CN116823203A - Recruitment system and recruitment method based on AI large language model - Google Patents
Recruitment system and recruitment method based on AI large language model Download PDFInfo
- Publication number
- CN116823203A CN116823203A CN202310873093.8A CN202310873093A CN116823203A CN 116823203 A CN116823203 A CN 116823203A CN 202310873093 A CN202310873093 A CN 202310873093A CN 116823203 A CN116823203 A CN 116823203A
- Authority
- CN
- China
- Prior art keywords
- job
- recruitment
- post
- language model
- current task
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000007115 recruitment Effects 0.000 title claims abstract description 86
- 238000000034 method Methods 0.000 title claims abstract description 45
- 230000006870 function Effects 0.000 claims abstract description 59
- 239000013598 vector Substances 0.000 claims description 23
- 239000000284 extract Substances 0.000 claims description 8
- 238000009826 distribution Methods 0.000 claims description 7
- 238000005520 cutting process Methods 0.000 claims description 5
- 238000012163 sequencing technique Methods 0.000 claims description 5
- 230000000903 blocking effect Effects 0.000 claims description 3
- 238000005516 engineering process Methods 0.000 abstract description 5
- 238000012216 screening Methods 0.000 abstract description 5
- 230000004044 response Effects 0.000 abstract description 2
- 230000008569 process Effects 0.000 description 8
- 238000004458 analytical method Methods 0.000 description 7
- 238000004364 calculation method Methods 0.000 description 7
- 230000008901 benefit Effects 0.000 description 6
- 238000012549 training Methods 0.000 description 6
- 239000000463 material Substances 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 238000004891 communication Methods 0.000 description 3
- 230000003993 interaction Effects 0.000 description 3
- 230000009191 jumping Effects 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 230000006399 behavior Effects 0.000 description 2
- 230000000295 complement effect Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000014509 gene expression Effects 0.000 description 2
- 238000012821 model calculation Methods 0.000 description 2
- 238000005316 response function Methods 0.000 description 2
- 238000003860 storage Methods 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 239000008358 core component Substances 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000036651 mood Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000003252 repetitive effect Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
Abstract
The invention discloses a recruitment system and a recruitment method based on an AI large language model, which divide complex functions into a plurality of single-function combinations, so that the function logic of the AI large language model is clearer, and the accuracy of the large language model is improved; the AI large language model adopts the requirement interrupt judgment and the requirement execution parallel operation, thereby reducing the feedback time delay. The invention is suitable for internet recruitment, and text understanding, logical reasoning and text generation technologies of the AI large language model are respectively applied to the job hunting terminal and the recruitment terminal to interact with the user, so that the functions of job hunting and recruitment are realized. The job hunting end functions comprise labor regulation consultation, resume generation, post search, resume delivery and automatic response. Recruitment end functions include labor regulation consultation, post generation, resume screening, resume abstract and intelligent customer service. The invention reduces a great amount of manual work in the links of job hunting and recruitment, shortens the recruitment period and reduces the operation threshold of internet job hunting.
Description
Technical Field
The invention belongs to the technical field of internet recruitment, and particularly relates to an AI-based large language model recruitment system and method.
Background
The traditional internet recruitment method comprises the following steps: the recruiter issues recruitment posts on the recruitment platform, the job seeker delivers resume according to recruitment information on the platform, the recruiter screens according to the delivered resume, multiple rounds of interviews are performed after candidate persons are screened, and finally recruitment is completed. In the recruitment process, the recruiter needs to do a lot of repeated work, for example, hundreds of resumes received in one day at hot posts, so that the recruiter can read hundreds of resumes to screen proper talents, namely time and labor are wasted, and meanwhile under the condition of large workload, excellent talents are easily missed.
In the traditional internet recruitment process, a job seeker needs to complete a resume first, then search for an intention post on a recruitment platform for delivery, and then wait for an interview opportunity. The process is relatively easy for the personnel seeking staff who are highly educated, but is relatively difficult for part of blue-led personnel seeking staff, and the personnel often cannot better write own resume, cannot perform a series of operations of recruitment platforms such as post search, delivery and the like, and the part of personnel seeking staff is excluded from a user group by the traditional internet recruitment platform.
In the traditional internet recruitment method, most recruitment platforms do not provide direct communication channels for the job seeker and the recruiter before entering the interview process, so that the job seeker cannot further know the conditions of the company and the position, and the recruiter cannot know the job seeker in advance. In view of this, some recruitment platforms also provide a channel for direct communication between the job seeker and the recruiter, but often many of the problems of the two-way communication are repetitive, such as "what is the company now on a scale? "how much ratio of the accumulation fund? "is there a double-break or a single-break? "what is you away from job? "," how often you can go into job? ", can not accept overtime when the business is busy? "etc., wastes a lot of time for the user to answer questions like this repeatedly asked question, and if an intelligent assistant is available to answer the questions for the user, there is no doubt a lot of time and effort.
In recent years, the development of artificial intelligence technology, especially the appearance of a large artificial intelligence language model, is bringing subversion changes to various industries. The large language model utilizes the technologies of deep learning, natural language processing and the like, has mass knowledge storage, and has higher efficiency and accuracy in many professional fields. How to improve the efficiency of the recruitment system based on a large language model and reduce the operation threshold of the recruitment system becomes a problem to be solved urgently.
