] In some embodiments of the first aspect of this application, the generation of structured personnel profile tags includes parsing the unstructured text of a resume and generating structured tags. Addressing the issue of low tag coverage in existing technologies, the personnel profile tag generation module employs a large language model-based enhanced personnel profile tag generation method, and processes the resume according to a three-stage processing pipeline of "rule preprocessing, LLM semantic extraction, and post-validation." Specifically, the resume text is obtained by parsing the format and restoring the text of the unstructured resume document; structured information is extracted from the resume text based on the large language model to obtain structured personnel profile tags; and rule post-validation and semantic normalization are performed on the structured personnel profile tags to update them.

[0013] In some embodiments of the first aspect of this application, based on the situation where there are different textual expressions between personnel profile tags and demand profile tags, a tag alignment mapping between structured personnel profile tags and structured demand profile tags is established to determine the correspondence between personnel profile tags and demand profile tags.

[0014] In some embodiments of the first aspect of this application, the structured personnel profile tags and structured demand profile tags are matched based on the tag alignment mapping to obtain a matching result, and a candidate set is determined based on the matching result. The candidates in the candidate set are then sorted, and a recommended resume list is output.

[0015] In some embodiments of the first aspect of this application, the intelligent matching recommendation module determines the total matching score based on the matching information corresponding to each candidate in the candidate set, sorts the candidates according to the total matching score, and forms a structured recommendation result containing matching information when outputting the recommended resume list.

[0016] In some embodiments of the first aspect of this application, a large-model intelligent retrieval agent is used as the interaction entry point between the user and the intelligent resume retrieval system. This agent obtains the user's input natural language retrieval request, determines a retrieval strategy based on the natural language retrieval request, and provides the retrieval strategy to the intelligent matching and recommendation module. Further, the large-model intelligent retrieval agent module obtains the user's natural language feedback on the recommended resume list, and updates at least one of the retrieval strategy and matching weight based on the natural language feedback to re-execute the matching process and update the recommended resume list.

[0017] In some embodiments of the first aspect of this application, the large model intelligent retrieval agent module is also used to generate search condition adjustment suggestions based on the current search strategy when a list of recommended resumes is not obtained, so as to assist the user in adjusting the search conditions.

[0018] To achieve the above and other related objectives, a second aspect of this application provides a resume intelligent retrieval system based on a large language model, comprising: a personnel profile tag generation module, a requirement profile tag generation module, an intelligent matching and recommendation module, and a large model intelligent retrieval agent module.

[0019] The personnel profile tag generation module is used to parse the format of unstructured resume documents to obtain resume text, and extract structured information from the resume text based on a large language model to obtain structured personnel profile tags.

[0020] The requirement profile tag generation module is used to extract structured information from unstructured job requirement description text based on a large language model to obtain structured requirement profile tags, and to establish a tag alignment mapping between the structured personnel profile tags and the structured requirement profile tags based on preset comparison rules.

[0021] The large-scale intelligent retrieval agent module is used to obtain the natural language retrieval request input by the user, determine the retrieval strategy based on the natural language retrieval request, and provide the retrieval strategy to the intelligent matching recommendation module; it is also used to obtain the natural language feedback input by the user for the current recommended resume list, update at least one of the retrieval strategy and matching weight based on the natural language feedback, and provide the updated result to the intelligent matching recommendation module.

[0022] The intelligent matching and recommendation module is used to match the structured personnel profile tags with the structured demand profile tags based on the retrieval strategy and the tag alignment mapping to obtain matching results; determine a set of candidates based on the matching results; calculate the total matching score for each candidate in the set; sort the candidates according to the total matching score; and output a list of recommended resumes.

[0023] To achieve the above and other related objectives, a third aspect of this application provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the intelligent resume retrieval method based on a large language model as described in any of the above embodiments.

[0024] To achieve the above and other related objectives, a fourth aspect of this application provides a computer program product comprising computer program code, which, when executed on a computer, enables the computer to implement the intelligent resume retrieval method based on a large language model as described in any of the above embodiments.

[0025] To achieve the above and other related objectives, a fifth aspect of this application provides an electronic terminal, including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the method.

[0026] As described above, the intelligent resume retrieval method and system based on a large language model in this application has the following beneficial effects: By extracting structured information from resume text and job requirement description text respectively based on a large language model, structured personnel profile tags and structured requirement profile tags are generated, improving the information processing capability of unstructured text; by establishing a tag alignment mapping between structured personnel profile tags and structured requirement profile tags, the tag inconsistency problem caused by different text expressions is reduced, improving the matching ability between personnel profile tags and requirement profile tags; by adopting multi-path recall and comprehensive ranking methods to determine the candidate matching degree, and combining statistical information, preset matching rules and matching models for comprehensive ranking, the accuracy and interpretability of recommendation results are improved; by receiving user natural language retrieval requests through the large model intelligent retrieval agent module, and dynamically adjusting the retrieval strategy and matching weight according to user feedback, the retrieval process adjustment based on natural language interaction is realized, improving the flexibility of the retrieval process; by generating retrieval strategy adjustment suggestions when no recommended resume list is obtained, users can optimize retrieval conditions according to the adjustment suggestions, improving the retrieval interaction experience. Attached Figure Description

[0027] Figure 1 A flowchart illustrating the intelligent resume retrieval method based on a large language model provided in this application embodiment;

[0028] Figure 2 A schematic diagram of the architecture of a resume intelligent retrieval system based on a large language model provided in an embodiment of this application;

[0029] Figure 3 This is a schematic diagram of the tag generation and tag data management structure provided in an embodiment of this application;

[0030] Figure 4 A flowchart illustrating the label alignment and mapping construction process provided in this application embodiment;

[0031] Figure 5 A schematic block diagram of an electronic terminal provided in an embodiment of this application. Detailed Implementation

[0032] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0033] In the embodiments of this application, terms such as "first" and "second" are used to distinguish similar or corresponding technical features. Those skilled in the art will understand that terms such as "first" and "second" do not limit the quantity or execution order, and that terms such as "first" and "second" do not necessarily imply that there is a difference between corresponding objects.

[0034] It should be noted that, in the embodiments of this application, the words "exemplary" or "for example" indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0035] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following associated objects have an "or" relationship.

[0036] Before providing a further detailed description of this application, the nouns and terms used in the embodiments of this application are explained, and the nouns and terms used in the embodiments of this application shall be interpreted as follows:

[0037] <1> Large Language Model (LLM)

[0038] A large language model refers to a model trained on large-scale data for natural language understanding and generation. In this embodiment, a large language model is used to extract structured information from resume text and job requirement description text to generate structured personnel profile tags and structured requirement profile tags.

[0039] <2> Structured personnel profile tags

[0040] Structured personnel profile tags refer to a set of tags used to describe candidate information obtained by extracting structured information from unstructured resume text using a large language model. In this embodiment, structured personnel profile tags are used to represent personnel-related information extracted from the resume.

[0041] <3> Structured demand profile tags

[0042] Structured requirement profile tags refer to a set of information tags used to describe job requirements, obtained by extracting structured information from unstructured job requirement description text using a large language model. In this embodiment, structured requirement profile tags are used to represent job requirement information in the job requirement description text.

