Scientific and technological talent cultivation model construction method based on multivariate reinforcement learning

Abstract

The invention relates to a scientific and technological talent cultivation model construction method based on multivariate reinforcement learning, and the method comprises the steps: S1, an employer calls talents in all needed fields, enables the talents to complete personnel resumes set for all the fields, and obtains a resume evaluation result; s2, enabling each talent to wear an electroencephalogram device in post work of the field of the employer where the talent is located, and collecting electroencephalogram data within a preset time period to obtain an electroencephalogram evaluation result; s3, collecting an expert resume, making an expert answer sheet by an employment unit, extracting expert description from the expert answer sheet through a pre-natural language model, and obtaining whether the expert description accords with the employment standard of the unit or not through a pre-trained expert model so as to determine whether to select an expert for evaluating the talent or not; and S4, setting and constructing an intelligent personnel model of multivariate reinforcement learning. The ability change of talents can be reflected more objectively, accurately and dynamically through multivariate reinforcement learning of personnel resumes, expert evaluation and electroencephalogram evaluation.

Description

Technical Field

[0001] The present invention relates to a scientific and technological talent training model construction, and in particular to a scientific and technological talent training model construction method based on multivariate reinforcement learning, and belongs to the field of intelligent talent models. Background Art

[0002] The personnel model refers to the selection of talents and the prediction of matching positions. Generally, the resume of a talent can be understood. The existing reinforcement learning only maps the resume of a talent to the knowledge graph, and does not give the specific composition of the knowledge graph in detail. In addition, the resume of a talent cannot fully evaluate a talent.

[0003] EEG is a means of evaluating a person's performance when facing difficulties, challenges, or calmness. Through real-time computer testing, it can determine the performance of different talents when facing specific things. However, EEG cannot fully determine a person's true inner subtle feelings. Therefore, how to describe the relationship between EEG and the real thoughts of the person being tested when facing various things is a better way to reflect the talent's practical process.

[0004] In addition, the personal experiences or research preferences of different experts may also lead to deviations in the evaluation of the same person. How to conduct more matched expert evaluations based on employer expectations and find suitable experts is a refined talent evaluation. Summary of the invention

[0005] In view of the above-mentioned problems of the prior art, the present invention will consider the following aspects for reinforcement learning: first, consider the design of a unified form of personnel resume sub-items, adopt the resume items and question-answering forms preferred by employers; second, comprehensively consider the three aspects of EEG processing results, the real description of the person being tested, and the results of handling things, and conduct a comprehensive evaluation of the EEG sub-item; third, collect expert resumes, and employers formulate expert answers. Through the model construction of expert sub-items, an expert optimization prediction model that meets the employer's requirements is formed; fourth, the above three sub-items are summed up and modeled to complete multivariate reinforcement training.

[0006] Based on the above considerations, one aspect of the present invention provides a method for constructing a scientific and technological talent training model based on multivariate reinforcement learning, comprising the following steps: S1 The employer gathers talents in the required fields and asks them to complete the personnel resumes set for each field. The employer uses a pre-defined natural language model to extract the description of the personnel resume, and then obtains the resume evaluation results based on the description and the pre-trained resume prediction model. S2 requires each talent to wear an EEG device while working in the field of the employer, collects EEG data within a preset time period, and then collects the wearer's description of his / her feelings during the work, and further extracts the description of feelings based on the pre-trained natural language model, and then obtains the EEG evaluation result based on the description of feelings and the pre-trained EEG prediction model; S3 collects expert resumes, and the employer prepares expert answers. The employer uses a pre-natural language model to extract expert descriptions from the expert answers. The employer uses a pre-trained expert model to determine whether the expert meets the employer's employment standards, so as to decide whether to select the expert to conduct the talent evaluation. S4 builds an intelligent personnel model based on multi-reinforcement learning ,in They are the pre-trained resume prediction model, EEG prediction model, and expert evaluation. is the weighting coefficient and .

