An automatic recruitment robot system based on large model RAG and fine-tuning

Abstract

The present invention relates to the field of administrative recruitment technology. The present invention discloses an automatic recruitment robot system based on a large model RAG and fine-tuning; the system comprises the following steps: identifying recruitment features, expanding the recruitment features into retrieval features, retrieving original retrieval data, obtaining comprehensive weight data, calculating recruitment correlation values, screening out effective retrieval data, identifying target data corresponding to the recruitment features, importing the effective retrieval data and the target data into a report position respectively, and generating a recruitment report sheet; compared with the prior art, the present invention can retrieve original retrieval data that preliminarily meets the recruitment needs of a user end through RAG technology, and train a retrieval recognition model that can identify target data for specific and special recruitment needs through fine-tuning technology, thereby meeting the requirements for accurate and rapid retrieval and recognition operations for data that meets specific recruitment needs of the user end, and facilitating the human resources department of an enterprise to identify suitable target data from massive data according to recruitment needs.

Description

Technical Field

[0001] The present invention relates to the field of administrative recruitment technology, and more specifically, to an automatic recruitment robot system based on a large model RAG and fine-tuning. Background Art

[0002] At present, more and more companies and organizations choose to recruit talents through the Internet. The main process of online recruitment includes: HR prepares recruitment documents, publishes recruitment needs on recruitment websites and receives resumes submitted by job seekers, conducts comprehensive analysis, and decides whether an interview is needed. In order to improve the retrieval and recognition accuracy of recruitment data and improve the efficiency of corporate recruitment, it is necessary to use an automatic recruitment robot system to retrieve and identify recruitment data, thereby achieving efficient and accurate recruitment results for the company.

[0003] The patent application with reference publication number CN113537920A discloses an intelligent personnel recruitment system, which is based on big data algorithms, AI intelligent matching and tripartite SaaS cooperation, focusing on improving the recruitment efficiency of the HR department and the full-process business connection, which can improve the HR efficiency by more than three times, so that enterprises can have their own recruitment portals. At the same time, enterprises can realize digital talent management and improve management efficiency. It can provide a full range of data services for talent recruitment from the aspects of application, interview and later talent tracking and evaluation. By establishing parsing algorithms, knowledge graphs and big data analysis, it provides services for corporate personnel recruitment, so as to stably match the employment needs of enterprises and optimize the use of talents.

[0004] When the existing automatic recruitment robot system searches and identifies massive recruitment data, it can screen out target data that matches the recruitment needs by comparing the association between the recruitment data and the recruitment needs one by one to achieve the recruitment effect. Since the types of recruitment data stored in the database are complex and cover a wide range, only the recruitment data with obvious features can be identified through a single comparison method, so that some recruitment data with unclear features cannot be retrieved and identified, resulting in the target data retrieved and identified being not accurate enough. At the same time, it is also impossible to perform targeted data retrieval and identification operations based on specific or special needs in the recruitment needs. It is easy for the target data retrieved and identified to be different from the company's recruitment needs, which is not conducive to the company's efficient and accurate recruitment.

[0005] In view of this, the present invention proposes an automatic recruitment robot system based on large model RAG and fine-tuning to solve the above problems. Summary of the invention

[0006] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solutions: an automatic recruitment robot system based on a large model RAG and fine-tuning, applied to a retrieval and recognition server, comprising:

[0007] A feature recognition module is used to obtain recruitment data from the user end, extract features to be recognized from the recruitment data, and recognize recruitment features from the features to be recognized;

[0008] The first retrieval module is used to expand the recruitment features into retrieval features based on the expansion criteria, and retrieve the original retrieval data corresponding to the retrieval features through the RAG technology. The expansion criteria are: expanding the structure value and the attribute value of the retrieval feature into a standard structure and a standard attribute respectively;

[0009] The calculation and screening module is used to obtain the comprehensive weight data of the original search data, which includes the effective character ratio and the effective update time, calculate the recruitment relevance value of the original search data, and screen out the effective search data from the original search data;

[0010] The second retrieval module is used to train a retrieval recognition model through fine-tuning technology, and identify target data corresponding to the recruitment characteristics through the retrieval recognition model;

[0011] The recruitment report module is used to construct a basic report form with a reporting position, which includes a first reporting position and a second reporting position, and import the valid search data and the target data into the reporting position respectively to generate a recruitment report form.

[0012] Furthermore, the method for extracting the features to be identified includes:

[0013] Mark all the stored data in the enterprise database one by one, and query the IP numbers of all the stored data and the IP numbers of the user terminals one by one through the address management system, record the stored data whose IP numbers are consistent with the IP numbers of the user terminals as recruitment data, and obtain i recruitment data;

[0014] Record job title, education, salary, region and benefits as standard semantics;

[0015] The recruitment semantics of i recruitment data are identified one by one through natural language processing technology, and the recruitment semantics with any standard semantics are recorded as target semantics to obtain p target semantics;

[0016] The recruitment data corresponding to the p target semantics are split one by one by using word segmentation technology to generate p character lines, and the characters in the p character lines are numbered in sequence;

[0017] The numbers corresponding to the p target semantic characters are recorded as target numbers, and the two numbers adjacent to the target number are recorded as adjacent numbers;

[0018] After summarizing the characters corresponding to the target number and the adjacent numbers in sequence, p features to be recognized are obtained.