[0043] <4> Label alignment mapping

[0044] Tag alignment mapping refers to a processing method that establishes a correspondence between structured personnel profile tags and structured demand profile tags based on preset matching rules. In this embodiment, the correspondence is determined through a thesaurus of personnel and demand tags and a tag hierarchy diagram to reduce inconsistencies between personnel profile tags and demand profile tags caused by different textual expressions.

[0045] <5> Large Model Intelligent Retrieval Agent Module

[0046] The large-scale intelligent retrieval agent module refers to a functional module that implements user interaction and retrieval strategy adjustment based on a large language model. In this embodiment, the large-scale intelligent retrieval agent module is used to obtain the natural language retrieval request input by the user, determine the retrieval strategy based on the natural language retrieval request, and adjust the retrieval strategy based on user feedback.

[0047] In some embodiments of this application, the intelligent resume retrieval method based on a large language model provided in this application can be applied to scenarios requiring resume retrieval, job requirement analysis, and personnel-job matching, such as corporate recruitment, talent screening, and human resource services. The following describes the intelligent resume retrieval method based on a large language model provided in the embodiments of this application with specific method steps.

[0048] The intelligent resume retrieval method based on a large language model provided in this application includes the following steps:

[0049] Step S11: Parse the unstructured resume document to obtain the resume text; extract structured information from the resume text and the unstructured job requirement description text based on the large language model to obtain structured personnel profile tags and structured requirement profile tags.

[0050] In some embodiments, personnel profile tag generation employs a three-stage processing pipeline of "rule preprocessing, LLM semantic extraction, and post-validation." First, unstructured resume documents are processed using a multi-format parsing adapter, which incorporates PDF and Word document parsing engines to convert unstructured resume documents of the corresponding formats into plain text streams. For image-based resumes, OCR (Optical Character Recognition) technology is used for text reconstruction. Subsequently, LLM extracts structured information from the obtained text stream and guides the LLM to output structured personnel profile tags using a dedicated prompt word engineering template. These structured personnel profile tags include current city, desired city, age, education, gender, current industry, current job function, current salary, and language proficiency. Education includes school, major, and degree, and the current salary is standardized to a monthly salary value.

[0051] In some embodiments, during the conversion of unstructured resume documents into text streams, the extraction results can be preprocessed according to the resume document's layout structure. For example, for resume documents with a two-column or multi-column layout, different text blocks in the document can be identified, and the corresponding reading order can be determined based on the positional relationship of each text block on the page, thus rearranging the extracted text content. Simultaneously, headers, footers, table boundaries, or other layout information unrelated to the generation of personnel profile tags can be identified and removed to reduce the impact of complex layout on subsequent structured information extraction. As an example, a PDF document parsing tool can be used to extract text from PDF format resumes, and the above-mentioned text block identification and reading order adjustment can be completed in conjunction with page layout information.

[0052] In some implementations, after obtaining the structured personnel profile tags, the method further includes: performing rule-based post-validation and semantic normalization on the structured personnel profile tags; validating the fields in the structured personnel profile tags according to preset validation rules; standardizing and mapping fields with semantic ambiguity; and updating the structured personnel profile tags based on the results of field validation and standardization mapping.

[0053] Specifically, after obtaining the structured personnel profile tags, rule-based post-validation is performed on the tags. A hard-condition regular expression validation layer is set after the LLM output. This layer includes a validator based on regular expressions and business rules, using the corresponding regular expressions and business rules as preset validation rules to validate the fields in the structured personnel profile tags. For example, the age field is validated to verify whether the age is within a reasonable range of 18 to 65 years old; the salary field is validated to verify whether the salary eligibility is valid. When a field validation fails, the LLM retry mechanism can be triggered, or the corresponding field can be marked as "awaiting manual confirmation".

[0054] Furthermore, in some embodiments, the confidence level of each field in the structured personnel profile labels extracted from the large language model can be evaluated, and whether further verification is needed can be determined based on the confidence level of each field. When the confidence level of a target field is lower than a preset confidence threshold, the target field is marked as a field to be confirmed, and a manual review entry is provided so that users can confirm or correct the field to be confirmed; the corresponding structured personnel profile labels are updated based on the manual review results. By combining the model extraction results, the post-rule verification results, and the field confidence levels, labeling errors caused by ambiguity in unstructured resumes or uncertainty in model extraction can be reduced.

[0055] Building upon the completed field validation, further semantic normalization is performed on fields in the structured personnel profile tags that exhibit semantic ambiguity or differ in wording. This involves standardizing the mapping according to the semantics expressed by the corresponding fields. For example, for the salary field, the contextual understanding capabilities of LLM are used to semantically normalize salary information expressed in different forms, such as "50,000 yuan per month" and "600,000 yuan per year," and the current salary is standardized to a monthly salary value. For the current job title field, different wordings, such as "Java developer" and "backend engineer," are aligned to the job title. Through this standardized mapping, fields expressed in different wordings but with corresponding semantics are transformed into a unified structured representation.

[0056] Based on the above field validation results and standardized mapping results, the corresponding structured personnel profile tags are updated, thereby completing the rule-based post-validation and semantic normalization processing of the structured personnel profile tags. By combining rule validation with the contextual understanding capabilities of LLM, the differences in tags caused by field anomalies and different textual expressions can be reduced, improving the accuracy and tag coverage of structured personnel profile tags.

[0057] In other embodiments, for the aforementioned structured information extraction process, a fine-tuned, small-parameter dedicated model can be used for field extraction to obtain field results and field confidence levels. Based on the field confidence levels and rule-based post-validation results, normal fields, abnormal fields, or low-confidence fields are determined. Normal fields undergo subsequent standardization mapping, while abnormal or low-confidence fields are re-extracted or corrected by a large language model. For example, a resume might state: "Joined Company A in 2019, joined Company B three years later." The small-parameter dedicated model might not accurately identify the temporal relationship between "three years later" and "2019," resulting in a deviation in the time field. In this case, the abnormal field can be identified based on time logic validation rules, and the relevant local text, necessary contextual information, and field results can be sent to the large language model for re-extraction or correction, without reprocessing the entire resume text. In some embodiments, when extracting structured information from job requirement description text, the job requirement description text in natural language form input by the recruiter can be input into a large language model. A large language model information extraction framework corresponding to the personnel profile tags is then used to semantically analyze the job requirements contained in the job requirement description text to obtain structured requirement profile tags. The job requirements may include hard requirements and soft requirements. Hard requirements may include, for example, education level, years of work experience, city, age, job grade, and explicitly required skills. Soft requirements may include, for example, skills, project experience, and other ability requirements reflected in the job description.

[0058] Specifically, the large language model performs contextual semantic understanding on the job requirement description text, identifies the fields used to describe recruitment conditions and their corresponding values, and structures the identification results according to a pre-defined requirement profile tag structure. The recruitment conditions may include educational requirements, years of work experience, work city, core skills, project experience, and job level. Through this processing, the recruitment conditions, originally existing in natural language form in the job requirement description text, are transformed into structured requirement profile tags that can participate in subsequent tag alignment, candidate recall, and matching degree calculation.

