Human-post matching large language model actuarial method based on intelligent retrieval

Abstract

The invention discloses an actuarial method for a man-post matching large language model based on intelligent retrieval, and the method comprises the steps: extracting key information related to post information from a candidate resume document through constructing the man-post matching large language model according to a post candidate information problem; inputting the extracted key information into the constructed post structure constraint tree, and distributing a corresponding hierarchical semantic ID according to an inherent hierarchical structure of the post structure constraint tree; and finally, according to the hierarchical semantic ID contained in each candidate resume document, retrieving a resume which is suitable for a post and ranks in the top from the candidate resume set. According to the method, post matching degree prediction and talent information retrieval are combined together in a sequence-to-sequence mode in a large language model generation retrieval framework, dual prediction of the post matching degree and the talent information is completed at the same time in single inference, the difference between the prediction and the actual post matching degree is effectively reduced, the talent retrieval result is optimized, and the retrieval efficiency is improved. Therefore, the matching accuracy and consistency are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically a method for precise calculation of a large language model for matching people and jobs based on intelligent retrieval. Background Technology

[0002] In the field of talent-job matching data processing, talent information retrieval and job matching degree prediction are the core tasks of intelligent talent-job matching systems. Accurately determining whether talent and job possess matching characteristics is crucial for establishing their relevance in talent-job retrieval. However, current talent retrieval research has significant limitations: some studies neglect the importance of job matching degree prediction, while others attempt to use implicit training objectives, hoping to measure the matching degree by the distance between talent and job in the vector space. However, these methods lack clear modeling basis for the consistency of talent-job matching, resulting in poor accuracy of matching results and a lack of transparency in the retrieval process. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a precise calculation method for job matching based on a large language model of intelligent retrieval. Within the retrieval framework generated by the large language model, job matching degree prediction and talent information retrieval are combined in a sequence-to-sequence manner. In a single inference, both job matching degree and talent information are predicted simultaneously, which effectively reduces the difference between the prediction and the actual job matching degree, optimizes the talent retrieval results, and thus improves the accuracy and consistency of matching.

[0004] The technical solution of this invention is as follows:

[0005] A method for precise calculation of a large language model for matching people and jobs based on intelligent retrieval, specifically including the following steps:

[0006] (1) Construct a set of candidate resumes D, Set up a question x to provide information about job candidates. Question x includes specific job information.

[0007] (2) Construct a large language model for person-job matching. The large language model for person-job matching includes an information extraction module, a job structure constraint tree, an information indexing module, and a resume retrieval module. The information extraction module is used to retrieve candidate resume documents d from the candidate resume set D based on the job candidate information question x. i Extracting key information related to job information from the information extraction module; the job structure constraint tree is constructed using an inherent hierarchical structure of department-job-skill; the information index module is used to extract candidate resume documents from the information extraction module. i Key information is input into the job structure constraint tree, and corresponding hierarchical semantic IDs are assigned based on the inherent hierarchical structure of the job structure constraint tree; the resume retrieval module retrieves information based on each candidate's resume document d. iThe hierarchical semantic IDs contained therein are used to retrieve resumes that are suitable for the position and ranked highly from the candidate resume set D.

[0008] (3) The constructed large language model for matching people and jobs is trained based on the joint optimization loss function to obtain the trained large language model for matching people and jobs. Then, the set job candidate information question x is input into the trained large language model for matching people and jobs to perform precise matching and obtain the job matching answer corresponding to the job candidate information question x and output it.

[0009] The information extraction module is used to extract candidate resume documents d from candidate resume set D based on the job candidate information question x. i Extracting key information related to job postings involves the following steps:

[0010] S11. Construct a corpus B = {B1, B2, B3} based on job requirements for each job position. B1 is constructed from the job title and its lexical variants, B2 is constructed from the skill requirements in the job description and their lexical variants, and B3 is constructed from the embedding vectors of the job description.