Disclosure of Invention
In order to solve the defects of the prior art and achieve the purpose of improving the efficiency of job application and recruitment, the invention adopts the following technical scheme:
a recruitment system based on an AI large language model is applied to a server, the server comprises an AI large language model (LLM, large Language Model), a group of single requirements corresponding to single job-seeking/recruitment functions are extracted from text information acquired by a job-seeking end/recruitment end and used for constructing and executing tasks, meanwhile, whether a current task is interrupted or not is judged according to the real-time text information of the job-seeking end/recruitment end, if the current task is interrupted by a new task before execution is completed, the current task is terminated, the new task is executed as the current task, otherwise, the current task is continuously executed, the job-seeking/recruitment functions corresponding to the output requirements are realized, and the job-seeking end/recruitment end is fed back.
The method has the advantages that the functions based on the language model are simplified, the requirements corresponding to the functions are extracted as singly as possible through the simplified instructions, the complex functions are reversely split into a plurality of single-function combinations, the text is possibly analyzed for a plurality of times by the input language model, a certain computational effort is increased, the logic complexity of the calculation of the language model is reduced, the consumed computational effort is offset to a certain extent, meanwhile, the multiple analysis and clearer calculation logic of the language model are realized, the benefit of the model analysis is improved, and the accuracy of the function realization of the language model is improved. On the other hand, job seeking or recruitment demands acquired based on multiple rounds of conversations are likely to change in the middle, parallel operation of demand interruption judgment and demand execution is adopted, whether the current demand execution result is fed back or not is determined based on the judgment result, and partial calculation force is consumed to replace reduction of feedback time delay of a job seeking end/recruitment end.
Further, cutting and blocking the labor regulation based on regulations and single text quantity, and vectorizing each text message; the method comprises the steps of obtaining text information and scene information of consultation of labor regulations by a job hunting end/recruitment end, comparing similarity between the text information and vectors of the labor regulations after vectorization, obtaining labor regulation text information corresponding to the consultation text information, and generating answering information of consultation of the job hunting end/recruitment end by the AI large language model based on the labor regulation text information and the scene information.
Because the content of the text is more and the longer the text content is processed by the labor regulation, the overflow of the display memory or the memory is easy to cause, the text needs to be cut, the content of each text block can be ensured to be relatively independent by cutting based on regulations, and the cut text content can be as complete as possible by the segmentation based on the single text quantity. Based on the similarity matching of the vectors, the calculation amount required by text information matching can be reduced while the matching precision is maintained.
Further, according to the job requirements of job seekers, the job retrieval task is executed, and the retrieved jobs are ordered based on the job scores, wherein the job scores are as follows:
score= co_s×co _w+sal_s×sal_w+tm_s×tm_w + rx_s×rx_w
wherein co_s represents a unit part to which a post belongs, sal_s represents a salary part, tm_s represents an age part, rx_s represents a resume delivery part, and co_w, sal_w, tm_w and rx_w are weights of the unit part, the salary part, the age part and the resume delivery part respectively;
the unit score is scored based on unit attributes;
the salary score is calculated according to salary levels of posts in the same type of posts;
the time score is distributed according to the post;
and the resume delivery score is used for scoring according to the satisfaction degree of the resume with the post requirements, wherein the satisfaction degree=the delivered quantity of the current post resume/the recruiter of the post, so that more posts with less delivery quantity have more opportunities to be exhibited in front of the search result.
And matching the similarity of the post vectors can be introduced, key fields (such as unit attributes) of the post information are vectorized, and the final ordering of vectors corresponding to the delivered posts is adjusted based on the similarity of the vectors. The method has the advantages that the traffic convenience level of the company scale, company welfare and company legal risk is more, and the demands of different job seekers are relatively abstract, the score is uniformly distributed, the demands of different job seekers are difficult to react, the complement can be made up by vector comparison and adjustment sequencing, the intention of subconscious consciousness of the job seekers is more similar, and for the job seekers, a mutual compromises relationship can exist among scoring items, the ratio obtained after the compromises cannot be embodied on the score, the compromises ratio can be changed along with the resume browsed by the job seekers, the fluctuation of the comprehension ratio can be well adapted through vectorization ratio influence, and the positions wanted in subconscious consciousness of the job seekers are better matched. Thereby further improving the hit rate of job seeker post retrieval and the success rate of post delivery.
Further, introducing randomness to the post scores, subjecting the final score score_n for ranking to a normal distribution:
，
wherein, the original score is taken as the mean value of normal distribution, and sigma represents standard deviation and is used for controlling randomness.
The positions obtained through direct calculation are ranked, so that some positions with low scores have no opportunity to be displayed all the time, and the positions with high scores are ranked in front of the positions with higher probability by introducing a random factor into the ranking of the search results of each position, so that the positions with low scores are provided with the opportunity to be displayed.
Further, acquiring and adjusting a standard deviation sigma based on job-seeking post searching frequency, browsing and delivering conditions; if the post searching frequency exceeds the frequency threshold and the browsed post exceeds the upper threshold, but the delivered post is lower than the lower threshold, the standard deviation is increased.
Because the job seeker browses and delivers according to the search post sequence, the job seeker is not delivered based on certain search and browsing, and the job seeker is not careful about the posts in the sorting, and at the moment, the introduced low-score posts are increased by increasing the standard deviation and the randomness.
Further, a user preference is set and updated based on the unit attribute of the delivered post, the salary level, the release time, the satisfaction of the post demand resume, for adjusting the weights of the company score, the salary score, the time score, and the resume delivery score.
Based on subconscious behaviors of browsing and delivering of job seekers, the weight is adjusted, and the matching degree of the job seekers and posts is improved. Some posts may have low scores due to smaller scale, inconvenient transportation, longer release time and larger satisfaction of delivery requirements of the company, but have higher payroll level, and the job seeker may be subconscious and more conscious of the payroll, so that the post adjusted by the user preference can just meet the requirements of the job seeker.