[0019] Further, the methods for identifying recruitment characteristics include:

[0020] Through the permission management system, query the permission remark boxes of p features to be identified one by one, and mark the form remark value and the permission remark value in the permission remark box;

[0021] The features to be identified whose form remark value and permission remark value are both high are recorded as features to be verified, and s features to be verified are obtained;

[0022] The features to be verified that have repeated target semantics are recorded as repeated features, and the recording times of the target semantics in the x repeated features are queried one by one through the timestamps, which are recorded as x recording times;

[0023] Arrange the x recorded moments in chronological order, remove duplicate features except the first recorded moment, and record the remaining duplicate features and features to be verified as recruitment features to obtain m recruitment features.

[0024] Furthermore, the extended method of retrieving features includes:

[0025] Through the structure management system and attribute management system, the structure and attributes of historical data in the big database are queried respectively to obtain the standard structure and standard attributes;

[0026] The structures and attributes of m recruitment features are queried one by one through the structure management system and the attribute management system, recorded as feature structures and feature attributes, and the structure values ​​of the feature structures and the attribute values ​​of the feature attributes are marked to obtain m structure values ​​and m attribute values;

[0027] Recording the recruitment feature whose structure value is not the standard structure as the first invalid feature, eliminating the structure value in the first invalid feature, and importing the standard structure into the feature structure of the first invalid feature to obtain the first valid feature;

[0028] Recording the recruitment feature whose attribute value is not the standard attribute as the second invalid feature, eliminating the attribute value in the second invalid feature, and importing the standard attribute into the feature attribute of the second invalid feature to obtain the second valid feature;

[0029] The first effective feature, the second effective feature and the remaining recruitment features are recorded as retrieval features, and m retrieval features are obtained.

[0030] Furthermore, the retrieval method of the original retrieval data includes:

[0031] Divide the historical data in the large database into m data sets, and randomly correspond the m data sets to the m search features one by one;

[0032] Mark the data input ports and data output ports of the m data sets respectively to obtain m input ports and m output ports;

[0033] Import m search features into m input ports one by one, filter out historical data with the same recruitment semantics as the m search features from the m data sets through RAG technology, and compress the retrieved historical data to generate m original data packets;

[0034] After the m original data packets are output through the m output ports and decompressed, we can obtain The original search data.

[0035] Furthermore, the method for obtaining the effective character ratio includes:

[0036] Through word segmentation technology, The original search data is split into original characters and counted The number of original characters in the original search data is obtained Total value;

[0037] The user-preset standard characters are queried from the large database, the original characters that are repeated with the standard characters are recorded as valid characters, and the number of valid characters is counted to obtain Valid values;

[0038] Will The valid values ​​are The total value is compared to obtain Percentage of valid characters;

[0039] The expression for the percentage of valid characters is:

[0040] ;

[0041] In the formula, For the The percentage of valid characters in the original search data. For the The valid value of the original search data, For the The total value of the original search data.

[0042] Furthermore, the expression of recruitment association value is:

[0043] ;

[0044] In the formula, For the Recruitment related values ​​of the original search data, For the The effective update time of the original search data, , is a proportional coefficient greater than 0;

[0045] Filtering methods to effectively retrieve data include:

[0046] Will Recruitment related value Sequentially associated with the preset recruitment threshold Compare;

[0047] when Greater than or equal to When The original search data are recorded as effective search data, and q effective search data are obtained;

[0048] when Less than When The original search data is recorded as the effective search data.

[0049] Furthermore, the training method of the retrieval recognition model includes:

[0050] Aggregate recruitment features and effective retrieval data to generate a feature data set, convert the feature data set into a corresponding set of feature vectors, construct the feature data set and the corresponding target data into a sample set of the retrieval recognition model, and divide the sample set into training samples and test samples;

[0051] The feature vectors in the training samples are used as the input of the retrieval and recognition model, and the target data corresponding to each set of feature data sets is used as the output of the retrieval and recognition model. The target data is used as the prediction target, and minimizing the sum of prediction errors of all training samples is used as the training target. The retrieval and recognition model is trained and tested with test samples until the sum of prediction errors converges. The training is stopped to obtain a retrieval and recognition model that can predict the target data.

[0052] Furthermore, the construction method of the basic report sheet includes:

[0053] Mark all report form templates in the report form database one by one, identify the template attributes of all report form templates, and record the report form templates with the template attribute as valid templates as valid reports, and obtain t valid report forms;

[0054] The generation time of t valid report sheets is queried one by one through the timestamp, and after arranging the t generation time in chronological order, the valid report sheet corresponding to the last generation time is selected as the target report sheet;

[0055] Draw two report positions arranged vertically in the target report form, and record the two report positions as the first report position and the second report position from top to bottom to obtain a basic report form.