[0059] As one specific implementation method, when the job description text entered by the recruiter is "Senior Java backend engineer, 3+ years of experience, Bachelor's degree or above, familiar with Spring Cloud microservices, practical experience in distributed systems, located in City A", the structured information of the job description text can be extracted based on the large language model to obtain structured requirement profile tags including education level no less than Bachelor's degree, work experience no less than 3 years, city A, core skills including Java, Spring Cloud and distributed systems, and job level senior.

[0060] Therefore, job requirement description texts composed of multiple natural language phrases, skill descriptions, and job conditions can be uniformly semantically parsed and structured using a large language model. This allows different types of job requirements to be represented in the form of structured tags, which serve as the basis for subsequent tag alignment between personnel profile tags and requirement profile tags, as well as for candidate screening.

[0061] Step S12: Based on preset comparison rules, establish a tag alignment mapping between the structured personnel profile tags and the structured demand profile tags; the preset comparison rules are used to determine the correspondence between the structured personnel profile tags and the structured demand profile tags.

[0062] In some implementations, the preset matching rules are determined based on a thesaurus of synonyms for personnel and demand tags and a tag hierarchy graph. The thesaurus of synonyms for personnel and demand tags is used to determine the synonym tag correspondence between the structured personnel profile tags and the structured demand profile tags. The tag hierarchy graph is used to determine the hierarchical tag correspondence between the structured personnel profile tags and the structured demand profile tags. According to the synonym tag correspondence and the hierarchical tag correspondence, a tag correspondence between the structured personnel profile tags and the structured demand profile tags is established, and the tag alignment mapping is established according to the tag correspondence.

[0063] like Figure 4 As shown, in some embodiments, after obtaining structured personnel profile tags and structured requirement profile tags, a tag alignment mapping between the structured personnel profile tags and the structured requirement profile tags is established based on preset matching rules. The preset matching rules are used to determine the correspondence between personnel profile-side tags and requirement profile-side tags. In this embodiment, the preset matching rules are determined based on a pre-established thesaurus of personnel and requirement tags and a tag hierarchy diagram. The thesaurus of personnel and requirement tags is used to determine the synonymous tag correspondence between personnel profile-side tags and requirement profile-side tags, and the tag hierarchy diagram is used to determine the hierarchical tag correspondence between personnel profile-side tags and requirement profile-side tags. For example, based on the thesaurus of personnel and requirement tags and the tag hierarchy diagram, synonym matching and hierarchical matching can be performed on personnel profile tags and requirement profile tags, and a tag correspondence can be established based on the matching results, thereby forming, for example, a tag alignment mapping relationship between personnel tag A and requirement tag A′.

[0064] Specifically, for the personnel and requirement tag synonym library, different tag expressions with the same or corresponding semantics can be pre-defined to determine the synonym tag correspondence between structured personnel profile tags and structured requirement profile tags. For example, for job function tags, tags with the same or corresponding semantics such as "backend developer," "server engineer," and "server developer" can be pre-defined in the personnel and requirement tag synonym library. When structured personnel profile tags and structured requirement profile tags use different textual forms to represent the same or corresponding job function information, the corresponding synonym tag correspondence can be determined according to the personnel and requirement tag synonym library, so that different textual expressions can correspond to the corresponding tag semantics, reducing the tag inconsistency caused by different textual expressions used in personnel resumes and job requirements.

[0065] For the aforementioned tag hierarchy diagram, the hierarchical correspondence between different tags with hierarchical relationships can be pre-recorded to determine the hierarchical tag correspondence between structured personnel profile tags and structured requirement profile tags. For example, when the requirement profile side uses relatively general higher-level tags, while the personnel profile side uses relatively specific lower-level tags, the hierarchical association between the two can be determined according to the tag hierarchy diagram, enabling the relatively general tags on the requirement profile side to establish a correspondence with the corresponding specific tags on the personnel profile side.

[0066] After determining the synonym tag correspondence and the hierarchical tag correspondence respectively, the tag correspondence between structured personnel profile tags and structured requirement profile tags is determined based on the synonym tag correspondence and the hierarchical tag correspondence. For the same structured personnel profile tag or structured requirement profile tag, the synonym correspondence and / or hierarchical correspondence with the other tag can be determined based on its correspondence in the personnel and requirement tag synonym database and the tag hierarchy relationship diagram, and the corresponding tag correspondence is formed accordingly.

[0067] Furthermore, a tag alignment mapping is established between structured personnel profile tags and structured demand profile tags based on the aforementioned tag correspondence. In other words, when there are differences between the tag expressions on the demand profile side and the tag expressions on the personnel profile side, the association between the two can be determined based on preset tag synonym relationships and hierarchical relationships, thus establishing a mapping between personnel profile tags that use different word expressions or are at different tag levels but have corresponding semantics and the demand profile tags.

[0068] For example, when "server-side development" appears in the job description, the synonymous or corresponding semantic relationship between "server-side development" and the corresponding job function tag can be determined based on the aforementioned personnel and requirement tag thesaurus. This, combined with the tag hierarchy diagram, establishes a corresponding relationship between "server-side development" and the "Java / Go / C++ backend" tag in the personnel profile, achieving automatic association between requirement profile tags and personnel profile tags. Through this tag alignment mapping, semantic differences arising from different word expressions or tag levels between the personnel profile side and the job requirement side can be addressed, reducing semantic differences between supply and demand tag spaces.

[0069] In some embodiments, the personnel and demand tag thesaurus and tag hierarchy diagram can be pre-established. Specifically, for tags with the same or corresponding semantics but expressed in different textual forms, corresponding tag associations can be established in the personnel and demand tag thesaurus; for tags with hierarchical relationships, the hierarchical correspondence between related tags can be recorded through the tag hierarchy diagram. Consistent tag correspondences and normalization rules can be used in the personnel profile tag generation, demand profile tag generation, and tag alignment processes, enabling tags expressed in different textual forms on the personnel profile side and the demand profile side to be processed under a unified tag correspondence, and providing a tag foundation for subsequent candidate retrieval and matching.

[0070] Furthermore, in some embodiments, the synonym database for personnel and requirement tags, as well as the corresponding normalization rule set, can be shared during the personnel profile tag generation, requirement profile tag generation, and subsequent matching processes. For example, during the personnel profile tag generation process, functional information expressed in different textual forms in the resume can be standardized according to the normalization rules; during the requirement profile tag generation and tag alignment processes, the correspondence between requirement profile tags and personnel profile tags can be determined based on the same tag synonym relationships and normalization rules. By using consistent tag synonym relationships and normalization rules in different processing processes, structured personnel profile tags and structured requirement profile tags can have a mutually corresponding tag basis for subsequent candidate recall and matching.

[0071] Step S13: Based on the label alignment mapping, match the structured personnel profile labels with the structured demand profile labels to obtain the matching results.