Further, a user preference post set is set and updated based on the information of the delivered posts, randomness is introduced into the post groups, and the posts which are sequenced by introducing the randomness are sequenced, if the posts are in the user preference post set, the post sequencing is performed, otherwise, the post is not performed. Therefore, the positions which are introduced with randomness and are sequenced are positions preferred by job seekers.
The recruitment system based on the AI large language model is applied to a client, and the client comprises a job hunting end and/or a recruitment end;
sending text information of a user so that an AI large language model of a server side extracts a group of single requirements corresponding to a single job-seeking/recruiting function and is used for constructing and executing a task, continuously judging whether the current task is interrupted or not according to the text information, terminating the current task if the current task is interrupted by a new task, executing the new task as the current task, and otherwise, continuously executing the current task;
and acquiring a result of realizing the job hunting/recruitment function corresponding to the requirement output by the server.
A recruitment method based on an AI large language model comprises the following steps:
step one: extracting a group of single requirements corresponding to single job seeking/recruiting functions from text information acquired by a job seeking end/recruiting end;
step two: constructing and executing a task based on a single requirement, continuously judging whether the current task is interrupted or not according to text information of a job application end/recruitment end, terminating the current task if the current task is interrupted by a new task, executing the new task as the current task, otherwise, continuously executing the current task, and outputting a job application/recruitment function realization result corresponding to the requirement;
step three: and feeding back the function realization result to the job hunting terminal/recruitment terminal.
A language model based system, applied to a server:
and extracting a group of single requirements corresponding to a single function from the text information acquired by the client for constructing and executing the task, continuously judging whether the current task is interrupted or not according to the text information of the client, terminating the current task if the current task is interrupted by the new task, executing the new task as the current task, otherwise, continuously executing the current task, outputting the function realization result corresponding to the requirements, and feeding back the client.
The invention has the advantages that:
according to the recruitment system and the recruitment method based on the AI large language model, the AI large language model is used as a core component of the whole recruitment system, the requirement corresponding to the function is extracted as singly as possible through the simplified instruction, the logic complexity of language model calculation is reduced, meanwhile, multiple analysis and clearer language model calculation logic are realized, the benefit of model analysis is improved, and the accuracy of function realization of the language model is improved. And the parallel operation of demand interruption judgment and demand execution is adopted, so that the feedback time delay of the job hunting end/recruitment end is reduced, and the job hunting and recruitment efficiency is improved.
Drawings
FIG. 1 is a block diagram of a system in an embodiment of the invention.
Fig. 2 is a schematic diagram of a job-seeking end interface in an embodiment of the present invention.
FIG. 3 is a flow chart of a post search function in an embodiment of the invention.
Fig. 4 is a flowchart of a resume screening function in an embodiment of the present invention.
FIG. 5 is a flow chart of a method in an embodiment of the invention.
Description of the embodiments
The following describes specific embodiments of the present invention in detail with reference to the drawings. It should be understood that the detailed description and specific examples, while indicating and illustrating the invention, are not intended to limit the invention.
As shown in fig. 1, a recruitment system based on AI large language model includes 2 clients: job hunting terminals and recruitment terminals. The client is realized by adopting a mobile phone App, a job-seeking interface design is shown in fig. 2, and the recruitment interface is similar to the job-seeking interface and comprises a content display interface, a text input box and a corresponding sending button thereof, and a voice acquisition and sending button. The user interacts with the recruitment system in the form of chat questions and answers by using natural language, not only can directly input text, but also can send voice, and the system converts the voice recognition into text by using a voice recognition technology and then processes the text. The system can return the result in a Text mode, or can use TTS (Text To Speech) technology To convert the Text into Speech and then return the Speech To the user. To describe this, the following description is not directed to text-to-speech conversion problems each time, and the natural language of the user's interaction with the system includes speech by default.
According to the embodiment of the invention, the MOSS model with the large open source is selected as a pre-training model, and then the model is subjected to fine tuning training by using a database, so that the model is more suitable for job hunting and recruitment scenes of the system. The fine tuning training adopts the LoRA (Low-Rank Adaptation of Large Language Models) method, so that the hardware requirement of the fine tuning MOSS is reduced, and the fine tuning time is shortened. The language library used for the fine tuning training is related text materials of job seeking and recruitment prepared in advance, and most of the text materials are presented in a question-answer dialogue mode, so that all aspects of a business scene are covered as much as possible. Because the number of model parameters is up to 160 hundred million, the occupied storage resources and operation resources are huge, and the model cannot be deployed on a terminal, so the model is deployed on a cloud GPU server.
The relational database adopted by the embodiment of the invention adopts MySql, and the vector database adopts Qdrant. The relational database is used for storing job hunting and recruitment information, and the vector database stores a labor regulation knowledge base.
The background operation of the system is carried out around the dialogue with the large language model, according to the context scene, the system inputs the prompt words (comprising the natural language description of the scene and the instructions sent to the large language model) and the text of the user into the large language model to process, the large language model returns the analysis result, if necessary, the system calls the corresponding plug-in to complete specific function execution according to the returned result, for example, in the post searching process, the large language extracts the searching condition of the post from the input text of the job seeker, and the post searching plug-in is provided for searching the post library.