[0056] Furthermore, the method for generating the recruitment report sheet includes:

[0057] Draw q independent data units in the first reporting position to obtain q first units;

[0058] Import q valid search data into q first units one by one, and annotate the keyword "valid" in the first report position after the import of the valid search data;

[0059] Drawing z independent data units in the second reporting position to obtain z second units;

[0060] Import z target data into z second units one by one, and annotate the keyword "target" in the second report position after importing the target data;

[0061] Draw a table unit on the recruitment report form above the first report position, and note the keyword "recruitment report" in the table unit to generate a recruitment report form.

[0062] The technical effects and advantages of the automatic recruitment robot system based on large model RAG and fine-tuning of the present invention are as follows:

[0063] The present invention obtains recruitment data from a user end, extracts features to be identified from the recruitment data, identifies recruitment features from the features to be identified, expands the recruitment features into retrieval features based on an extension criterion, retrieves original retrieval data corresponding to the retrieval features through a RAG technology, obtains comprehensive weight data of the original retrieval data, calculates recruitment association values ​​of the original retrieval data, and screens out effective retrieval data from the original retrieval data, trains a retrieval recognition model through a fine-tuning technology, identifies target data corresponding to the recruitment features through the retrieval recognition model, constructs a basic report sheet with a reporting position, and imports effective retrieval data and target data into the reporting position respectively, to generate a recruitment report sheet; compared with the prior art, by expanding the recruitment features into retrieval features, the retrieval features can be used to identify the recruitment data of the user end, and the recruitment data can be used to identify the recruitment data of the user end. It can accurately represent the recruitment needs of the user end, and at the same time provide reasonable and clear retrieval standards for RAG technology, so as to retrieve the original retrieval data that initially meets the recruitment needs of the user end and has a large coverage range, and calculate the recruitment correlation value in combination with the comprehensive proportion data, and screen out effective retrieval data with a small coverage range from the original retrieval data, avoiding the omission of data with unclear features, and training a retrieval and recognition model that can identify target data for specific and special recruitment needs through fine-tuning technology, which meets the precise and rapid retrieval and recognition operations of data that meets the specific recruitment needs of the user end, and facilitates the human resources department of the enterprise to identify suitable target data from massive data according to recruitment needs, thereby improving the two-way interaction effect of data between enterprises and personnel, and effectively improving the recruitment success rate of enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A schematic diagram of an automatic recruitment robot system based on a large model RAG and fine-tuning provided in Embodiment 1 of the present invention;

[0065] Figure 2 A flowchart of an automatic recruitment robot method based on a large model RAG and fine-tuning provided in the second embodiment of the present invention;

[0066] Figure 3 A schematic diagram of a structural diagram of an electronic device provided in Embodiment 3 of the present invention;

[0067] Figure 4 A schematic diagram of the structure of a computer-readable storage medium provided in Embodiment 4 of the present invention. DETAILED DESCRIPTION

[0068] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] Example 1: Please refer to Figure 1 As shown, the automatic recruitment robot system based on the large model RAG and fine-tuning described in this embodiment is applied to the retrieval and recognition server, including:

[0070] The feature recognition module obtains the recruitment data from the user, extracts the features to be recognized from the recruitment data, and recognizes the recruitment features from the features to be recognized;

[0071] The user end refers to the port that can use the automatic recruitment robot to publish recruitment information and screen and identify recruitment information, and serves as the data interaction end of the automatic recruitment robot. Specifically, the user end is the human resources department of the enterprise responsible for personnel recruitment. Recruitment data refers to the original recruitment information input by the user end to the automatic recruitment robot in order to meet the recruitment needs, which can be used as the pre-data basis for subsequent personnel recruitment;

[0072] Since recruitment data is the original information input by the user, the information and content contained in the recruitment data are large in quantity and complex in type. In order to accurately simplify the complex recruitment data, it is necessary to extract the features to be identified from the recruitment data so that the features to be identified can perform a preliminary simplified representation of the information and content in the recruitment data.

[0073] The extraction methods of the features to be identified include:

[0074] Mark all the stored data in the enterprise database one by one, and query the IP numbers of all the stored data and the IP numbers of the user terminals one by one through the address management system, record the stored data whose IP numbers are consistent with the IP numbers of the user terminals as recruitment data, and obtain i recruitment data;

[0075] Position, education, salary, region and benefits are recorded as standard semantics; standard semantics refers to the semantics that can express the true meaning of the recruitment needs of the user side and serve as the basis for subsequent identification and judgment of identification features;

[0076] The recruitment semantics of i recruitment data are identified one by one through natural language processing technology, and the recruitment semantics with any standard semantics are recorded as target semantics to obtain p target semantics;

[0077] The recruitment data corresponding to the p target semantics are split one by one by using word segmentation technology to generate p character lines, and the characters in the p character lines are numbered in sequence;

[0078] The numbers corresponding to the p target semantic characters are recorded as target numbers, and the two numbers adjacent to the target number are recorded as adjacent numbers;

[0079] After summarizing the characters corresponding to the target number and the adjacent numbers in sequence, p features to be recognized are obtained.