[0072] In some implementations, the matching includes multi-path recall, and the ranking of candidates is a comprehensive ranking. The multi-path recall includes: matching the hard tags in the structured personnel profile tags and the structured demand profile tags based on an inverted index to obtain a first candidate result; vectorizing the structured personnel profile tags and the structured demand profile tags based on a vector model, and matching them according to the similarity between corresponding vectors to obtain a second candidate result; determining the matching result based on the first candidate result and the second candidate result, and determining a candidate set based on the matching result. The comprehensive ranking includes: for each candidate in the candidate set, determining a statistical matching score, a rule matching score, and a model matching score based on statistical information, preset matching rules, and a matching model; weighting the statistical matching score, the rule matching score, and the model matching score according to preset weights to obtain a total matching score for the corresponding candidate, and ranking them according to the total matching score.

[0073] In some embodiments, after establishing a label alignment mapping between structured personnel profile labels and structured requirement profile labels, personnel profiles are matched with job requirements based on the label alignment mapping. The matching adopts a multi-path recall method, which includes inverted index recall and vector recall. The corresponding candidate results are obtained through two different recall methods, and the matching result is determined based on the candidate results obtained by different recall methods, so as to reduce the omission of candidates that may be caused by using only a single matching method.

[0074] Specifically, in the multi-path recall process, inverted index recall is performed first. Elasticsearch (a distributed search and analytics engine) is used to match the hard tags in the structured personnel profile tags with the hard tags in the structured requirement profile tags. Elasticsearch is used to quickly retrieve and match the corresponding tags based on a pre-built index. The inverted index is used to establish an index relationship between the retrieval tags and the personnel profiles containing the corresponding tags, thereby enabling the rapid retrieval of the corresponding personnel profiles based on the requirement profile tags to be matched. The hard tags can include tags with explicit constraints such as city and education level. Based on the matching results of the hard tags, the first candidate results are obtained. Through inverted index recall, candidates who meet the requirements of the corresponding hard tags can be included in the candidate range. For example, when the hard tags in the job requirements include "City: City A" and "Education: Bachelor's degree or above", candidates whose work city is City A and whose education level is Bachelor's degree or above can be recalled and used as the first candidate results.

[0075] On the other hand, vector recall is performed. Based on the embedding model, personnel profiles and job requirements are vectorized. The embedding model maps textual information to vector representations, and the generated vectors represent the semantic features of the corresponding text. This ensures that semantically similar texts have high similarity in the vector space, facilitating semantic matching based on vector similarity. Matching is performed based on the similarity between the vectors corresponding to the personnel profile and the vectors corresponding to the job requirements to obtain a second candidate result. This second candidate result represents candidates who semantically match the job requirements, supplementing semantically similar candidates that are difficult to recall through hard label matching alone. For example, when the job requirement description includes "Java backend development," although a candidate's personnel profile may not directly contain the label "Java backend development," it may contain semantically related information such as "Spring Boot" and "microservice development." Vector similarity matching can still recall this candidate, making them a second candidate result.

[0076] In some implementations, inverted index recall and vector recall can be carried out in parallel or cascaded. When using parallel recall, the first candidate result can be obtained through inverted index recall and the second candidate result can be obtained through vector recall. The first candidate result and the second candidate result are then merged and deduplicated to form a candidate set.

[0077] When using a cascaded recall approach, a first candidate result can be obtained based on hard label matching. Then, the candidate range of vector recall can be limited by the first candidate result. Based on the vector similarity between the personnel profile and the job requirements, a second candidate result can be obtained from the first candidate result. The candidate set can then be determined based on the second candidate result.

[0078] Therefore, by combining inverted index recall with vector recall, candidate identification can consider both the matching of hard labels such as city and education level, and the semantic similarity between personnel profiles and job requirements, thus balancing the accuracy and comprehensiveness of candidate recall. After determining the candidate set, each candidate in the candidate set is comprehensively ranked. In some embodiments, the comprehensive ranking adopts a hybrid ranking method combining "statistics, rules, and models," determining the statistical matching score, rule matching score, and model matching score of the corresponding candidate based on statistical information, preset matching rules, and matching models, respectively, and determining the candidate's total matching score based on the statistical matching score, rule matching score, and model matching score.

[0079] Specifically, the statistical matching score can be determined based on the overlap between the structured personnel profile tags corresponding to the candidate and the structured requirement profile tags corresponding to the job requirements. For example, the tag intersection or Jaccard similarity can be calculated for the skill tag set to reflect the statistical matching degree between the candidate's actual skills and the job requirements. Jaccard similarity measures the degree of overlap between two sets and can be determined based on the ratio of the number of elements in the intersection to the number of elements in the union of the two sets; a higher similarity indicates a higher degree of overlap between the two sets.

[0080] The rule matching score can be determined based on preset matching rules. For example, it can be determined whether there is any overlap between the salary range corresponding to the candidate and the salary range corresponding to the position, and whether the candidate's job level matches the job level required by the position. The rule matching score is then determined based on the satisfaction of the corresponding rules. The rule matching score can further reflect the degree to which the candidate meets the preset business matching conditions.

[0081] The model matching score can be determined through a matching model. For example, job description information and candidate resume summaries can be input into the Sentence-BERT Chinese vector model to obtain corresponding vector representations. The model matching score is then determined based on the semantic similarity between the two, reflecting the degree of matching between the job requirements and the candidate resume at the overall semantic level. The Sentence-BERT Chinese vector model is a semantic representation model used to map sentences or text into vector representations. The generated vectors can characterize the overall semantic features of the text, enabling semantically similar texts to have high similarity in the vector space.

[0082] After obtaining the statistical matching score, rule matching score, and model matching score respectively, the statistical matching score, rule matching score, and model matching score are weighted according to preset weights to obtain the total matching score of the corresponding candidate. As a specific implementation, the weight corresponding to the statistical matching score can be set to 0.4, the weight corresponding to the rule matching score to 0.3, and the weight corresponding to the model matching score to 0.3. The total matching score of the candidate is then calculated based on the weights of each matching score. The above weights are only used to illustrate one specific implementation; in different implementations, corresponding weights can be set according to actual matching needs.

[0083] After obtaining the total matching score for each candidate, the candidates are ranked according to the total matching score. For example, the candidates can be arranged in descending order of total matching score, and the top N candidates are selected as the recommendation results from the ranked candidate set. Thus, by determining the candidate set through multi-path recall, and further combining statistical information, preset matching rules, and matching models to comprehensively rank the candidates, the determination and ranking of candidates can simultaneously consider the combined influence of hard tag matching, semantic similarity, and different matching factors.

[0084] In some implementations, the method further includes: using a large-model intelligent retrieval agent based on the ReAct framework to obtain the natural language retrieval request input by the user; determining a retrieval strategy based on the natural language retrieval request, wherein the retrieval strategy includes one or more retrieval conditions; and recording the retrieval strategy, weights used to calculate the total matching score, and user feedback information in multiple rounds of dialogue through a session state machine; obtaining the natural language feedback input by the user for the current recommended resume list; recording the natural language feedback as the user feedback information; updating at least one of the retrieval strategy and the weights based on the user feedback information; and re-performing the matching, total matching score calculation, and sorting based on the updated results to update the recommended resume list.