[0011] S12. Decompose candidate resume documents at the lexical level. i Specifically, it is based on B1 and B2 in corpus B from candidate resume document d. i The keyword βb was selected as the key information related to the job information, as shown in the following formula (1):

[0012]

[0013] In equation (1), j∈{1,2}; K j The hyperparameter represents the number of keywords selected, and freq() represents the time frequency; from the candidate resume document d i The selected keyword βb includes βb1 and βb2;

[0014] S13. Segmenting candidate resume documents at the sentence level. i Specifically, it is based on B1 and B2 in corpus B from candidate resume document d. i The key sentence βs is selected as the key information related to the job information, as shown in the following formula (2):

[0015]

[0016] In equation (2), b1 and b2 represent candidate resume documents d. iIn this context, represents the words in sentence s; freq() represents the frequency of occurrence; length(s) represents the length of sentence s; emb() represents the embedding vector; cos() represents cosine similarity; μ represents the set balance parameter; K s This is a hyperparameter that controls the number of key sentences selected.

[0017] S14. From candidate resume document d i The key information selected that is relevant to the job information corresponds to the skills in the resume that are relevant to the job information.

[0018] The candidate resume document d i Each key piece of information related to the job information selected in the process, namely each skill, corresponds to a hierarchical semantic ID. The hierarchical semantic ID is represented in a prefix plus suffix format. The prefix is ​​determined based on the position of the skill in the job structure constraint tree in the candidate's resume document. That is, after finding the position corresponding to the skill in the job structure constraint tree, the skill is represented according to the inherent hierarchical structure of this position. That is, hierarchical semantic ID = number represented by the company root node - number represented by the department node - number represented by the job node - number represented by the skill node. The suffix is ​​the unique identifier set for this skill node in the candidate's resume document.

[0019] The above is based on each candidate's resume document d i The hierarchical semantic IDs included retrieve suitable and top-ranked resumes from the candidate resume set D. Specifically, after the model is trained, beam search is used. In each step of the search, multiple candidate sequences with high probability are retained, and the search continues along these beams to query and retrieve candidate resume documents in an end-to-end manner. Finally, the multiple optimal sequences returned by the beam search are integrated to output the person-job matching answer corresponding to the job candidate information question x.

[0020] The joint optimization loss function is shown in Equation (3) below:

[0021]

[0022] In equation (3), L represents the joint optimization loss function. Indicates indexing loss. Indicates retrieval loss. Indicates a correction of loss;

[0023] The index loss The specific calculation is shown in the following formula (4):

[0024]

[0025] In equation (4), id i Representative candidate resume document d iThe hierarchical semantic ID obtained from the index, P(id) i |d i ) indicates that in a given candidate resume document d i Under these conditions, candidate resume document d i Belongs to ID i The conditional probability;

[0026] The aforementioned retrieval loss The specific calculation is shown in the following formula (5):

[0027]

[0028] In equation (5), P(id) i |x) represents the condition that, given the job candidate information problem x, the job candidate information problem x belongs to id. i The conditional probability;

[0029] The aforementioned correction loss The specific calculation is shown in the following formula (6):

[0030]

[0031] In equation (6), λ represents the hyperparameter; k represents the k optimal sequences returned by the retrieval; R t Indicates a reward; Representative job matching large language model prediction of candidate resume document d i Corresponding hierarchical semantic ID; id i For candidate resume document d i The true hierarchical semantic ID; L represents the length of the hierarchical semantic ID, and the predicted hierarchical semantic ID. and the real hierarchical semantic ID id i The length of each document is L; N is the number of candidate resume documents d in the candidate resume set D. i The total number; exp() is the exponential function; sigmoid represents the sigmoid normalization function, which makes the reward R... t The value range is -1 to 1.

[0032] Advantages of this invention:

[0033] (1) The job matching big language model of the present invention constructs a corpus based on job requirements for each job and mines deep semantic information by combining the big language model. Thus, based on the job candidate information question, it can extract key information related to job information from the candidate resume document and use this key information as the retrieval basis to generate a common and information-rich document representation for the two tasks (constructing hierarchical semantic ID and question retrieval matching), thereby improving the retrieval efficiency.

[0034] (2) Based on the inherent hierarchical structure of department-position-skill, this invention constructs a job structure constraint tree. By constructing an information index module automatically extracted based on a large language model, and representing each candidate's resume document as a hierarchical semantic ID based on the job structure constraint tree, the dual prediction of job matching degree and relevant talent information can be completed simultaneously in a single person-job matching.