[0007] Optionally, the description of the personnel resume may include academic qualifications, work experience, projects undertaken, awards received, and answers to test questions set by the employer.

[0008] Preferably, the test questions include multiple-choice questions and true-or-false questions, and the employer scores the answers to the test questions based on the standard answers for the job positions in various fields that the employer needs, and also scores the academic qualifications, work experience, projects undertaken, and awards received.

[0009] It is easy to understand that a scoring principle may include: the higher the academic qualifications, the higher the score; those who have worked in famous companies in their work experience will score higher; those who have worked in the same company for longer periods of time will score higher; those who have worked in different companies but not for more than 1 or 2 years in each company will have appropriate points deducted or very few points added; the more projects they have undertaken, the higher the score, and there will be more key projects.

[0010] Optionally, resume prediction model The pre-training method includes: S1-1 Based on the description of the personnel resume Construct a mapping to form the corresponding evaluation value , expressed as In fact, it is to establish arrive Pointer to .

[0011] S1-2 builds a multi-class support vector machine (Multi-class SVM) to describe Personnel resume description objects in each setting Manual scoring , Number the objects and count them to form a mapping description , and according to each Get the corresponding manual evaluation value ,Right now ,in and They are manual scoring rules and pre-set manual evaluation algorithms; specifically they can be arithmetic mean, weighted mean, etc.

[0012] S1-3 To make every talent Divide into training set and validation set , the training set is used for multi-class support vector machine , and get the predicted evaluation value ,use right Verify and define the accuracy as And the loss function , where the accuracy is a subset of the validation set of all correct predictions The model and The modulus ratio is optimized using the loss function Parameters, when , both , then stop training, It is a natural number indicating the number of optimization parameters.

[0013] It should be understood that the manual evaluation results of daily human resource resume screening are used as data for multi-category support vector machine training, so that the description interpreted by the natural language model is mapped into a machine evaluation that simulates manual work through the multi-category support vector machine.

[0014] Optionally, describe The description of the project in the personnel resume is the natural language description extracted by the natural language model NLP based on the resume, according to the preset rules For scoring, consider the natural language model NLP, and the resume prediction model is written as , For a resume of talent, , thus obtaining the resume evaluation results.

[0015] For example, employers can give scores to judgment questions. For example, if the standard answer is correct and the talent also answers "yes" in his / her resume, then the NLP analysis shows "yes" and the corresponding score will be given, otherwise no score will be given.

[0016] Optionally, descriptions of work feelings include six categories: underperforming, subpar, average, exciting, exceptional, and uncertain.

[0017] Optionally, the EEG data includes a time t-frequency freq spectrum within a preset time period.

[0018] Optionally, the training process of the EEG prediction model includes the following steps: S2-1 collects multiple talents with poor, average, normal, excellent and leading job performance in various fields, and asks them to wear EEG devices to work during a preset time period; S2-2 For each type of performance, a convolutional neural network model CNN is constructed. The EEG spectra of multiple talents in each type of performance are divided into a training set and a validation set, with the ratio of the two being (6-5):(1-3). The training set is used to train CNN, and the softmax function is used to predict the six types of descriptions. The validation set is used to verify the accuracy, and the loss function is calculated to optimize the CNN parameters. When the loss function stabilizes, the training is stopped, thus obtaining a set of five models They represent poor, average, normal, excellent, and leading respectively; S2-3: Let underperformance, poor performance, normal performance, excitement, extraordinary performance, and uncertainty be described as numerical functions , , respectively representing the six types of descriptions, assuming that the actual work experience of talents is numerically expressed , the measured EEG spectrum of talents is ,but is the weighting coefficient, and .