[0080] After the features to be identified are extracted, the features to be identified can comprehensively represent the user-side recruiter's demand information and the information provided by the recruiter, so that the automatic recruitment robot can make a broad and conceptual representation of the user-side recruitment information. In order to represent the recruitment information more accurately and concisely, it is necessary to identify the recruitment features from the features to be identified.

[0081] Methods for identifying recruitment characteristics include:

[0082] Through the permission management system, query the permission remark boxes of p features to be identified one by one, and mark the form remark value and the permission remark value in the permission remark box; the permission remark box is a content box used to remark the form state and the permission state of the feature to be identified, and the form remark value and the permission remark value are specific representations of the form state and the permission state in the permission remark box; the form remark value and the permission remark value both include high values ​​and low values; a high value indicates that the form state and the permission state are both in a state of passing the permission record, and a low value indicates that the form state and the permission state are both in a state of not passing the permission record;

[0083] The features to be identified whose form remark value and permission remark value are both high are recorded as features to be verified, and s features to be verified are obtained;

[0084] The features to be verified that have repeated target semantics are recorded as repeated features, and the recording times of the target semantics in the x repeated features are queried one by one through the timestamps, which are recorded as x recording times;

[0085] Arrange the x recorded moments in chronological order, remove duplicate features except the first recorded moment, and record the remaining duplicate features and features to be verified as recruitment features to obtain m recruitment features.

[0086] It should be noted that recruitment features are features used to present the recruitment information on the user side in a real and concise manner, and serve as the basis for the subsequent initial screening of recruitment personnel information, that is, to initially screen out data information that meets the recruitment needs of the user side from the external database.

[0087] The first retrieval module, based on the expansion criteria, expands the recruitment features into retrieval features, and retrieves the original retrieval data corresponding to the retrieval features through the RAG technology;

[0088] After the recruitment features are obtained, they can only represent the recruitment needs of the user, and there may be incompatibility and mismatch between them and the stored data in the external database. In order to improve the convenience of information data retrieval in the external database and avoid retrieval errors, it is necessary to expand the recruitment features so that the recruitment features are expanded into retrieval features that can match the stored data in the external database;

[0089] When expanding recruitment features into search features, the search features generated after the expansion need to be able to directly match the stored data in the external database, so it is necessary to use the expansion criteria to expand the recruitment features;

[0090] The extension criteria are: expand the structure value and attribute value of the search feature to the standard structure and standard attribute respectively; this ensures that the structure value and attribute value in the search feature are consistent with the standard structure and standard attribute, and provides a good foundation for the subsequent search and identification operations of the RAG technology;

[0091] Extension methods for retrieving features include:

[0092] Through the structure management system and attribute management system, the structure and attributes of historical data in the big database are queried respectively to obtain the standard structure and standard attributes;

[0093] The structures and attributes of m recruitment features are queried one by one through the structure management system and the attribute management system, recorded as feature structures and feature attributes, and the structure values ​​of the feature structures and the attribute values ​​of the feature attributes are marked to obtain m structure values ​​and m attribute values;

[0094] Recording the recruitment feature whose structure value is not a standard structure as the first invalid feature, removing the structure value in the first invalid feature, and importing the standard structure into the feature structure of the first invalid feature to obtain the first valid feature; so that the recruitment feature corresponding to the first invalid feature can be expanded to generate features that meet subsequent needs;

[0095] Record the recruitment feature whose attribute value is not the standard attribute as the second invalid feature, remove the attribute value in the second invalid feature, and import the standard attribute into the feature attribute of the second invalid feature to obtain the second valid feature; so that the recruitment feature corresponding to the second invalid feature can be expanded to generate a feature that meets subsequent needs;

[0096] The first effective feature, the second effective feature and the remaining recruitment features are recorded as retrieval features, and m retrieval features are obtained.

[0097] When the search features are expanded, the original search data that matches the search features can be retrieved from the large database based on the search features, so that the original search data can correspond to the recruitment information on the user side as a whole and be provided to the user side for preliminary screening;

[0098] The retrieval methods for raw retrieval data include:

[0099] The historical data in the large database is equally divided into m data sets, and the m data sets are randomly matched with the m search features one by one; by equally dividing into data sets, it can be ensured that the number of historical data in each data set is consistent, and there is no phenomenon that the number of historical data in a certain data set is too much or too little. The massive historical data in the large database can be divided into multiple independent sets, achieving the effect of breaking the whole into parts, and also reducing the burden of data retrieval in each data set;

[0100] Mark the data input ports and data output ports of the m data sets respectively to obtain m input ports and m output ports;

[0101] Import m search features into m input ports one by one, select historical data with the same recruitment semantics as the m search features from the m data sets through RAG technology, and compress the retrieved historical data to generate m original data packets; through compression, a large amount of retrieved historical data can be kept together, avoiding the loss or confusion of historical data, and improving the stability of historical data retrieval;

[0102] After the m original data packets are output through the m output ports and decompressed, we can obtain The original search data.