To accomplish a task for a user, multiple rounds of dialogue are often required, specific to each round of interaction, the large language model functions as single as possible, comparing the following 2 large language model instructions: "extract post search condition from job seeker's input text, including post name, job site, salary information, and output in json format" and "analyze his demand from job seeker's input text, extract post search condition from text if post search demand is, including post name, job site, salary information, and output in json format, if resume generation demand is, extract personal information, including name, birth year, month, native place … … (omitted)", the former instruction is concise and clear, the function required to be completed is also single and clear, the latter combines multiple single functions, the logic is more complex. The embodiment adopts a method for simplifying instructions by a large language model, splits complex functions into a plurality of single-function combinations, so that a text input by a user can be input into the large language model for analysis for a plurality of times, and a little computation power is consumed, but the gain is more obvious, the functional logic of a background is clearer, further, fine adjustment training is easier, and the accuracy of the large language model is improved.
When a function of job seeking or recruitment is implemented, a user needs to perform multiple rounds of conversations to complete the current flow, but the user does not run through the current flow every time, and may have other requirements in the middle to interrupt the current flow, the normal processing steps are as follows: firstly judging that the user inputs a new task to interrupt the current task flow, then if the user does not interrupt the current task flow, the large language model continues to analyze the user text according to the instruction of the original flow, otherwise, the user enters the new task flow. Because the requirement of the large language model on calculation force is often high, the time of each execution needs at least several seconds, and the mode of sequentially executing 2 steps obviously prolongs the waiting time of user interaction. To overcome this drawback, the present embodiment employs a two-step asynchronous concurrency approach, pseudo-code as follows:
Begin
Cur = Null
while until the user exits
Inputting user text into the large language model to judge whether the user puts forward a new task (asynchronous execution) (1)
If Cur ≠ Null
Flow operation (asynchronous execution) of current task by inputting user text into large language model (2)
Endif
Waiting for the returned results of steps (1) and (2)
If returns a result according to (1) that the user has new needs
Terminating the current task and assigning a new task to Cur
Else
Continuing the current task, and displaying (2) the execution result to the user
End
In the pseudo code, the step (1) is a branch judgment, the original step (2) is to judge whether the execution is needed or not according to the result of the step (1), the step (2) is advanced to be executed together with the step (1) at present, if the branch of the step (2) is hit, the execution result can be returned more quickly, and the waiting time of the user is shortened.
The functions of the job hunting terminal of the system comprise: labor regulation consultation, resume generation, post search, resume delivery and automatic response. The recruitment end comprises the following functions: labor regulation consultation, post generation, resume screening, resume abstract and intelligent customer service.
The implementation method of the labor regulation consultation function comprises the following steps:
firstly, a labor regulation knowledge base is established, and the method comprises the following steps:
s1, collecting and arranging related labor regulations;
s2, cutting and blocking the rule text according to regulations and paragraphs;
s3, calling a text steering vector plug-in to vector each text, and storing the vector and the corresponding text into a vector database, wherein the vector is used as an index value;
the longer the text large language model is processed more slowly, the more the OOM (memory overflow) is caused, so the large language model cannot analyze all the labor regulations at one time, and the large language model needs to be cut into blocks in advance. The rule of the method for cutting and partitioning the labor rule in the step S2 of establishing the labor rule knowledge base is as follows:
1. each text block does not exceed 2000 words;
2. each text does not exceed 1 chapter;
3. if chapter 1 exceeds 2000 words, then the chapter is evenly split while ensuring that the cut is at the boundaries of the section.
Rule 2 ensures that the content of each text block is relatively independent, and rule 3 leaves the cut text content as complete as possible.
After the labor rule knowledge base is established, the labor rule consultation can be carried out, and the method comprises the following steps:
s1, a user expresses consultation requirements of labor regulations by natural language and describes a scene;
s2, after the large language model recognizes that the requirement of a user is labor regulation consultation, the system calls a text turning quantity plug-in to vectorize the description;
s3, searching a vector database of the labor regulation knowledge base by using the vector obtained in the step S2, finding out a text vector and a text of which the cosine distance is smaller than a specified threshold value, wherein the text and the user description have higher correlation degree.
S4, inputting the text obtained in the step S3 and the scene description of the user into a large language model for analysis, obtaining solutions or suggestions, and displaying the solutions or suggestions to the user in a natural language mode.
The vector database stored in the labor regulation knowledge base adopts Qdrant, which is also responsible for the similarity matching search of text vectors. The job hunting terminal and the recruitment terminal use the same labor rule knowledge base to carry out consultation.
The method for realizing the resume generating function comprises the following steps:
s1, expressing the requirement of generating resume by using natural language by a job seeker;
s2, after the large language model recognizes the requirement of a user for generating resume through semantic understanding, inquiring whether personal information and job-seeking information in a database are complete, if so, jumping to a step S4, and if not, prompting the user that supplementary information is needed;
s3, the user supplements personal information and job seeking information by using natural language, the large language model solution analyzes the information from the dialogue, and then stores the information into a database, and the step S2 is skipped to continuously check the information integrity;
s4, obtaining complete personal information and job seeking information of the user from the database, and calling a function plug-in generated by the related resume to generate the resume by using the information as parameters.
The personal information in the realization method of resume generation refers to personal conditions including name, gender, birth month, cultural degree and the like which need to be disclosed in job hunting. Job hunting information is descriptive information including job hunting posts, desired workplaces, salary requirements and the like and some condition requirements related to job hunting, further including work experiences, project experiences and the like. The generated resume data is stored in a relational database in a system in a way of being divided into fields, and the resume data is expressed as a record in the database. The user can select the template to generate the resume, the template of the resume is in the html format, and the resume generation can be realized by calling the resume generation plug-in to replace corresponding fields in the template with corresponding fields of resume data. After the resume is generated, the Apache POI with an open source can be further called, and the Apache POI is converted into common document formats such as word, pdf and the like for export.