[0035] (3) The present invention trains the constructed human-job matching big language model based on the joint optimization loss function, which effectively reduces the difference between the predicted and actual job matching IDs, optimizes the talent retrieval results, and thus improves the accuracy and consistency of the human-job matching big language model. Attached Figure Description

[0036] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] A method for precise calculation of a large language model for matching people and jobs based on intelligent retrieval includes the following steps:

[0039] (1) Construct a set of candidate resumes D, Set up a question x to provide information about job candidates. Question x includes specific job information.

[0040] (2) Construct a large language model for matching people and jobs. The large language model for matching people and jobs includes an information extraction module, a job structure constraint tree, an information index module, and a resume retrieval module.

[0041] The information extraction module is used to extract candidate resume documents d from the candidate resume set D based on the job candidate information question x. i Extracting key information related to job postings involves the following steps:

[0042] S11. Construct a corpus B = {B1, B2, B3} for each job position based on job requirements. B1 is constructed from the job title and its lexical variations. For example, for the skill "software development" in the job requirements, stop words are split and removed, and the remaining part and its lexical variations are added to B1. B2 is constructed from the skill requirements in the job description and their lexical variations. The definitions of skill requirements in the job description are split and stop words are removed, and the remaining part and its lexical variations are added to B2. To enhance the corpus, a Large Language Model (LLM) designed for the human resources domain is used to obtain the definitions of skill requirements in the job description. After obtaining the LLM results, stop words are removed and the results are incorporated into B2. B3 is constructed from the embedding vectors of the job description.

[0043] S12. Deconstruct candidate resume documents at the lexical level. i Using Large Language Modeling (LLM) hints designed for the human resources domain, the definitions of skill requirements in job descriptions are obtained. After obtaining the LLM results, lexical corpus information is acquired; specifically, based on B1 and B2 in corpus B, lexical information is obtained from candidate resume documents d. i The keyword βb was selected as the key information related to the job information, as shown in the following formula (1):

[0044]

[0045] In equation (1), j∈{1,2}; K j The hyperparameter represents the number of keywords selected, and freq() represents the time frequency; from the candidate resume document d i The selected keyword βb includes βb1 and βb2;

[0046] S13. Segmenting candidate resume documents at the sentence level. i Using Large Language Modeling (LLM) hints designed for the human resources domain, the definitions of skill requirements in job descriptions are obtained. After obtaining the LLM results, sentence-level corpus information is acquired; specifically, based on B1 and B2 in corpus B, information is extracted from candidate resume documents d. i The key sentence βs is selected as the key information related to the job information, as shown in the following formula (2):

[0047]

[0048] In equation (2), b1 and b2 represent candidate resume documents d. iIn this context, represents the words in sentence s; freq() represents the frequency of occurrence; length(s) represents the length of sentence s; emb() represents the embedding vector; cos() represents cosine similarity; μ represents the set balance parameter; K s This is a hyperparameter that controls the number of key sentences selected.

[0049] S14. From candidate resume document d i The key information related to the job information selected in the resume corresponds to the skills related to the job information in the resume;

[0050] The job structure constraint tree is constructed using the inherent hierarchical structure of department-job-skill;

[0051] The information indexing module is the candidate resume document extracted by the information extraction module. i Key information from the candidate resume document is input into the job structure constraint tree. Based on the inherent hierarchical structure of the job structure constraint tree, a corresponding hierarchical semantic ID is assigned. i Each key piece of information related to the job information selected in the document, i.e., each skill, corresponds to a hierarchical semantic ID. The hierarchical semantic ID is represented in a prefix plus suffix format. The prefix is ​​determined according to the position of the skill in the candidate's resume document in the job structure constraint tree. That is, after finding the position corresponding to the skill in the job structure constraint tree, the skill is represented according to the inherent hierarchical structure of this position. That is, hierarchical semantic ID = number represented by the company root node - number represented by the department node - number represented by the job node - number represented by the skill node. For example, the skill "Java development" involved in the resume belongs to the inherent hierarchical structure of "technical department - software development - programming language". Therefore, the prefix of the skill involved in this resume is 0-2-5-269. 0 represents the root node of the job structure constraint tree, i.e., the company root node. 2 represents the number represented by the technical department in the department branch structure. 5 represents the number represented by the software development position among all positions in the technical department. 269 represents the number represented by the programming language in the skill range involved in the software development position. The suffix is ​​the unique identifier set by the candidate's resume document under this skill node.