[0103] It should be noted that the original search data is used to match the search features and is used as the object of subsequent data re-screening. Therefore, the amount of original search data is also relatively large to ensure that the original search data can contain all the data associated with the search features.

[0104] The calculation and screening module obtains the comprehensive weight data of the original search data, calculates the recruitment relevance value of the original search data, and screens out the effective search data from the original search data;

[0105] Comprehensive weight data refers to data that can express the correlation between the original search data and the actual recruitment characteristics of the user end, and is used as the basis for calculating the recruitment correlation value, so as to accurately screen out effective search data;

[0106] The comprehensive weight data includes the effective character ratio and effective update duration;

[0107] The effective character ratio value refers to the ratio of the number of characters in the original search data that overlap with the standard characters to the total number of characters, which can be used to represent the correlation between the original search data and the recruitment characteristics. The larger the effective character ratio value, the more characters in the original search data that overlap with the standard characters, the greater the correlation between the original search data and the recruitment characteristics, and the greater the recruitment correlation value;

[0108] Methods for obtaining the effective character ratio include:

[0109] Through word segmentation technology, The original search data is split into original characters and counted The number of original characters in the original search data is obtained Total value;

[0110] The user-preset standard characters are queried from the large database, the original characters that are repeated with the standard characters are recorded as valid characters, and the number of valid characters is counted to obtain A valid value; the preset standard character refers to the character preset by the user end to represent the valid character and serves as the basis for valid character recognition;

[0111] Will The valid values ​​are The total value is compared to obtain Percentage of valid characters;

[0112] The expression for the percentage of valid characters is:

[0113] ;

[0114] In the formula, For the The percentage of valid characters in the original search data. For the The valid value of the original search data, For the The total value of the original search data.

[0115] The effective update time refers to the time length of the collection process of the original search data into the big database, which can be used to indicate the agility of the collection process of the original search data. The longer the effective update time is, the longer the collection process of the original search data into the big database is, and the smaller the agility of the original search data to the recruitment characteristics is, and the smaller the recruitment correlation value is. The effective update time is obtained by querying the time length between the first data update time and the last data update time in the original search data through the data update system.

[0116] When the effective character ratio and effective update time are obtained, the recruitment correlation value indicating the degree of correlation between the original search data and the recruitment characteristics can be calculated according to the effective character ratio and effective update time, and the recruitment correlation value can be used as a screening basis for the original search data;

[0117] The expression of recruitment association value is:

[0118] ;

[0119] In the formula, For the Recruitment related values ​​of the original search data, For the The effective update time of the original search data, , is a proportional coefficient greater than 0;

[0120] in, , exemplary, is 0.52, is 0.48; , The setting is to balance the proportion of valid characters and the proportion of effective update time in the recruitment association value, which can ensure that changes in the proportion of valid characters and the effective update time can cause corresponding changes in the recruitment association value.

[0121] After the recruitment correlation value is calculated, it is necessary to judge and screen whether the original search data meets the subsequent recruitment special needs based on the size of the recruitment correlation value, so as to screen out the effective search data with a greater correlation with the recruitment characteristics from the original search data with a larger coverage;

[0122] Filtering methods to effectively retrieve data include:

[0123] Will Recruitment related value Sequentially associated with the preset recruitment threshold Comparison; The preset recruitment relevance threshold is used to limit the minimum value of the recruitment relevance value recorded as effective search data, that is, to limit the critical data of the degree of relevance between the original search data and the recruitment characteristics; The preset recruitment relevance threshold is obtained by collecting a large number of minimum values ​​of recruitment relevance values ​​recorded as effective search data in history and calculating their average value;

[0124] when Greater than or equal to , explain The recruitment relevance value of the original search data is greater than or equal to the preset recruitment relevance threshold. If the correlation between the first original search data and the recruitment characteristics is greater, the first The original search data are recorded as effective search data, and q effective search data are obtained;

[0125] when Less than , explain The recruitment relevance value of the original search data is less than the preset recruitment relevance threshold. If the correlation between the first original search data and the recruitment characteristics is small, the first The original search data is recorded as the valid search data, and there is no valid search data.

[0126] The second retrieval module trains a retrieval recognition model through fine-tuning technology, and identifies target data corresponding to recruitment characteristics through the retrieval recognition model;

[0127] The retrieval recognition model refers to the ability to use recruitment features as retrieval input conditions and identify target data corresponding to the recruitment features from valid retrieval data. With the help of fine-tuning technology, it can train a special retrieval recognition model to meet the information retrieval of different recruiters according to the specific restrictions and requirements of the user end for recruiters.