[0085] In some embodiments, a large-scale intelligent retrieval agent based on the ReAct (Reasoning + Acting) framework interacts with the user. The ReAct framework is an intelligent agent implementation framework that combines information analysis, task decision-making, and external tool invocation, enabling the large language model to determine the operation to be performed based on the user's request and adjust subsequent processing based on the operation result. The agent is an intelligent interaction component built on the large language model, used to understand the user's natural language request, determine the retrieval strategy, invoke the corresponding retrieval or recommendation processing, and adjust the retrieval process based on user feedback. The agent receives the natural language retrieval request input by the user and invokes the large language model to perform intent recognition and slot filling on the natural language retrieval request to determine the corresponding retrieval strategy, which includes one or more retrieval conditions determined from the natural language retrieval request.

[0086] For example, when a user enters "Looking for several master's degree holders who know Python, preferably in City A," the Agent performs semantic parsing on this natural language search request, identifying search information such as skills, education level, and city. This identified information is then filled into the corresponding query slots, which are preset fields used to carry different types of search conditions, such as skill slots, education level slots, and city slots, forming the current query parameters and search strategy for candidate retrieval. Based on this search strategy, the Agent invokes intelligent matching and recommendation processing to obtain a list of candidates who match the current search conditions and their corresponding recommended resumes.

[0087] Furthermore, the Agent maintains a session state machine to record and maintain the retrieval state during multi-turn dialogues. The retrieval state includes the current retrieval strategy, current query parameters, weights used in the aforementioned weighted scoring model to calculate the total matching score, and user feedback information. Through this session state machine, when obtaining subsequent natural language feedback from the user, the corresponding retrieval conditions can be updated based on the retrieval state already recorded in the current session, without needing to regenerate a new retrieval strategy based on all retrieval conditions.

[0088] In some embodiments, the retrieval strategy can also be represented in a structured manner using a retrieval DSL (Domain-Specific Language). The retrieval DSL describes retrieval criteria, matching rules, and / or ranking parameters in a structured form. The Agent can generate a corresponding retrieval DSL based on the initial natural language retrieval request input by the user, and invoke matching processing based on the retrieval DSL. The retrieval DSL can serve as a structured representation of the current retrieval strategy and is associated with the retrieval state in the current session, allowing for subsequent modifications based on user feedback.

[0089] After outputting the current recommended resume list to the user, the Agent obtains the user's natural language feedback on the current recommended resume list and records the natural language feedback as user feedback information to participate in the subsequent search strategy and / or weight adjustment.

[0090] For example, when a user comments "Too senior, only 3-5 years of experience is needed," the agent uses a large language model to parse this natural language feedback, identifies the user's intent as adjusting the work experience requirement, and extracts a new work experience requirement: "3 ≤ work experience ≤ 5." Subsequently, based on the current search strategy recorded in the session state machine, the agent modifies the original work experience requirement to 3 to 5 years, while keeping other search conditions in the current search strategy unchanged. For implementations using a search DSL to represent the current search strategy, the search conditions corresponding to the work experience requirement in the search DSL can be modified accordingly.

[0091] After updating the search strategy based on user feedback, the Agent can re-process the matching based on the updated strategy. When the updated search criteria can be processed based on the current candidate set, filtering can be re-executed only for the current candidate set without having to perform a full recall of all resume data. For example, if the user only adjusts work experience, salary range, or other filter criteria, the current candidate set can be re-filtered based on the updated search criteria while keeping the current session state and other unchanged search criteria unchanged, thus achieving incremental retrieval based on user feedback.

[0092] In addition to updating the search strategy, some embodiments can also update the weights used to calculate the total match score based on user feedback. For example, when a user provides feedback such as "I value language skills more" during a conversation, the matching factors that need to be adjusted can be determined based on this user feedback, and the corresponding weights can be adjusted so that the subsequent calculation of the total match score can reflect the matching requirements that the user is currently focusing on.

[0093] After updating at least one of the search strategy and the weights, the matching, total matching score calculation, and sorting are re-performed based on the updated results. Specifically, candidates who meet the corresponding search conditions can be re-determined according to the updated search strategy, and the total matching score of the corresponding candidates can be recalculated according to the updated weights. Then, the candidates can be sorted according to the recalculated total matching score to obtain the updated recommended resume list.

[0094] For example, if a user changes their work experience requirement from "more than 3 years" to "3 to 5 years", the current candidate set can be re-filtered based on the updated work experience requirement, and the total matching score and ranking results can be recalculated for the filtered candidates, thereby achieving incremental retrieval and incremental re-ranking. If the user indicates that they "value language ability more", the corresponding weights used for calculating the total matching score can be adjusted, and the total matching score for each candidate can be recalculated and the candidate ranking can be refreshed based on the adjusted weights.

[0095] In this way, the large-scale intelligent retrieval agent can inherit the currently determined retrieval strategy, the weights used to calculate the total matching score, and user feedback information through the session state machine during multi-round dialogues. It can also make targeted adjustments to at least one of the retrieval strategy and the weights based on the user's subsequent input of natural language feedback. Furthermore, it can re-match, calculate the total matching score, and sort based on the updated results to continuously update the recommended resume list. This allows users to gradually adjust the candidate screening and sorting results through continuous natural language interaction without having to repeatedly input complete search conditions.

[0096] Furthermore, in some embodiments, in addition to dynamically adjusting the retrieval strategy and / or weights based on user feedback as described above, user interaction behavior information during the resume retrieval process can also be recorded. This interaction behavior information includes at least one of the user's actions regarding candidates: viewing, selecting, skipping, and / or modifying search criteria. This interaction behavior information is then associated and stored with the corresponding retrieval strategy, candidates, and matching results. Furthermore, based on the interaction behavior information and user feedback information accumulated within a preset time range, the relevant models used for matching processing and / or the weights used to calculate the total matching score can be updated, so that subsequent retrieval processes can utilize historical interaction behavior and user feedback to adjust the corresponding matching processing.

[0097] In some implementations, intelligent guidance for no results is also included, which includes: obtaining the current search strategy, which includes one or more search conditions; when a candidate retrieval based on the current search strategy fails to obtain a corresponding list of recommended resumes, while keeping other search conditions unchanged, relaxing individual search conditions in the current search strategy and determining the estimated recall after relaxing the corresponding search conditions, where the estimated recall represents the expected number of candidates obtained; and generating corresponding search strategy adjustment suggestions based on the estimated recall corresponding to each search condition.

[0098] In some embodiments, the current search strategy can be a search strategy determined based on a natural language search request input by the user, or it can be a search strategy updated based on user feedback during multiple rounds of interaction. The search criteria may include, for example, at least one of education level, years of work experience, skill requirements, and city.

[0099] When no corresponding list of recommended resumes is obtained based on the current search strategy, the Agent initiates a no-result intelligent guidance mechanism to analyze one or more search conditions in the current search strategy. Specifically, for the target search condition to be analyzed, while keeping other search conditions unchanged, the target search condition is simulated to be relaxed, and the estimated recall after relaxing the target search condition is determined; when the current search strategy includes multiple search conditions, the above simulated relaxation process can be performed on each search condition separately to obtain the estimated recall corresponding to each search condition.