[0052] When a resume includes multiple skills, multiple semantic IDs are assigned to the resume involving multiple skills. Each semantic ID corresponds to one skill. For example, if a resume contains the skills "Java development" and "data analysis", then the resume has two semantic IDs: 0-2-5-267-725 and 0-2-5-269-725.

[0053] The resume retrieval module searches for each candidate's resume document d iThe hierarchical semantic ID contained therein retrieves resumes that are suitable for the position and ranked highly from the candidate resume set D. Specifically, after the model is trained, beam search is used. In each step of the search, multiple candidate sequences with high probability are retained and the search continues along these beams to query and retrieve candidate resume documents in an end-to-end manner. Finally, the multiple optimal sequences returned by the beam search are integrated to output the person-job matching answer corresponding to the job candidate information question x.

[0054] (3) Train the constructed large language model for matching people and jobs based on the joint optimization loss function to obtain the trained large language model for matching people and jobs. Then, input the set job candidate information question x into the trained large language model for matching people and jobs to perform precise matching and obtain the job matching answer corresponding to the job candidate information question x and output it.

[0055] The joint optimization loss function is shown in equation (3) below:

[0056]

[0057] In equation (3), L represents the joint optimization loss function. Indicates indexing loss. Indicates retrieval loss. Indicates a correction of loss;

[0058] The index loss The specific calculation is shown in the following formula (4):

[0059]

[0060] In equation (4), id i Representative candidate resume document d i The hierarchical semantic ID obtained from the index, P(id) i |d i ) indicates that in a given candidate resume document d i Under these conditions, candidate resume document d i Belongs to ID i The conditional probability;

[0061] The aforementioned retrieval loss The specific calculation is shown in the following formula (5):

[0062]

[0063] In equation (5), P(id) i |x) represents the condition that, given the job candidate information problem x, the job candidate information problem x belongs to id. i The conditional probability;

[0064] The aforementioned correction loss The specific calculation is shown in the following formula (6):

[0065]

[0066] In equation (6), λ represents the hyperparameter; k represents the k optimal sequences returned by the retrieval; R t Indicates a reward; Representative job matching large language model prediction of candidate resume document d i Corresponding hierarchical semantic ID; id i For candidate resume document d i The true hierarchical semantic ID; L represents the length of the hierarchical semantic ID, and the predicted hierarchical semantic ID. and the real hierarchical semantic ID id i The length of each document is L; N is the number of candidate resume documents d in the candidate resume set D. i The total number; exp() is the exponential function; sigmoid represents the sigmoid normalization function, which makes the reward R... t The value range is -1 to 1.

[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for precise calculation of a large language model for matching people and jobs based on intelligent retrieval, characterized in that: Specifically, it includes the following steps: (1) Construct a set of candidate resumes D, Set up a question x to provide information about job candidates. Question x includes specific job information. (2) Construct a large language model for person-job matching. The large language model for person-job matching includes an information extraction module, a job structure constraint tree, an information indexing module, and a resume retrieval module. The information extraction module is used to retrieve candidate resume documents d from the candidate resume set D based on the job candidate information question x. i Extracting key information related to job information from the information extraction module; the job structure constraint tree is constructed using an inherent hierarchical structure of department-job-skill; the information index module is used to extract candidate resume documents from the information extraction module. i Key information is input into the job structure constraint tree, and corresponding hierarchical semantic IDs are assigned based on the inherent hierarchical structure of the job structure constraint tree; the resume retrieval module retrieves information based on each candidate's resume document d. i The hierarchical semantic IDs contained therein are used to retrieve resumes that are suitable for the position and ranked highly from the candidate resume set D. (3) The constructed large language model for matching people and jobs is trained based on the joint optimization loss function to obtain the trained large language model for matching people and jobs. Then, the set job candidate information question x is input into the trained large language model for matching people and jobs to perform precise matching and obtain the job matching answer corresponding to the job candidate information question x and output it.