[0128] The target data refers to the data corresponding to the recruitment features identified from the effective search data, and is used as the output of the search and recognition model, so as to be the information of the recruiters corresponding to the different recruitment features of the user end, so as to meet the recruitment needs of the user end; the target data is obtained by collecting a large number of historical recruitment features of the user end and the data received by the user end corresponding to the effective search data;

[0129] The training methods for the retrieval recognition model include:

[0130] Aggregate recruitment features and effective retrieval data to generate a feature data set, and convert the feature data set into a corresponding set of feature vectors; construct the feature data set and the corresponding target data into a sample set of the retrieval recognition model, and divide the sample set into training samples and test samples, where the training samples account for 70% of the sample set and the test samples account for 30% of the sample set;

[0131] The feature vectors in the training samples are used as the input of the retrieval recognition model, and the target data corresponding to each set of feature data sets are used as the output of the retrieval recognition model. The target data is used as the prediction target, and the sum of the prediction errors of all training samples is minimized as the training target. The retrieval recognition model is trained, and the retrieval recognition model is tested with test samples until the sum of the prediction errors reaches convergence, and the training is stopped to obtain a retrieval recognition model that can predict the target data.

[0132] Exemplarily, the retrieval recognition model is any one of a CNN neural network model or AlexNet;

[0133] Other model parameters of the retrieval recognition model, target loss value, optimization algorithm, ratio of training samples to test samples, and optimization of loss function are all achieved through actual engineering and obtained through continuous experimental tuning.

[0134] After the retrieval and recognition model is trained, the user's real-time recruitment features and effective retrieval data can be input into the retrieval and recognition model, so as to retrieve and identify the corresponding z target data. At this time, the z target data can be used as recruitment information corresponding to the user's recruitment features, and as data subsequently provided to the user for display.

[0135] The recruitment report module constructs a basic report form with a report position, and imports the valid search data and target data into the report position respectively to generate a recruitment report form;

[0136] After the target data is identified, the automatic recruitment robot automatically retrieves the effective search data and target data related to the recruitment characteristics according to the recruitment needs of the user. Since the effective search data is used to meet the overall recruitment needs of the user, the scope of recruiters covered is relatively large, while the target data is used to meet the specific recruitment needs of the user, and the scope of recruiters covered is relatively small, it is necessary to distinguish and display the effective search data and the target data;

[0137] The basic report sheet is used to provide an original and basic display structure for the data identified by the search, and to make corresponding optimization changes to the basic report sheet according to the different types of data identified by the search, so as to meet the requirements of importing effective search data and target data;

[0138] The reporting position includes a first reporting position and a second reporting position;

[0139] The basic report form construction method includes:

[0140] Mark all report form templates in the report form database one by one, identify the template attributes of all report form templates, and record the report form templates with template attributes as valid templates as valid report forms, and obtain t valid report forms; the report form template refers to a blank report form without any substantial data in the report form database, and the template attribute is used to represent the attribute of the report form template and to determine whether the report form template can be used subsequently. Specifically, the template attribute includes a valid template and an invalid template;

[0141] The generation time of t valid report sheets is queried one by one through the timestamp, and after arranging the t generation time in chronological order, the valid report sheet corresponding to the last generation time is selected as the target report sheet;

[0142] Draw two report positions arranged vertically in the target report form, and record the two report positions as the first report position and the second report position from top to bottom to obtain a basic report form. The report position is used to provide a location for data import for different data in the report form to ensure that different types of data in the report form can remain relatively independent;

[0143] After the basic report form is generated, the effective search data and the target data can be correspondingly imported into the first report position and the second report position in the basic report form, so that the basic report form can be filled with data, so that the basic report form can generate a recruitment report form and meet the recruitment needs of the user end, thereby realizing the automatic retrieval and recognition effect of the automatic recruitment robot on the recruitment information, and improving the recruitment efficiency of the user end;

[0144] The methods for generating a recruitment report sheet include:

[0145] Draw q independent data units in the first report position to obtain q first units; the data unit is used to indicate the position of a specific data in the report sheet, and ensure the independence of each specific data in the report sheet, and also facilitate the subsequent partitioning and display of the data in the report sheet;

[0146] Import q valid search data into q first units one by one, and annotate the keyword "valid" in the first report position after the import of the valid search data;

[0147] Drawing z independent data units in the second reporting position to obtain z second units;

[0148] Import z target data into z second units one by one, and annotate the keyword "target" in the second report position after importing the target data;

[0149] Draw a table unit on the recruitment report form above the first report position, and note the keyword "recruitment report" in the table unit to generate a recruitment report form.

[0150] In this embodiment, by acquiring recruitment data from the user end, extracting features to be identified from the recruitment data, and identifying recruitment features from the features to be identified, based on the extension criteria, the recruitment features are expanded into retrieval features, and the original retrieval data corresponding to the retrieval features are retrieved through the RAG technology, the comprehensive weight data of the original retrieval data is obtained, the recruitment association value of the original retrieval data is calculated, and the effective retrieval data is screened from the original retrieval data, the retrieval recognition model is trained through the fine-tuning technology, and the target data corresponding to the recruitment features are identified through the retrieval recognition model, a basic report form with a reporting position is constructed, and the effective retrieval data and the target data are respectively imported into the reporting position to generate a recruitment report form; compared with the prior art, by expanding the recruitment features into retrieval features, The retrieval features can not only accurately represent the recruitment needs of the user side, but also provide reasonable and clear retrieval standards for RAG technology, that is, to retrieve the original retrieval data that initially meets the recruitment needs of the user side and has a large coverage range, and combine the comprehensive proportion data to calculate the recruitment correlation value, and screen out effective retrieval data with a small coverage range from the original retrieval data, and through fine-tuning technology, train a retrieval recognition model that can identify target data for specific and special recruitment needs, which meets the precise and fast retrieval and recognition operations of data that meets the specific recruitment needs of the user side, and facilitates the human resources department of the enterprise to identify suitable target data from massive data according to recruitment needs, thereby improving the two-way interaction effect of data between enterprises and personnel, and effectively improving the recruitment success rate of enterprises.