[0100] For example, if the educational requirement in the current search strategy is "Master's degree" and no corresponding list of recommended resumes is obtained according to the current search strategy, the educational requirement can be simulated to be relaxed to "Bachelor's degree" while keeping the years of work experience, skill requirements, city, and other search conditions unchanged, and the number of candidates expected to be obtained after relaxing the educational requirement can be determined.

[0101] Furthermore, based on the estimated recall corresponding to each search condition, the search conditions to be adjusted and the corresponding adjustment methods are determined, and search strategy adjustment suggestions are generated. For example, if relaxing the educational requirement from "Master's" to "Bachelor's" can increase the number of candidates, a search strategy adjustment suggestion can be generated: "It is recommended to adjust the educational requirement to Bachelor's, which is expected to increase the number of candidate resumes by 15."

[0102] After providing the user with the suggested adjustments to the search strategy, the Agent can also obtain the user's confirmation of the proposed adjustments. When the user confirms the corresponding suggested adjustments, the Agent updates the current search strategy based on the user's confirmation and re-executes candidate screening based on the updated search strategy to obtain an updated list of recommended resumes.

[0103] In other embodiments, when the number of candidates obtained based on the current retrieval strategy is lower than a preset threshold, the above method can be used to simulate the relaxation of one or more retrieval conditions in the current retrieval strategy, and retrieval strategy adjustment suggestions can be generated based on the corresponding estimated recall volume to expand the recall scope of candidates.

[0104] Step S14: Determine the candidate set based on the matching results, calculate the total matching score for each candidate in the candidate set, sort the candidates according to the total matching score, and output a recommended resume list.

[0105] In some implementations, when calculating the total matching score for each candidate, the sub-item scores for each matching dimension are determined, and the total matching score is determined based on the sub-item scores for each matching dimension; when outputting the recommended resume list, the total matching score and the sub-item scores for each matching dimension are used as the matching details for the corresponding candidate to form a structured recommendation result.

[0106] In some embodiments, the matching dimensions may include at least one of education, skills, experience, and salary. Based on the matching status of the candidate's structured personnel profile tags and the job requirements' structured requirement profile tags across each matching dimension, corresponding sub-item scores can be determined, and the candidate's overall matching score can be determined based on these sub-item scores. Thus, while obtaining the overall matching degree of the candidate, it also reflects the candidate's specific matching status across different matching dimensions.

[0107] Furthermore, the structured recommendation results may also include fields such as name, gender, age, matching degree, and resume link. The matching degree includes the total matching degree score and matching degree details formed by the scores of sub-items corresponding to each matching dimension, and the resume link is used to locate the corresponding resume information.

[0108] To help those skilled in the art better understand the technical solution of the present invention, the following description, in conjunction with a specific application scenario, further illustrates the intelligent resume retrieval method based on a large language model provided in this application embodiment.

[0109] In one specific implementation, recruiters need to screen candidates for the "Senior Java Backend Engineer" position. The input job description text includes requirements such as "3+ years of experience, Bachelor's degree or above, familiar with Spring Cloud microservices, practical experience in distributed systems, and location in Hangzhou." After obtaining the job description text, the system performs structured information extraction using a large language model to obtain structured requirement profile tags that include educational requirements of Bachelor's degree or above, work experience of 3+ years, city of Hangzhou, core skills including Java, Spring Cloud and distributed systems, and job level of senior.

[0110] After obtaining the structured requirement profile tags, a tag alignment mapping is established between the structured requirement profile tags and the corresponding structured personnel profile tags of the candidates, according to the aforementioned tag alignment method. Then, the candidate set is determined based on inverted index recall and vector recall using the aforementioned multi-path recall method. In this embodiment, 128 candidate resumes are obtained through multi-path recall.

[0111] For the recalled candidate set, the total matching score for each candidate is determined according to the aforementioned comprehensive ranking method. In one specific implementation, the weights corresponding to the statistical matching score, rule matching score, and model matching score can be set to 0.4, 0.3, and 0.3, respectively, and the total matching score of the candidate is calculated based on each matching score and its corresponding weight.

[0112] Furthermore, the overall match score can be displayed in detail. For example, for a candidate with an overall match score of 87%, the match details could include 100% matching in education, 92% in skills, 85% in experience, and 78% in salary. After completing the comprehensive scoring, candidates are sorted according to their overall match scores, and the top N candidates are selected from the sorted set to form a recommended resume list. In this embodiment, the top 20 candidates can be returned, and for each candidate, a candidate summary, overall match score, match details, and a resume preview link can be displayed.

[0113] In another specific embodiment, the candidate screening criteria can be updated based on the initial candidate resume screening results, following the aforementioned method of dynamically adjusting the search strategy based on user natural language feedback. For example, after outputting a list of recommended resumes to recruiters, the recruiter inputs natural language feedback such as "too senior, only 3-5 years of experience is needed." The Agent parses the natural language feedback and updates the work experience requirement in the current search strategy from "more than 3 years" to "3 to 5 years," while keeping other search criteria unchanged.

[0114] When the updated search criteria can be processed based on the current candidate set, the current candidate set can be re-filtered according to the aforementioned incremental search method, and the total matching score and sorting can be recalculated without having to completely recall all resume data again. In this embodiment, after re-screening and sorting candidates based on the updated years of work experience criteria, 18 candidate resumes that meet the updated search criteria can be obtained.

[0115] Furthermore, when no suitable candidates are found according to the updated search criteria or the number of candidates is insufficient, the aforementioned intelligent guidance mechanism for no results can be activated. Based on the estimated recall volume after the relaxation of each search criterion, a search strategy adjustment suggestion can be generated, and after obtaining user confirmation, the candidates can be screened and sorted again based on the adjusted search strategy.

[0116] Therefore, in the above specific embodiments, the process can first generate requirement profile tags, align tags, recall candidates, perform comprehensive scoring, and output a list of recommended resumes based on the job description text. When the recommended results cannot meet the current screening requirements of the recruiters, the large-scale model intelligent retrieval agent obtains the natural language feedback input by the recruiters, makes targeted adjustments to the current search conditions, and continuously updates the list of recommended resumes through incremental filtering, incremental sorting, and intelligent guidance when no results are found, thus forming an iterative screening process for candidate resumes based on natural language interaction.

[0117] The preceding text, with reference to specific embodiments, has provided a detailed explanation of the intelligent resume retrieval method based on a large language model provided in this application. The following text, with reference to other embodiments, will further describe the intelligent resume retrieval system, medium, program product, and terminal based on a large language model provided in this application.

[0118] Based on the above method embodiments, this application also provides a resume intelligent retrieval system based on a large language model. The system includes a personnel profile tag generation module, a requirement profile tag generation module, an intelligent matching and recommendation module, and a large model intelligent retrieval agent module. Each module can be used to execute the corresponding processing procedures in the aforementioned method embodiments. The following, in conjunction with... Figure 2 and Figure 3 The intelligent resume retrieval system based on a large language model provided in the embodiments of this application will be further described.