The post search function is implemented as shown in fig. 3, and the method comprises the following steps:
s1, expressing the requirement of a job seeker for finding work by using natural language;
s2, after the large language model judges the job searching requirement of the job seeker, prompting the job seeker to input post searching conditions;
s3, describing post search conditions by using natural language by a user, and analyzing the search conditions from the description by using a large language model solution;
s4, judging whether the search conditions are sufficient, if not, prompting, and jumping to the step S3 to enable the job seeker to further describe;
s5, matching a plurality of posts with high coincidence degree from a post library according to search conditions obtained from the description of the job seeker by the large language model, and giving the job seeker.
The post information in the post search function implementation method comprises post names, workplaces, salary levels, working year requirements and other fields. S2, judging the job searching requirement of the job seeker, and designing the corpus corresponding to the function to accurately make judgment for a large language model, wherein more expression modes such as ' I want to find a job "," want to find a security job in Hangzhou ', ' Hangzhou, security job ', ' find a security job, hangzhou, salary over 5000 and the like are all the requirements of the job searching. Similarly, when other user requirements are determined, various expressions and moods are covered as much as possible, so that the large language model achieves better fine tuning effect on the function. The "necessary search condition" described in step S4 refers to the job name and job site in this implementation, and other job information fields are optional. Because the post names are different in terms of the names, searching is carried out only from text matching and is easy to miss, synonym replacement processing is needed, such as searching for the post of a cleaner, and searching for the post named as a cleaner can be carried out.
The number of post search results is often large, the post search results need to be ordered and displayed in a paging mode, and attention and resume delivery are easier to obtain for the posts arranged in front. The basis of the post ordering is post division, and the calculation formula is as follows:
score= co_s×co _w+sal_s×sal_w+tm_s×tm_w + rx_s×rx_w+pref_s×pref_w
the post total score is 100 minutes, the score is composed of 4 parts, the company score (co_s) to which the post belongs, the salary score (sal_s), the time-efficiency score (tm_s) and the resume delivery score (rx_s), and each part is 100 minutes. co_w, sal_w, tm_w and rx_w are the score weights of the 4 parts, respectively, and the sum of the weights is 1. The factors of the corporate evaluation are: company scale, company welfare, company legal risk, office place traffic convenience degree and the like, and the factors are comprehensively evaluated to obtain scores of the company, so that the method is not described again. The pay score is scored according to the level of the same type of post, the higher the pay score is, the more the pay score is, the score reaching 2 times of the average pay score can be set, and the score 0 is obtained when the average pay is 0.5 times. The time score is calculated according to the post release time, and the score is higher for the post which is newly released, and the score of the release time within 3 days can be set to be more than 0 score of 2 months. The resume delivery score is determined according to the satisfaction degree of the position demand resume, namely the delivered quantity of the current position resume is/the recruitment number of the position, the score is lower when the value is larger, the score is set to be 0, the score is 1, and the score is 0 when the value is larger than or equal to 10, so that more opportunities are exhibited in front of the search result at the positions with less delivery quantity.
And matching the similarity of the post vectors can be introduced, key fields (such as unit attributes) of the post information are vectorized, and the final ordering of vectors corresponding to the delivered posts is adjusted based on the similarity of the vectors. The method has the advantages that the traffic convenience level of the company scale, company welfare and company legal risk is more, and the demands of different job seekers are relatively abstract, the score is uniformly distributed, the demands of different job seekers are difficult to react, the complement can be made up by vector comparison and adjustment sequencing, the intention of subconscious consciousness of the job seekers is more similar, and for the job seekers, a mutual compromises relationship can exist among scoring items, the ratio obtained after the compromises cannot be embodied on the score, the compromises ratio can be changed along with the resume browsed by the job seekers, the fluctuation of the comprehension ratio can be well adapted through vectorization ratio influence, and the positions wanted in subconscious consciousness of the job seekers are better matched. Thereby further improving the hit rate of job seeker post retrieval and the success rate of post delivery.
The fact that the posts calculated directly by the formula are ranked also causes a problem that some posts with low scores are not shown at all times, so that the strategy of ranking posts with high scores in the front is changed into ranking posts with higher scores in the front in the embodiment of the invention. The method for realizing the strategy is to introduce randomness into the original score so that the score score_n is subjected to normal distributionWherein the normal distribution has a mean value score and a standard deviation sigma. In the implementation process, the randomness is controlled by adjusting the size of sigma, and the larger the sigma is, the larger the randomness is. The score_n is obtained by introducing a random factor method into the sorting of the search results of each post, obeys a normal distribution with the average score, and is sorted by the size of the score_n, so that the posts with high scores are ranked in front of the higher probability, and the posts with low scores are given a chance of being displayed.
Acquiring and adjusting standard deviation sigma based on job-seeking post searching frequency, browsing and delivering conditions; if the post searching frequency exceeds the frequency threshold and the browsed post exceeds the upper threshold, but the delivered post is lower than the lower threshold, the standard deviation is increased.
Because the job seeker browses and delivers according to the search post sequence, the job seeker is not delivered based on certain search and browsing, and the job seeker is not careful about the posts in the sorting, and at the moment, the introduced low-score posts are increased by increasing the standard deviation and the randomness.
And setting and updating a user preference based on the unit attribute of the delivered post, the payroll level, the release time and the satisfaction of the post demand resume, and adjusting weights of company points, payroll points, time-lapse points and resume delivery points.
Based on subconscious behaviors of browsing and delivering of job seekers, the weight is adjusted, and the matching degree of the job seekers and posts is improved. Some posts may have low scores due to smaller scale, inconvenient transportation, longer release time and larger satisfaction of delivery requirements of the company, but have higher payroll level, and the job seeker may be subconscious and more conscious of the payroll, so that the post adjusted by the user preference can just meet the requirements of the job seeker.