[0019] Therefore, we use digital functions to quantify the feelings of each performance category and perform a weighted average with the numerical values ​​of the talents' real feelings to obtain a more objective description. If the internal score is significantly higher than the classified performance level, it indicates that the talent is making dynamic progress and showing a positive side, otherwise the performance is regressing. Therefore, through comprehensive consideration of EEG and the oral description of the talent, the present invention highlights the fact that the actual performance of the current work is superimposed on the past resume history, forming a dynamic evaluation based on the resume. Therefore, the current actual work ability will not be ignored because of a mediocre resume, nor will the temporary poor practical ability of the talent be ignored because the resume seems excellent. The emphasis on the current and temporary here also indicates a dynamic evaluation.

[0020] Optionally, the preset time period includes monthly, quarterly, semi-annual, or annual.

[0021] In order to prevent the resume from being exaggerated, the present invention also considers the industry evaluation sub-item of the expert.

[0022] Optionally, the expert answer sheet includes the expert's academic qualifications, graduation school, areas of expertise, published papers, published works, projects hosted or participated in, and at least one multiple-choice question and / or true-or-false question set by the employer.

[0023] The expert model also uses a multi-class support vector machine. The training process includes: S3-1 Screen candidates based on expert resumes and use NLP to extract expert descriptions Construct a mapping to form the corresponding evaluation value , respectively, are selected and excluded, and the preset mapping is expressed as ; S3-2 builds a multi-class support vector machine (Multi-class SVM) to describe The expert answer sheet description object for each setting Manual scoring , Number the objects and count them to form a mapping description , and according to each Get the corresponding manual evaluation value ,Right now ,in and They are manual scoring rules and pre-set manual evaluation algorithms; specifically they can be arithmetic mean, weighted mean, etc.

[0024] S3-3 Each expert’s Divide into training set and validation set , the training set is used for multi-class support vector machine , and get the predicted evaluation value ,use right Verify and define the accuracy as And the loss function , where the accuracy is a subset of the validation set of all correct predictions The model and The modulus ratio is optimized using the loss function Parameters, when , then stop training, It is a natural number indicating the number of optimization parameters.

[0025] Optionally, It is to select the top 3-5 experts with the highest predicted evaluation value or the top 3-5 experts from the candidates based on the multi-category support vector machine to replace the manual evaluation value. ,thus .

[0026] Another object of the present invention is to provide a device for constructing an intelligent science and technology talent model based on multivariate reinforcement learning, including a computer system having a non-temporary memory in which a computer system readable program capable of implementing the above method is stored.

[0027] The solution of the present invention has the following beneficial effects: Through NLP, CNN, and multi-category support vector machines, we conduct expert evaluation on personnel resumes and expert selection, as well as multivariate reinforcement learning of EEG evaluation, thus providing a dynamic evaluation of talents. This can more objectively, accurately, and dynamically reflect the changes in talents' abilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The main flow chart of the method for constructing a scientific and technological talent training model based on multivariate reinforcement learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] Figure 1 The main process of the construction method of the scientific and technological talent training model based on multivariate reinforcement learning is given, which includes two parts. The first part corresponds to S1 and S3, and the other part includes step S2, where S1 and S3 are for talent resume evaluation and expert selection respectively. Due to the similarity of the algorithms, the superscript e of the formula symbol in the multi-category support vector machine used for expert selection is omitted. The specific steps include the following: Resume prediction model The pre-training method includes: S1-1 Based on the description of the personnel resume Construct a mapping to form the corresponding evaluation value , expressed as ; S1-2 builds a multi-class support vector machine (Multi-class SVM) to describe Personnel resume description objects in each setting Manual scoring , Number the objects and count them to form a mapping description , and according to each Get the corresponding manual evaluation value ,Right now ,in and They are manual scoring rules and pre-set manual evaluation algorithms; specifically they can be arithmetic mean, weighted mean, etc.

[0030] S1-3 To make every talent Divide into training set and validation set , the training set is used for multi-class support vector machine , and get the predicted evaluation value ,use right Verify and define the accuracy as And the loss function , where the accuracy is a subset of the validation set of all correct predictions The model and The modulus ratio is optimized using the loss function Parameters, when , both , then stop training, It is a natural number indicating the number of optimization parameters.