[0119] like Figure 2 As shown in the embodiments of this application, the intelligent resume retrieval system based on a large language model includes a personnel profile tag generation module, a requirement profile tag generation module, an intelligent matching and recommendation module, and a large model intelligent retrieval agent module.

[0120] The system includes: a personnel profile tag generation module for processing unstructured resume documents to generate structured personnel profile tags; a requirement profile tag generation module for processing job requirement description text to generate structured requirement profile tags; an intelligent matching and recommendation module for performing tag alignment, candidate matching, total matching score calculation, and sorting based on the structured personnel profile tags and the structured requirement profile tags to generate a recommended resume list; and a large-scale intelligent retrieval agent module for acquiring natural language retrieval requests input by users and natural language feedback input to the recommended resume list, and determining or adjusting corresponding retrieval strategies and / or weights for calculating the total matching score based on the natural language retrieval requests and natural language feedback.

[0121] like Figure 3 As shown, the tag generation and tag data management structure provided in this embodiment includes a personnel profile tag generation module 301, a requirement profile tag generation module 302, and a tag data management module 303. The personnel profile tag generation module 301 is used to parse, extract information, validate rules, and normalize unstructured resume documents to generate structured personnel profile tags. The requirement profile tag generation module 302 is used to process and extract information from job requirement description text to generate structured requirement profile tags. The personnel profile tag generation module 301 and the requirement profile tag generation module 302 process unstructured resume documents and job requirement description text respectively, and provide the generated structured personnel profile tags and structured requirement profile tags to the tag data management module 303 for storage and management. The tag data management module 303 is also used to store and manage a thesaurus of personnel and requirement tags and a tag hierarchy relationship diagram for subsequent use in establishing tag alignment mapping between structured personnel profile tags and structured requirement profile tags.

[0122] Specifically, the personnel profile tag generation module 301 includes a document parsing unit 3011, a text extraction unit 3012, an information extraction unit 3013, a rule verification and normalization unit 3014, and a personnel profile tag output unit 3015; the requirement profile tag generation module 302 includes a job requirement text processing unit 3021, an information extraction unit 3022, and a requirement profile tag output unit 3023.

[0123] The personnel profile tag output unit 3015 is used to output structured personnel profile tags obtained after rule validation and normalization processing, and provide the structured personnel profile tags to the tag data management unit 303; the demand profile tag output unit 3023 is used to output structured demand profile tags obtained after extracting structured information, and provide the structured demand profile tags to the tag data management unit 303. The tag data management unit 303 stores and manages the structured personnel profile tags and the structured demand profile tags.

[0124] In some embodiments, the tag data management 303 is further used to store and manage a thesaurus of personnel and demand tags, as well as a tag hierarchy graph. The thesaurus of personnel and demand tags records tag relationships with the same or corresponding semantics between structured personnel profile tags and structured demand profile tags; the tag hierarchy graph records hierarchical correspondences between different tags. The thesaurus of personnel and demand tags and the tag hierarchy graph are used to subsequently determine the correspondence between structured personnel profile tags and structured demand profile tags, and to establish tag alignment mappings based on these correspondences.

[0125] In some embodiments, the intelligent resume retrieval system based on a large language model can be deployed on a server or server cluster using a microservice architecture, with data interaction between functional modules via interface calls and / or message passing. Further, in some embodiments, the tag data management 303 can implement corresponding data storage and management functions based on the system's data storage layer. The data storage layer can be used to store original resumes, structured personnel profile tags, structured requirement profile tags, a thesaurus of personnel and requirement tags, a tag hierarchy diagram, user interaction information, and user feedback information. For example, MongoDB can be used to store original resumes and their corresponding parsing results, Redis can be used to cache frequently accessed tag data, and PostgreSQL can be used to store user interaction logs and feedback data; functional modules can make interface calls via RESTful APIs and asynchronously transmit data through message queues. The above database, interface forms, and message passing methods are merely examples and are not limited in this application.

[0126] In some embodiments, the system may also set up a vector database to store vector representations corresponding to structured personnel profile tags, structured requirement profile tags, resume text, and / or job requirement description text. The vector representations can be obtained by vectorizing the corresponding structured tags or text information using a vector model. When performing vector recall, the intelligent matching and recommendation module can perform similarity retrieval based on the vector representations of the objects to be matched in the vector database to obtain semantically similar candidates and / or candidate positions. The specific implementation of the vector database is not limited.

[0127] Figure 5 This is a schematic block diagram of the electronic terminal provided in an embodiment of this application. Figure 5 As shown, the electronic terminal 500 includes at least one processor 501, a memory 502, at least one communication interface 503, and a user interface 505. The various components in the electronic terminal 500 are coupled together via a bus system 504. It is understood that the bus system 504 is used to implement communication between these components. In addition to a data bus, the bus system 504 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in… Figure 5 The general will label all buses as bus systems.

[0128] The user interface 505 may include a monitor, keyboard, mouse, trackball, buttons, touchpad, or touch screen.

[0129] It is understood that memory 502 can be volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM) or programmable read-only memory (PROM). By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM) and synchronous static random access memory (SSRAM). The memories described in the embodiments of this application are intended to include, but are not limited to, these and any other suitable categories of memory.

[0130] In this embodiment, the memory 502 is used to store the computer program 5021 to support the operation of the electronic terminal 500. The computer program 5021 is stored in the memory 502, and when the processor 501 executes the computer program 5021, it implements the intelligent resume retrieval method based on a large language model as described in any of the above embodiments.

[0131] The methods disclosed in the embodiments of this application can be applied to processor 501, or implemented by processor 501. Processor 501 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in processor 501 or by instructions in the form of software. The processor 501 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 501 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. General-purpose processor 501 may be a microprocessor or any conventional processor, etc. The steps of the resume intelligent retrieval method based on a large language model provided in the embodiments of this application can be directly reflected as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software modules may be located in a storage medium, which is located in memory. The processor reads the information in the memory and combines it with its hardware to complete the steps of the aforementioned method.

[0132] In an exemplary embodiment, the electronic terminal 500 may be implemented by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), or complex programmable logic devices (CPLDs) to perform the aforementioned method.

[0133] According to the method provided in the embodiments of this application, this application also provides a computer program product, which includes: computer program code, which, when run on a computer, causes the computer to perform the method of any of the embodiments described above.

[0134] According to the method provided in the embodiments of this application, this application also provides a computer-readable storage medium storing program code, which, when run on a computer, causes the computer to perform the method of any of the embodiments described above.

[0135] The terms "module," "unit," and "system" used in this specification refer to computer-related entities that implement corresponding functions, which can be implemented using hardware, software, firmware, or a combination of hardware and software. For example, one or more modules or units can be implemented by a processor executing a corresponding computer program, or by corresponding hardware circuitry. One or more modules or units can be integrated or configured separately, and can be implemented in one or more electronic devices as needed. The modules or units can interact with each other through corresponding communication methods to achieve the corresponding functions described in the embodiments of this application.