Setting and updating a user preference post set based on the information of the delivered posts, introducing randomness to the post groups, and improving the sequenced posts by introducing randomness, if the posts are in the user preference post set, improving the post sequencing, otherwise, not improving the posts. Therefore, the positions which are introduced with randomness and are sequenced are positions preferred by job seekers.
The method for realizing the automatic response function comprises the following steps:
s1, judging the type of a problem by using a large language model;
s2, if the answer is a preset question, a preset answer is given, and the answer is completed;
s3, if the answer is a personal information question, calling a database plug-in to search personal related information, and giving an answer to complete the answer;
s4, if the questions cannot be answered automatically, the questions are transferred to be answered by workers.
In the implementation step S1 of the automatic response function, the model does not judge the inputted question types at one time, but separately discriminates and judges 2 question types, and in order to save time, the 2 judgments are simultaneously inputted to the large language model. The preset questions and the preset answers in step S2 refer to some questions and answers frequently asked by the job seeker, which are input into the system in advance. When a question is asked, the large language model determines whether the question is a preset question, and this determination method does not mean that the question description text is completely consistent, but the meaning of the question is similar, such as "what is you away? "and" why do you leave? "the two questions may be considered identical and may be answered with the same answer.
The method for realizing the post generating function comprises the following steps:
s1, a user expresses and generates a post requirement by using natural language;
s2, after the large language model judges the requirement, a user is required to describe posts;
s3, describing post information by a user;
s4, judging whether the description of the post information by the user is finished or not by the large language model, and if the description is incomplete, turning to the step S3;
s5, after post description is completed, the large language model reorganizes the post description into a complete and orderly post description text;
s6, inquiring whether the user issues or not, and if yes, calling an issue plug-in and an issue post.
The post generation required fields include post names, job sites, salaries, academic requirements, post capability requirements, post responsibilities, etc. Through a dialogue mode, recruiters perfect the field information, and the large language model forms a complete post description text according to the information. The post field processing for non-descriptive classes is relatively simple and basically direct, but requires further processing for descriptive fields such as post capability requirements and post responsibilities, mainly because the user may only enter a few words with critical information, such as the "3-year Java multithreading MVC framework", for convenience, and the large language model needs to describe the "more than 3-year Java development experience, familiarity with multithreading programming, proficiency with MVC framework" according to the post requirements that the several words form a complete order. To enhance the large language model's ability to generate posts, we collected a large number of post descriptive text covering various industries as material to fine-tune the large language model MOSS. The sentences in the post text material are segmented, after the high-frequency words are removed, the left low-frequency words are keywords with larger information carrying capacity, the keywords are used as the input of fine tuning training, the original text is used as the output comparison target (Label) of fine tuning, and the method saves the workload of manually extracting the keywords to manufacture posts to generate a corpus.
The resume screening function is realized, as shown in fig. 4, and the method comprises the following steps:
s1, searching a resume delivery library for delivering specified posts, and ending the exit if the unprocessed resume number is 0;
s2, taking out an untreated resume from the post delivery library;
s3, comparing relevant fields of the current resume according to the rigid condition requirements of posts, if the conditions are not met, marking the resume as processed, discarding the resume, and jumping to the step S1;
s4, inputting descriptive fields such as education experiences, work experiences, project experiences and the like in post descriptions and resume into a large language model and requiring scoring of the matching degree of the post descriptions and the resume;
s5, if the matching score of descriptive fields in the post description and the resume is larger than a threshold value, the resume is passed, otherwise, the resume is eliminated, and the resume is marked as processed. Returning to step S1.
The hard condition of the step S3 in the resume screening function implementation specifically refers to a condition that can be subjected to simple logic comparison, such as fields of an academic calendar, a working life, and the like, and the results can be conveniently searched in a relational database through the fields. And for descriptive information which cannot be directly compared, understanding through a large language model, and scoring the matching degree.
Resume summary function: and automatically analyzing the resume and outputting the abstract. The method comprises the following steps:
s1, writing out an instruction text by using a pure natural language, wherein the instruction text is used for extracting a resume abstract, and further, the requirements of abstract extraction can be written;
s2, extracting each field element of the resume, splicing the field elements into a section of plain text, and splicing the section of plain text after the instruction text in the step S1;
s3, inputting the spliced text into a large language model, and outputting the abstract of the resume by the large language model.
Outputting a resume abstract, wherein the important point is to concentrate and refine the texts by a large language model aiming at the work experience and project experience parts with longer text amplitude in the resume, and highlighting the highlight part of the resume. In step S1, the instruction for extracting the abstract may additionally require, for example: "keep only the last 1 job experience", "ignore project experience", etc.
The method for realizing the intelligent customer service function comprises the following steps:
s1, judging the type of a problem by using a large language model;
s2, if the answer is a preset question, a preset answer is given, and the answer is completed;
s3, if the answer is related to the company or the post, calling a database plug-in to search the information of the company and the post, and giving an answer to finish the answer;
s4, if the questions cannot be answered automatically, the questions are transferred to be answered by workers.
In the implementation step S1 of the intelligent customer service function, the model does not judge the input problem types at one time, but separately discriminates and judges 2 problem types, and in order to save time, the 2 judgments are simultaneously and concurrently input to the large language model. The preset questions and the preset answers in step S2 refer to some questions and answers frequently asked by recruiters, which are input into the system in advance. When a question is asked, the large language model determines whether the question is a preset question, and this determination method does not mean that the question description text is completely consistent, but the meaning of the question is similar, such as "how many people the company is now? What are the "and" companies' current personnel sizes? "the two questions may be considered identical and may be answered with the same answer.