[0031] The expert model also uses a multi-class support vector machine. The training process includes: S3-1 Screen candidates based on expert resumes and use NLP to extract expert descriptions Construct a mapping to form the corresponding evaluation value , respectively, are selected and excluded, and the preset mapping is expressed as ; S3-2 builds a multi-class support vector machine (Multi-class SVM) to describe The expert answer sheet description object for each setting Manual scoring , Number the objects and count them to form a mapping description , and according to each Get the corresponding manual evaluation value ,Right now ,in and They are manual scoring rules and pre-set manual evaluation algorithms; specifically they can be arithmetic mean, weighted mean, etc.

[0032] S3-3 Each expert’s Divide into training set and validation set , the training set is used for multi-class support vector machine , and get the predicted evaluation value ,use right Verify and define the accuracy as And the loss function , where the accuracy is a subset of the validation set of all correct predictions The model and The modulus ratio is optimized using the loss function Parameters, when , then stop training, It is a natural number indicating the number of optimization parameters.

[0033] The second part S2 involves the training process of the EEG prediction model, which includes the following steps: S2-1 collects multiple talents with poor, average, normal, excellent and leading job performance in various fields, and asks them to wear EEG devices to work every quarter; S2-2 For each type of performance, a convolutional neural network model CNN is constructed. The EEG spectra of multiple talents in each type of performance are divided into a training set and a validation set, with a ratio of 5:2. The training set is used to train CNN, and the softmax function is used to predict the six types of descriptions. The validation set is used to verify the accuracy, and the loss function is calculated to optimize the CNN parameters. When the loss function stabilizes, the training is stopped, thus obtaining a set of five models They represent poor, average, normal, excellent, and leading respectively; S2-3: Let underperformance, poor performance, normal performance, excitement, extraordinary performance, and uncertainty be described as numerical functions , respectively representing the six types of descriptions, assuming that the actual work experience of talents is numerically expressed , the measured EEG spectrum of talents is ,but is the weighting coefficient, and In this embodiment .

[0034] Finally, step S4 is given to build an intelligent personnel model of multivariate reinforcement learning , in They are the pre-trained resume prediction model, EEG prediction model, and expert evaluation. is the weighting coefficient and In this embodiment .

[0035] An embodiment of the present invention also provides a device for constructing an intelligent science and technology talent model based on multivariate reinforcement learning, including a computer system, which can be an enterprise-level system or a network server system. The computer system has a non-temporary memory in which a computer system-readable program capable of implementing the above method is stored.

[0036] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A method for constructing a scientific and technological talent training model based on multivariate reinforcement learning, characterized in that: The steps include: S1 The employer gathers talents in the required fields and asks them to complete the personnel resumes set for each field. The employer uses a pre-defined natural language model to extract the description of the personnel resume, and then obtains the resume evaluation results based on the description and the pre-trained resume prediction model. S2 requires each talent to wear an EEG device while working in the field of the employer, collects EEG data within a preset time period, and then collects the wearer's description of his / her feelings during the work, and further extracts the description of feelings based on the pre-trained natural language model, and then obtains the EEG evaluation result based on the description of feelings and the pre-trained EEG prediction model; S3 collects expert resumes, and the employer prepares expert answers. The employer uses a pre-natural language model to extract expert descriptions from the expert answers. The employer uses a pre-trained expert model to determine whether the expert meets the employer's employment standards, so as to decide whether to select the expert to conduct the talent evaluation. S4 builds an intelligent personnel model based on multi-reinforcement learning ,in They are the pre-trained resume prediction model, EEG prediction model, and expert evaluation. is the weighting coefficient and .

2. The method according to claim 1, characterized in that The description of a personnel resume includes academic qualifications, work experience, projects undertaken, awards received, and answers to test questions set by the employer.

3. The method according to claim 1, characterized in that The test questions include multiple-choice questions and true-or-false questions. Employers score the answers to the test questions based on the standard answers for the job positions in each field they need, and also give scores to academic qualifications, work experience, projects undertaken, and awards received.