2. The method for precise calculation of a large language model for matching people and jobs based on intelligent retrieval as described in claim 1, characterized in that: The information extraction module is used to extract candidate resume documents d from candidate resume set D based on the job candidate information question x. i Extracting key information related to job postings involves the following steps: S11. Construct a corpus B = {B1, B2, B3} based on job requirements for each job position. B1 is constructed from the job title and its lexical variants, B2 is constructed from the skill requirements in the job description and their lexical variants, and B3 is constructed from the embedding vectors of the job description. S12. Deconstruct candidate resume documents at the lexical level. i Specifically, it is based on B1 and B2 in corpus B from candidate resume document d. i The keyword βb was selected as the key information related to the job information, as shown in the following formula (1): In equation (1), j∈{1,2}; K j The hyperparameter represents the number of keywords selected, and freq() represents the frequency of selection; from the candidate resume document d i The selected keyword βb includes βb1 and βb2; S13. Segmenting candidate resume documents at the sentence level. i Specifically, it is based on B1 and B2 in corpus B from candidate resume document d. i The key sentence βs is selected as the key information related to the job information, as shown in the following formula (2): In equation (2), b1 and b2 represent candidate resume documents d. i In this context, represents the words in sentence s; freq() represents the frequency of occurrence; length(s) represents the length of sentence s; emb() represents the embedding vector; cos() represents cosine similarity; μ represents the set balance parameter; K s This is a hyperparameter that controls the number of key sentences selected. S14, From candidate resume document d i The key information selected that is relevant to the job information corresponds to the skills in the resume that are relevant to the job information.

3. The method for precise calculation of a large language model for matching people and jobs based on intelligent retrieval, as described in claim 2, is characterized in that: The candidate resume document d i Each key piece of information related to the job information selected in the process, namely each skill, corresponds to a hierarchical semantic ID. The hierarchical semantic ID is represented in a prefix plus suffix format. The prefix is ​​determined based on the position of the skill in the job structure constraint tree in the candidate's resume document. That is, after finding the position corresponding to the skill in the job structure constraint tree, the skill is represented according to the inherent hierarchical structure of this position. That is, hierarchical semantic ID = number represented by the company root node - number represented by the department node - number represented by the job node - number represented by the skill node. The suffix is ​​the unique identifier set for this skill node in the candidate's resume document.

4. The method for precise calculation of a large language model for matching people and jobs based on intelligent retrieval as described in claim 1, characterized in that: The above is based on each candidate's resume document d i The hierarchical semantic IDs included retrieve suitable and top-ranked resumes from the candidate resume set D. Specifically, after the model is trained, beam search is used. In each step of the search, multiple candidate sequences with high probability are retained, and the search continues along these beams to query and retrieve candidate resume documents in an end-to-end manner. Finally, the multiple optimal sequences returned by the beam search are integrated to output the person-job matching answer corresponding to the job candidate information question x.

5. The method for precise calculation of a large language model for matching people and jobs based on intelligent retrieval, as described in claim 4, is characterized in that: The joint optimization loss function is shown in Equation (3) below: In equation (3), L represents the joint optimization loss function. Indicates indexing loss. Indicates retrieval loss. Indicates a correction of loss; The index loss The specific calculation is shown in the following formula (4): In equation (4), id i Representative candidate resume document d i The hierarchical semantic ID obtained from the index, P(id) i |d i ) indicates that in a given candidate resume document d i Under these conditions, candidate resume document d i Belongs to ID i The conditional probability; The aforementioned retrieval loss The specific calculation is shown in the following formula (5): In equation (5), P(id) i |x) represents the condition that, given the job candidate information problem x, the job candidate information problem x belongs to id. i The conditional probability; The aforementioned correction loss The specific calculation is shown in the following formula (6): In equation (6), λ represents the hyperparameter; k represents the k optimal sequences returned by the retrieval; R t Indicates a reward; Representative position matching large language model prediction of candidate resume document d i Corresponding hierarchical semantic ID; id i For candidate resume documents d i The true hierarchical semantic ID; L represents the length of the hierarchical semantic ID, and the predicted hierarchical semantic ID. and the real hierarchical semantic ID id i The length of each document is L; N is the number of candidate resume documents d in the candidate resume set D. i The total number; exp() is the exponential function; sigmoid represents the sigmoid normalization function, which makes the reward R t The value range is -1 to 1.

Images