[0151] Example 2: Please refer to Figure 2 As shown, the part not described in detail in this embodiment is described in the first embodiment, which provides an automatic recruitment robot method based on a large model RAG and fine-tuning, which is applied to a retrieval and recognition server and is implemented based on an automatic recruitment robot system based on a large model RAG and fine-tuning, including:

[0152] S1: Obtain recruitment data from the user, extract features to be identified from the recruitment data, and identify recruitment features from the features to be identified;

[0153] S2: Based on the extension criteria, the recruitment features are extended into retrieval features, and the original retrieval data corresponding to the retrieval features are retrieved through the RAG technology. The extension criteria are: the structure value and attribute value of the retrieval feature are extended into the standard structure and standard attribute respectively;

[0154] S3: Obtain comprehensive weight data of the original search data, the comprehensive weight data including the effective character ratio and effective update time, calculate the recruitment relevance value of the original search data, and filter out effective search data from the original search data;

[0155] S4: Train the retrieval recognition model through fine-tuning technology, and identify the target data corresponding to the recruitment characteristics through the retrieval recognition model;

[0156] S5: Construct a basic report form with a reporting position, the reporting position includes a first reporting position and a second reporting position, and import the valid search data and the target data into the reporting position respectively to generate a recruitment report form.

[0157] Example 3: Please refer to Figure 3 As shown, this embodiment discloses an electronic device, including a processor and a memory;

[0158] Wherein, the memory stores a computer program that can be called by the processor;

[0159] The processor executes the automatic recruitment robot method based on large model RAG and fine-tuning by calling the computer program stored in the memory.

[0160] Since the electronic device introduced in this embodiment is an electronic device used to implement an automatic recruitment robot method based on a large model RAG and fine-tuning in Example 2 of this application, based on the automatic recruitment robot method based on a large model RAG and fine-tuning introduced in the embodiment of this application, a person skilled in the art can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application will not be described in detail here. As long as a person skilled in the art implements an electronic device used in an automatic recruitment robot method based on a large model RAG and fine-tuning in the embodiment of this application, it belongs to the scope of protection of this application.

[0161] Example 4: Please refer to Figure 4 As shown, this embodiment discloses a computer-readable storage medium on which a rewritable computer program is stored;

[0162] When the computer program is executed, the automatic recruitment robot method based on large model RAG and fine-tuning is implemented.

[0163] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An automatic recruitment robot system based on large model RAG and fine-tuning, applied to retrieval and recognition server, characterized in that: include: A feature recognition module is used to obtain recruitment data from the user end, extract features to be recognized from the recruitment data, and recognize recruitment features from the features to be recognized; The first retrieval module is used to expand the recruitment features into retrieval features based on the expansion criteria, and retrieve the original retrieval data corresponding to the retrieval features through the RAG technology. The expansion criteria are: expanding the structure value and the attribute value of the retrieval feature into a standard structure and a standard attribute respectively; Extension methods for retrieving features include: Through the structure management system and attribute management system, the structure and attributes of historical data in the big database are queried respectively to obtain the standard structure and standard attributes; The structures and attributes of m recruitment features are queried one by one through the structure management system and the attribute management system, recorded as feature structures and feature attributes, and the structure values ​​of the feature structures and the attribute values ​​of the feature attributes are marked to obtain m structure values ​​and m attribute values; Recording the recruitment feature whose structure value is not the standard structure as the first invalid feature, eliminating the structure value in the first invalid feature, and importing the standard structure into the feature structure of the first invalid feature to obtain the first valid feature; Recording the recruitment feature whose attribute value is not the standard attribute as the second invalid feature, eliminating the attribute value in the second invalid feature, and importing the standard attribute into the feature attribute of the second invalid feature to obtain the second valid feature; The first effective feature, the second effective feature and the remaining recruitment features are recorded as retrieval features to obtain m retrieval features; The retrieval methods for raw retrieval data include: Divide the historical data in the large database into m data sets, and randomly correspond the m data sets to the m search features one by one; Mark the data input ports and data output ports of the m data sets respectively to obtain m input ports and m output ports; Import m search features into m input ports one by one, filter out historical data with the same recruitment semantics as the m search features from the m data sets through RAG technology, and compress the retrieved historical data to generate m original data packets; After the m original data packets are output through the m output ports and decompressed, we can obtain Original search data; The calculation and screening module is used to obtain the comprehensive weight data of the original search data, which includes the effective character ratio and the effective update time, calculate the recruitment relevance value of the original search data, and screen out the effective search data from the original search data; Methods for obtaining the effective character ratio include: Through word segmentation technology, The original search data is split into original characters and counted The number of original characters in the original search data is obtained Total value; The user-preset standard characters are queried from the large database, the original characters that are repeated with the standard characters are recorded as valid characters, and the number of valid characters is counted to obtain Valid values; Will The valid values ​​are The total value is compared to obtain Percentage of valid characters; The expression for the percentage of valid characters is: ； In the formula, For the The percentage of valid characters in the original search data. For the The valid value of the original search data, For the The total value of the original search data: The expression of recruitment association value is: ； In the formula, For the Recruitment related values ​​of the original search data, For the The effective update time of the original search data, , is a proportional coefficient greater than 0; Filtering methods to effectively retrieve data include: Will Recruitment related value Sequentially associated with the preset recruitment threshold Compare; when Greater than or equal to When The original search data are recorded as effective search data, and q effective search data are obtained; when Less than When The original search data are recorded as effective search data; The second retrieval module is used to train a retrieval recognition model through fine-tuning technology, and identify target data corresponding to the recruitment characteristics through the retrieval recognition model; The recruitment report module is used to construct a basic report form with a reporting position, which includes a first reporting position and a second reporting position, and import the valid search data and the target data into the reporting position respectively to generate a recruitment report form.