[0136] Those skilled in the art will recognize that the various illustrative logical blocks and steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0137] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0138] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0139] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0140] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0141] In the above embodiments, the functions of each functional unit can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. A computer program product includes one or more computer instructions (programs). When the computer program instructions (programs) are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0142] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0143] In summary, this application provides a method and system for intelligent resume retrieval based on a large language model. This application extracts structured information from unstructured resume documents and job requirement descriptions, generating personnel profile tags and requirement profile tags, and establishing a tag alignment mapping between these tags. Based on this, a candidate set is obtained through multi-path recall, and comprehensive matching and ranking are performed using statistical information, preset matching rules, and matching models. Simultaneously, a large-model intelligent retrieval agent dynamically adjusts the retrieval strategy and / or matching weights based on the user's natural language retrieval request and natural language feedback on the recommendation results, thereby improving the accuracy and flexibility of resume retrieval and candidate-job matching. Therefore, this application effectively overcomes the problems of insufficient structured resume information, lack of unified alignment between personnel profiles and job requirements, and insufficient flexibility in retrieval methods in existing technologies, and has high industrial application value.

[0144] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A resume intelligent retrieval method based on a large language model, characterized in that, include: Parse the format of unstructured resume documents to obtain resume text; Based on the large language model, structured information is extracted from the resume text and unstructured job requirement description text respectively to obtain structured personnel profile tags and structured requirement profile tags. Based on preset comparison rules, a tag alignment mapping is established between the structured personnel profile tags and the structured requirement profile tags; the preset comparison rules are used to determine the correspondence between the structured personnel profile tags and the structured requirement profile tags. Based on the label alignment mapping, the structured personnel profile labels and the structured requirement profile labels are matched to obtain the matching results; Based on the matching results, a set of candidates is determined, the total matching score for each candidate in the set is calculated, and the candidates are sorted according to the total matching score to output a list of recommended resumes.

2. The intelligent resume retrieval method based on a large language model according to claim 1, characterized in that, After obtaining the structured personnel profile tags, the process also includes: The structured personnel profile tags are subjected to rule-based validation and semantic normalization. Fields in the structured personnel profile tags are validated according to preset validation rules, and fields with semantic ambiguity are standardized and mapped. The structured personnel profile tags are updated based on the results of field validation and standardized mapping.

3. The intelligent resume retrieval method based on a large language model according to claim 1, characterized in that, The preset comparison rules are determined based on a thesaurus of synonyms for personnel and demand tags and a tag hierarchy relationship diagram. The personnel and demand tag thesaurus is used to determine the synonym tag correspondence between the structured personnel profile tags and the structured demand profile tags; The tag hierarchy diagram is used to determine the hierarchical tag correspondence between the structured personnel profile tags and the structured requirement profile tags; Based on the synonym tag correspondence and the hierarchical tag correspondence, establish the tag correspondence between the structured personnel profile tags and the structured requirement profile tags, and establish the tag alignment mapping based on the tag correspondence.

4. The intelligent resume retrieval method based on a large language model according to claim 1, characterized in that, The matching includes multi-path recall, and the candidates are sorted by comprehensive sorting; The multi-path recall includes: matching the hard tags in the structured personnel profile tags and the structured demand profile tags based on an inverted index to obtain a first candidate result; vectorizing the structured personnel profile tags and the structured demand profile tags based on a vector model, and matching them according to the similarity between the corresponding vectors to obtain a second candidate result; and determining the candidate personnel set based on the first candidate result and the second candidate result. The comprehensive ranking includes: for each candidate in the candidate set, determining a statistical matching score, a rule matching score, and a model matching score based on statistical information, preset matching rules, and a matching model; weighting the statistical matching score, the rule matching score, and the model matching score according to preset weights to obtain the total matching score of the corresponding candidate, and ranking them according to the total matching score.

5. The intelligent resume retrieval method based on a large language model according to claim 1, characterized in that, Also includes: A large-scale intelligent retrieval agent based on the ReAct framework is used to obtain the natural language retrieval request input by the user. The retrieval strategy is determined according to the natural language retrieval request. The retrieval strategy includes one or more retrieval conditions. The retrieval strategy, the weight used to calculate the total matching score, and the user feedback information in the multi-turn dialogue are recorded through a session state machine. Obtain natural language feedback from the user regarding the current recommended resume list, record the natural language feedback as the user feedback information, update at least one of the retrieval strategy and the weight based on the user feedback information, and re-perform the matching, total matching score calculation and sorting based on the update result to update the recommended resume list.

6. The intelligent resume retrieval method based on a large language model according to claim 5, characterized in that, It also includes intelligent guidance for no-result scenarios, which includes: Obtain the current search strategy, which includes one or more search conditions; When a candidate retrieval based on the current retrieval strategy fails to obtain a corresponding list of recommended resumes, while keeping other retrieval conditions unchanged, the individual retrieval conditions in the current retrieval strategy are relaxed respectively, and the estimated recall after relaxing the corresponding retrieval conditions is determined. The estimated recall is used to represent the expected number of candidates obtained. Based on the estimated recall for each search condition, corresponding search strategy adjustment suggestions are generated.

7. The intelligent resume retrieval method based on a large language model according to claim 1, characterized in that, When calculating the total matching score for each candidate, the sub-item score corresponding to each matching dimension is determined, and the total matching score is determined based on the sub-item score corresponding to each matching dimension. When outputting the recommended resume list, the total matching score and the sub-item scores corresponding to each matching dimension are used as the matching details of the corresponding candidates to form a structured recommendation result.

8. A resume intelligent retrieval system based on a large language model, characterized in that, include: The module includes a personnel profile tag generation module, a demand profile tag generation module, an intelligent matching and recommendation module, and a large-scale intelligent retrieval agent module. The personnel profile tag generation module is used to parse the format of unstructured resume documents to obtain resume text, and extract structured information from the resume text based on a large language model to obtain structured personnel profile tags. The requirement profile tag generation module is used to extract structured information from unstructured job requirement description text based on a large language model to obtain structured requirement profile tags, and to establish a tag alignment mapping between the structured personnel profile tags and the structured requirement profile tags based on preset comparison rules. The large-scale intelligent retrieval agent module is used to obtain the natural language retrieval request input by the user, determine the retrieval strategy based on the natural language retrieval request, and provide the retrieval strategy to the intelligent matching recommendation module; It is also used to obtain natural language feedback input by the user for the current recommended resume list, update at least one of the retrieval strategy and the weight used to calculate the total matching score based on the natural language feedback, and provide the updated result to the intelligent matching recommendation module; The intelligent matching and recommendation module is used to match the structured personnel profile tags with the structured demand profile tags based on the retrieval strategy and the tag alignment mapping to obtain matching results; when retrieving candidates based on the structured demand profile tags, the module determines the candidate set according to the matching results, calculates the total matching score for each candidate in the candidate set, sorts the candidates according to the total matching score, and outputs a recommended resume list.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the intelligent resume retrieval method based on a large language model as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The method includes a computer program that, when executed by a processor, implements the intelligent resume retrieval method based on a large language model as described in any one of claims 1 to 7.

11. An electronic terminal, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the computer program, it implements the intelligent resume retrieval method based on a large language model as described in any one of claims 1 to 7.

Images