The recruitment system based on the AI large language model is applied to a client, and the client comprises a job hunting end and/or a recruitment end;
sending text information of a user so that an AI large language model of a server side extracts a group of single requirements corresponding to a single job-seeking/recruiting function and is used for constructing and executing a task, continuously judging whether the current task is interrupted or not according to the text information, terminating the current task if the current task is interrupted by a new task, executing the new task as the current task, and otherwise, continuously executing the current task;
and acquiring a result of realizing the job hunting/recruitment function corresponding to the requirement output by the server.
As shown in fig. 5, the recruitment method based on the AI large language model includes the following steps:
step one: extracting a group of single requirements corresponding to single job seeking/recruiting functions from text information acquired by a job seeking end/recruiting end;
step two: constructing and executing a task based on a single requirement, continuously judging whether the current task is interrupted or not according to text information of a job application end/recruitment end, terminating the current task if the current task is interrupted by a new task, executing the new task as the current task, otherwise, continuously executing the current task, and outputting a job application/recruitment function realization result corresponding to the requirement;
step three: and feeding back the function realization result to the job hunting terminal/recruitment terminal.
A system based on a language model is applied to a server, a group of single requirements corresponding to a single function are extracted from text information acquired by a client side and used for constructing and executing a task, meanwhile, whether a current task is interrupted or not is continuously judged according to the text information of the client side, if the current task is interrupted by a new task, the current task is terminated, the new task is executed as the current task, otherwise, the current task is continuously executed, the function realization result corresponding to the requirements is output, and the client side is fed back.
The above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some or all of the technical features thereof can be replaced with equivalents; such modifications and substitutions do not depart from the spirit of the technical solutions according to the embodiments of the present invention.
Claims (10)
1. Recruitment system based on AI large language model is applied to the server, characterized in that: the server comprises an AI large language model, extracts a group of single requirements corresponding to single job-seeking/recruiting functions from text information acquired by the job-seeking/recruiting ends, and is used for constructing and executing tasks, meanwhile, judging whether the current task is interrupted or not according to the real-time text information of the job-seeking/recruiting ends, if the current task is interrupted by a new task before the execution of the current task is completed, terminating the current task, executing the new task as the current task, otherwise, continuing to execute the current task, outputting job-seeking/recruiting functions corresponding to the requirements, and feeding back the job-seeking/recruiting ends.
2. The recruitment system based on AI large language model of claim 1, wherein: cutting and blocking labor regulations based on regulations and single text quantity, and vectorizing each piece of text information; the method comprises the steps of obtaining text information and scene information of consultation of labor regulations by a job hunting end/recruitment end, comparing similarity between the text information and vectors of the labor regulations after vectorization, obtaining labor regulation text information corresponding to the consultation text information, and generating answering information of consultation of the job hunting end/recruitment end by the AI large language model based on the labor regulation text information and the scene information.
3. The recruitment system based on AI large language model of claim 1, wherein: according to the job requirements of job seekers, executing a job retrieval task, and sorting the retrieved jobs based on job scores, wherein the job scores are as follows:
score= co_s×co_w+sal_s×sal_w+tm_s×tm_w + rx_s×rx_w
wherein co_s represents a unit part to which a post belongs, sal_s represents a salary part, tm_s represents an age part, rx_s represents a resume delivery part, and co_w, sal_w, tm_w and rx_w are weights of the unit part, the salary part, the age part and the resume delivery part respectively;
the unit score is scored based on unit attributes;
the salary score is calculated according to salary levels of posts in the same type of posts;
the time score is distributed according to the post;
and the resume delivery score is calculated according to the satisfaction degree of the post demand resume, wherein the satisfaction degree=the delivered number of the current post resume/the recruiter of the post.
4. The AI-large language model based recruitment system of claim 3 wherein: introducing randomness to the post scores, and enabling the final score score_n used for ranking to be subjected to normal distribution:
，
wherein, the original score is taken as the mean value of normal distribution, and sigma represents standard deviation and is used for controlling randomness.
5. The AI-large language model based recruitment system of claim 4 wherein: acquiring and adjusting standard deviation sigma based on job-seeking post searching frequency, browsing and delivering conditions; if the post searching frequency exceeds the frequency threshold and the browsed post exceeds the upper threshold, but the delivered post is lower than the lower threshold, the standard deviation is increased.
6. The AI-large language model based recruitment system of claim 3 wherein: and setting and updating a user preference based on the unit attribute of the delivered post, the payroll level, the release time and the satisfaction of the post demand resume, and adjusting weights of company points, payroll points, time-lapse points and resume delivery points.
7. The AI-large language model based recruitment system of claim 3 wherein: setting and updating a user preference post set based on the information of the delivered posts, introducing randomness to the post groups, and improving the sequenced posts by introducing randomness, if the posts are in the user preference post set, improving the post sequencing, otherwise, not improving the posts.
8. Recruitment system based on AI large language model is applied to the customer end, its characterized in that: the client comprises a job hunting end and/or a recruitment end;
sending text information of a user so that an AI large language model of a server side extracts a group of single requirements corresponding to a single job-seeking/recruiting function and is used for constructing and executing a task, continuously judging whether the current task is interrupted or not according to the text information, terminating the current task if the current task is interrupted by a new task, executing the new task as the current task, and otherwise, continuously executing the current task;
and acquiring a result of realizing the job hunting/recruitment function corresponding to the requirement output by the server.