2. According to claim 1, an automatic recruitment robot system based on large model RAG and fine-tuning is characterized in that: The method for extracting the features to be identified includes: Mark all the stored data in the enterprise database one by one, and query the IP numbers of all the stored data and the IP numbers of the user terminals one by one through the address management system, record the stored data whose IP numbers are consistent with the IP numbers of the user terminals as recruitment data, and obtain i recruitment data; Record job title, education, salary, region and benefits as standard semantics; The recruitment semantics of i recruitment data are identified one by one through natural language processing technology, and the recruitment semantics with any standard semantics are recorded as target semantics to obtain p target semantics; The recruitment data corresponding to the p target semantics are split one by one by using word segmentation technology to generate p character lines, and the characters in the p character lines are numbered in sequence; The numbers corresponding to the p target semantic characters are recorded as target numbers, and the two numbers adjacent to the target number are recorded as adjacent numbers; After summarizing the characters corresponding to the target number and the adjacent numbers in sequence, p features to be recognized are obtained.

3. The automatic recruitment robot system based on large model RAG and fine-tuning according to claim 2 is characterized in that: The method for identifying the recruitment characteristics includes: Through the permission management system, query the permission remark boxes of p features to be identified one by one, and mark the form remark value and the permission remark value in the permission remark box; The features to be identified whose form remark value and permission remark value are both high are recorded as features to be verified, and s features to be verified are obtained; The features to be verified that have repeated target semantics are recorded as repeated features, and the recording times of the target semantics in the x repeated features are queried one by one through the timestamps, which are recorded as x recording times; Arrange the x recorded moments in chronological order, remove duplicate features except the first recorded moment, and record the remaining duplicate features and features to be verified as recruitment features to obtain m recruitment features.

4. The automatic recruitment robot system based on large model RAG and fine-tuning according to claim 3 is characterized in that: The training method of the retrieval recognition model includes: Aggregate recruitment features and effective retrieval data to generate a feature data set, convert the feature data set into a corresponding set of feature vectors, construct the feature data set and the corresponding target data into a sample set of the retrieval recognition model, and divide the sample set into training samples and test samples; The feature vectors in the training samples are used as the input of the retrieval and recognition model, and the target data corresponding to each set of feature data sets is used as the output of the retrieval and recognition model. The target data is used as the prediction target, and minimizing the sum of prediction errors of all training samples is used as the training target. The retrieval and recognition model is trained and tested with test samples until the sum of prediction errors converges. The training is stopped to obtain a retrieval and recognition model that can predict the target data.

5. The automatic recruitment robot system based on large model RAG and fine-tuning according to claim 4 is characterized in that: The method for constructing the basic report sheet includes: Mark all report form templates in the report form database one by one, identify the template attributes of all report form templates, and record the report form templates with the template attribute as valid templates as valid reports, and obtain t valid report forms; The generation time of t valid report sheets is queried one by one through the timestamp, and after arranging the t generation time in chronological order, the valid report sheet corresponding to the last generation time is selected as the target report sheet; Draw two report positions arranged vertically in the target report form, and record the two report positions as the first report position and the second report position from top to bottom to obtain a basic report form.

6. The automatic recruitment robot system based on large model RAG and fine-tuning according to claim 5 is characterized in that: The method for generating the recruitment report sheet includes: Draw q independent data units in the first reporting position to obtain q first units; Import q valid search data into q first units one by one, and annotate the keyword "valid" in the first report position after the import of the valid search data; Drawing z independent data units in the second reporting position to obtain z second units; Import the z target data into the z second units one by one, and annotate the keyword "target" in the second report position after the target data is imported; Draw a table unit on the recruitment report form above the first report position, and note the keyword "recruitment report" in the table unit to generate a recruitment report form.

Images