4. The method according to claim 3, characterized in that: Resume prediction model The pre-training method includes: S1-1 Based on the description of the personnel resume Construct a mapping to form the corresponding evaluation value , expressed as ; S1-2 builds a multi-class support vector machine (Multi-class SVM) to describe Personnel resume description objects in each setting Manual scoring , Number the objects and count them to form a mapping description , and according to each Get the corresponding manual evaluation value ,Right now ,in and They are manual scoring rules and pre-set manual evaluation algorithms; S1-3 To make every talent Divide into training set and validation set , the training set is used for multi-class support vector machine , and get the predicted evaluation value ,use right Verify and define the accuracy as And the loss function , where the accuracy is a subset of the validation set of all correct predictions The model and The modulus ratio is optimized using the loss function Parameters, when , both , then stop training, It is a natural number indicating the number of optimization parameters.

5. The method according to claim 4, characterized in that describe The description of the project in the personnel resume is the natural language description extracted by the natural language model NLP based on the resume, according to the preset rules For scoring, consider the natural language model NLP, and the resume prediction model is written as , For a resume of talent, , thus obtaining the resume evaluation results.

6. The method according to any one of claims 1 to 5, characterized in that The descriptions of work feelings include six categories: underperformance, poor performance, average performance, excitement, extraordinary performance, and uncertainty. The EEG data include the time t-frequency freq spectrum within a preset time period.

7. The method according to claim 6, characterized in that The training process of the EEG prediction model includes the following steps: S2-1 collects multiple talents with poor, average, normal, excellent and leading job performance in various fields, and asks them to wear EEG devices to work during a preset time period; S2-2 For each type of performance, a convolutional neural network model CNN is constructed. The EEG spectra of multiple talents in each type of performance are divided into a training set and a validation set, with the ratio of the two being (6-5):(1-3). The training set is used to train CNN, and the softmax function is used to predict the six types of descriptions. The validation set is used to verify the accuracy, and the loss function is calculated to optimize the CNN parameters. When the loss function stabilizes, the training is stopped, thus obtaining a set of five models They represent poor, average, normal, excellent, and leading respectively; S2-3: Let underperformance, poor performance, normal performance, excitement, extraordinary performance, and uncertainty be described as numerical functions , respectively representing the six types of descriptions, assuming that the actual work experience of talents is numerically expressed , the measured EEG spectrum of talents is ,but is the weighting coefficient, and .

8. The method according to claim 7, characterized in that The preset time periods include monthly, quarterly, semi-annual, or annual; the expert answers include the expert's academic qualifications, graduate school, areas of expertise, published papers, published works, projects hosted or participated in, and at least one multiple-choice question and / or judgment question set by the employer.

9. The method according to claim 7 or 8, characterized in that: The expert model also uses a multi-class support vector machine. The training process includes: S3-1 Screen candidates based on expert resumes and use NLP to extract expert descriptions Construct a mapping to form the corresponding evaluation value , respectively, are selected and excluded, and the preset mapping is expressed as ; S3-2 builds a multi-class support vector machine (Multi-class SVM) to describe The expert answer sheet description object for each setting Manual scoring , Number the objects and count them to form a mapping description , and according to each Get the corresponding manual evaluation value ,Right now ,in and They are manual scoring rules and pre-set manual evaluation algorithms; S3-3 Each expert’s Divide into training set and validation set , the training set is used for multi-class support vector machine , and get the predicted evaluation value ,use right Verify and define the accuracy as And the loss function , where the accuracy is a subset of the validation set of all correct predictions The model and The modulus ratio is optimized using the loss function Parameters, when , then stop training, It is a natural number indicating the number of optimization parameters.

10. The method according to claim 9, characterized in that It is to select the top 3-5 experts with the highest predicted evaluation value or the top 3-5 experts from the candidates based on the multi-category support vector machine to replace the manual evaluation value. ,thus .

Images