9. A recruitment method based on an AI large language model is characterized by comprising the following steps:
step one: extracting a group of single requirements corresponding to single job seeking/recruiting functions from text information acquired by a job seeking end/recruiting end;
step two: constructing and executing a task based on a single requirement, continuously judging whether the current task is interrupted or not according to text information of a job application end/recruitment end, terminating the current task if the current task is interrupted by a new task, executing the new task as the current task, otherwise, continuously executing the current task, and outputting a job application/recruitment function realization result corresponding to the requirement;
step three: and feeding back the function realization result to the job hunting terminal/recruitment terminal.
10. A language model based system for use with a server, comprising:
and extracting a group of single requirements corresponding to a single function from the text information acquired by the client for constructing and executing the task, continuously judging whether the current task is interrupted or not according to the text information of the client, terminating the current task if the current task is interrupted by the new task, executing the new task as the current task, otherwise, continuously executing the current task, outputting the function realization result corresponding to the requirements, and feeding back the client.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310873093.8A CN116823203A (en) | 2023-07-17 | 2023-07-17 | Recruitment system and recruitment method based on AI large language model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310873093.8A CN116823203A (en) | 2023-07-17 | 2023-07-17 | Recruitment system and recruitment method based on AI large language model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116823203A true CN116823203A (en) | 2023-09-29 |
Family
ID=88139140
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310873093.8A Pending CN116823203A (en) | 2023-07-17 | 2023-07-17 | Recruitment system and recruitment method based on AI large language model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116823203A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117495321A (en) * | 2023-11-03 | 2024-02-02 | 北京探也智能科技有限公司 | Self-help recruitment method and device based on large model |
| CN117524192A (en) * | 2023-11-08 | 2024-02-06 | 北京中科深智科技有限公司 | Speaker pause processing method in speech recognition |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110534108A (en) * | 2019-09-25 | 2019-12-03 | 北京猎户星空科技有限公司 | A kind of voice interactive method and device |
| CN111080234A (en) * | 2019-11-25 | 2020-04-28 | 苏州思必驰信息科技有限公司 | Heuristic dialogue recruitment method and system |
| CN113268352A (en) * | 2021-06-11 | 2021-08-17 | 中科院软件研究所南京软件技术研究院 | Multi-instruction response type task collaborative management method facing general service robot |
| CN113539261A (en) * | 2021-06-30 | 2021-10-22 | 大众问问(北京)信息科技有限公司 | Man-machine voice interaction method and device, computer equipment and storage medium |
-
2023
- 2023-07-17 CN CN202310873093.8A patent/CN116823203A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110534108A (en) * | 2019-09-25 | 2019-12-03 | 北京猎户星空科技有限公司 | A kind of voice interactive method and device |
| CN111080234A (en) * | 2019-11-25 | 2020-04-28 | 苏州思必驰信息科技有限公司 | Heuristic dialogue recruitment method and system |
| CN113268352A (en) * | 2021-06-11 | 2021-08-17 | 中科院软件研究所南京软件技术研究院 | Multi-instruction response type task collaborative management method facing general service robot |
| CN113539261A (en) * | 2021-06-30 | 2021-10-22 | 大众问问(北京)信息科技有限公司 | Man-machine voice interaction method and device, computer equipment and storage medium |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117495321A (en) * | 2023-11-03 | 2024-02-02 | 北京探也智能科技有限公司 | Self-help recruitment method and device based on large model |
| CN117524192A (en) * | 2023-11-08 | 2024-02-06 | 北京中科深智科技有限公司 | Speaker pause processing method in speech recognition |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109885672B (en) | Question-answering type intelligent retrieval system and method for online education | |
| CN116823203A (en) | Recruitment system and recruitment method based on AI large language model | |
| US7606700B2 (en) | Adaptive task framework | |
| KR102318103B1 (en) | Method for machine learning train set and recommendation systems to recommend the scores to match between the recruiter and job seekers, and to give the scores of matching candidates to recruiters and to give the pass scores to job seekers respectively | |
| CN111105209B (en) | Job resume matching method and device suitable for person post matching recommendation system | |
| KR102416684B1 (en) | Ai-based support system for writing resume and job referral | |
| JP7297458B2 (en) | Interactive content creation support method | |
| CN112765974B (en) | Service assistance method, electronic equipment and readable storage medium | |
| CN109325780A (en) | A kind of exchange method of the intelligent customer service system in E-Governance Oriented field | |
| CN111311180A (en) | Resume screening method and device | |
| CN116362699A (en) | Post matching report generation method | |
| CN110929007A (en) | Electric power marketing knowledge system platform and application method | |
| CN114357127A (en) | Intelligent question-answering method based on machine reading understanding and common question-answering model | |
| KR20200139008A (en) | User intention-analysis based contract recommendation and autocomplete service using deep learning | |
| CN113505606B (en) | Training information acquisition method and device, electronic equipment and storage medium | |
| WO2005038584A2 (en) | Matching job candidate information | |
| Lamba et al. | An integrated system for occupational category classification based on resume and job matching | |
| CN112650838A (en) | Intelligent question and answer method and system based on historical case big data | |
| CN113495963A (en) | Embedded representation method and device of network security knowledge graph | |
| CN117609477B (en) | Large model question-answering method and device based on domain knowledge | |
| CN115188381B (en) | Voice recognition result optimization method and device based on click ordering | |
| Miura et al. | Interactive Minutes Generation System Based on Hierarchical Discussion Structure | |
| Kim | The Future Past of Humanities Research: Musing Methodology in the Digital Convergence Era | |
| Mutiara et al. | The Crowdsourcing Method to Normalize “Bahasa Alay”, a Case of Indonesian Corpus | |
| Kukoyi et al. | Voice Information Retrieval In Collaborative Information Seeking |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |