A method and system for setting questions and composing examination papers based on examination paper composition algorithm
Through the paper grouping algorithm and knowledge graph model, test questions that meet the position requirements are generated, which solves the problem of improper setting of the difficulty of the existing technical middle school entrance examination questions, and achieves efficient and accurate recruitment screening results.
Patent Information
- Application Number
- CN202411165026.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-23
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-08-23
AI Technical Summary
It is difficult for the existing technology to quickly, efficiently and accurately generate test questions that meet the position requirements, resulting in improper difficulty setting during the recruitment process and may lead to unqualified talents passing.
Using a paper grouping algorithm method, by obtaining the historical test bank and adding the test bank, the initial test bank is constructed, and the competency model is configured in combination with the knowledge graph and the R-GNN model, a two-way detailed list of questions is generated, and randomly selected questions are selected to form papers. Adjust the difficulty of the test paper to meet the requirements.
It has achieved accurate construction of test questions based on position needs, improved the efficiency and accuracy of the recruitment process, ensured that the difficulty of the test paper meets preset standards, and enhanced the adaptability and practicality of the evaluation method.
Smart Images

Figure CN118691238B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of test paper management for social open recruitment, and in particular to a method and system for setting questions and testing papers based on a test paper setting algorithm. Background Art
[0002] In the process of selecting candidates who meet the job requirements and are an ideal match for enterprises and institutions, examinations are the most frequent and effective means of quickly screening applicants' job competencies, and are an important part of recruitment screening.
[0003] This type of question is highly targeted, highlighting the characteristics of enterprises and institutions in terms of fields, organizations and positions. The evaluation methods are more diverse, focusing on the examination of theoretical knowledge and practical ability. The selection is mainly based on "selecting the best", and the test questions are difficult. The difficulty setting will take into account the job requirements and the competitiveness of the applicants (such as the number of applicants, the source of applicants, etc.). Some enterprises and institutions will not easily lower the standards, and may set a basic passing line (for example, more than 70 points) in terms of admission to ensure that all personnel who enter the field of vision have considerable qualities, and would rather have nothing than have something unqualified.
[0004] Therefore, how to combine the examination dimensions, test points, question types, difficulty coefficients, examination outlines, and the degree of fit with the knowledge and ability quality requirements of the recruitment position to quickly, efficiently and accurately generate test questions is a technical problem that needs to be solved urgently.
[0005] To this end, the present invention provides a method and system for setting questions and composing examination papers based on an examination paper composition algorithm. Summary of the invention
[0006] Based on this, it is necessary to provide a method and system for setting questions and writing papers based on a paper-setting algorithm to address the above technical problems.
[0007] According to a first aspect of the present invention, a method for setting questions and composing papers based on a paper composition algorithm is provided, the method comprising: obtaining a historical question bank and a newly added question bank to construct an initial question bank; obtaining target position recruitment data and newly compiled examination syllabus information in enterprises and institutions, configuring corresponding competency models, and constructing a two-way detailed table of questions based on the configured competency models, wherein the two-way detailed table of questions includes information on the dimensions of ability and quality examinations required for the target position, information on the dimensions of knowledge examinations required for the target position, information on test point requirements, information on knowledge point requirements, information on question type requirements, information on the number of questions required, information on the difficulty of questions required, information on the total difficulty of the test paper required, information on the total score requirements for each question type, information on the score requirements for each question type, and information on the order of question arrangement required; according to the mapping relationship formed between the fields of the two-way detailed table of questions and the initial question bank, a natural language processor is used to process the questions. The test questions that meet the test paper composition requirements are extracted from the initial test question bank by using the random test paper composition algorithm and natural language processing technology, and the test questions are randomly extracted from the candidate test question bank to construct the candidate test papers, taking the examination dimension information, test point requirement information, question type requirement information, question quantity requirement information and question difficulty requirement information of the two-way detailed table of questions as constraints; the total difficulty value of the candidate test papers is calculated, and it is determined whether the total difficulty value of the candidate test papers meets the total difficulty requirement information of the papers; if so, the corresponding test papers are adjusted and generated according to the question type requirement information, question difficulty requirement information, total score requirement information of each question type, score requirement information of the corresponding questions of each question type and the question arrangement order requirement information; if not, some questions in the candidate test papers are adjusted until the total difficulty requirement information of the papers is met.
[0008] In some optional implementation methods of some embodiments, the obtaining of the historical question bank and the newly added question bank and constructing the initial question bank specifically includes: obtaining the historical question bank and the newly added question bank, marking the questions contained in the historical questions and the newly added questions, respectively obtaining the historical question-related information and the newly added question-related information, and constructing the initial question bank based on the historical question-related information and the newly added question-related information.
[0009] In some optional implementation methods of some embodiments, the information related to historical test questions includes historical test question type information, historical test question title information, historical test question answer information, test point information corresponding to the historical test questions, knowledge point information corresponding to the historical test questions and history test question difficulty coefficient information, and the information related to newly added test questions includes newly added test question type information, newly added test question title information, newly added test question answer information, test point information corresponding to the newly added test questions, knowledge point information corresponding to the newly added test questions and new test question difficulty coefficient information.
[0010] In some optional implementations of some embodiments, the process of determining the difficulty coefficient information of the newly added test questions specifically includes: according to the field mapping relationship between the test point information corresponding to the newly added test questions and the corresponding test point information in the historical test question bank, as well as the field mapping relationship between the knowledge point information corresponding to the newly added test questions and the corresponding knowledge point information in the historical test question bank, using natural language processing technology, respectively extracting historical test questions that meet the similarity with the test points of the newly added test questions up to a preset test point threshold and generating a historical test question test point set, and several historical test questions that meet the similarity with the knowledge points of the newly added test questions up to a preset knowledge point threshold and generating a historical test question knowledge point set; assigning values to the historical test question test point set and the historical test question knowledge point set in turn according to preset weights, and calculating the comprehensive similarity values of the historical test question test point set and the historical test question knowledge point set; sorting the comprehensive similarity values from high to low, and selecting the historical test question difficulty coefficient information corresponding to the historical test question with the highest comprehensive similarity value as the difficulty coefficient information of the newly added test questions.
[0011] In some optional implementations of some embodiments, the acquisition of target job recruitment data and new examination syllabus information in enterprises and institutions, and configuration of corresponding competency models specifically include: acquiring historical job recruitment data and historical examination syllabus information in enterprises and institutions, and performing information extraction processing on the historical job recruitment data and historical examination syllabus information, wherein the content of the information extraction processing includes: extracting relevant entities and relationship information between relevant entities, wherein the relevant entities include positions, competences and qualities required for positions, knowledge required for positions, and historical examination point requirement information; constructing a knowledge graph of the relationship between positions, competences and qualities required for positions, knowledge required for positions, and historical examination point requirement information; inputting the knowledge graph into a preset In the initial R-GNN model, the knowledge graph is encoded by the graph encoder in the initial R-GNN model to output the related entity type feature representation, which is normalized by the first softmax layer to generate the classification prediction probability of the related entity, and the loss value of the related entity classification is calculated according to the classification prediction probability of the related entity; based on the related entity type feature representation, a related entity type feature representation matrix is constructed, and the related entity type feature representation matrix is input into the graph decoder in the initial R-GNN model for link prediction, and is normalized by the second softmax layer to output the link prediction probability between related entities, and the link prediction probability between related entities is calculated. The measured loss value; based on the loss value of related entity classification and the loss value of link prediction between related entities, the total target loss value is calculated, and then the graph encoder network parameters and the graph decoder network parameters in the initial R-GNN model are updated through the gradient descent algorithm to optimize the total target loss value, and the optimal network parameters of the graph encoder and the optimal network parameters of the graph decoder in the R-GNN model are output, and the trained R-GNN model is obtained; the target position recruitment data and the newly compiled examination syllabus information in enterprises and institutions are obtained, and the target position recruitment data and the newly compiled examination syllabus information are subjected to information extraction processing, wherein the content of the information extraction processing includes: extracting the target position contained in the target position recruitment data, the target position The target position, the required abilities and qualities, the required knowledge for the target position and the newly compiled test point requirement information are input into the trained R-GNN model to obtain the test point requirement information corresponding to the abilities and qualities required for the target position and the test point requirement information corresponding to the knowledge required for the target position; based on the test point requirement information corresponding to the abilities and qualities required for the target position and the test point requirement information corresponding to the knowledge required for the target position, create the first item content of the key fields of the abilities and qualities required for the target position and the key fields of the knowledge required for the target position based on the columns in the table, configure the key fields of the test point requirement information corresponding to the first item content, and generate the corresponding competency model.
[0012] In some optional implementations of some embodiments, the configured competency model is used to construct a two-way detailed table of questions, specifically including: based on the competency model, creating a second project content of column-based key fields of ability and quality required for the target position and key fields of knowledge required for the target position in a table, and configuring the test point requirement information, knowledge point requirement information, question type requirement information, question quantity requirement information, question difficulty requirement information, total test paper difficulty requirement information, total score requirement information for each question type, score requirement information for each question type corresponding to the question, and question arrangement order requirement information corresponding to the second project content to generate a corresponding two-way detailed table of questions.
[0013] In some optional implementations of some embodiments, the mapping relationship formed between the fields of the two-way detailed table of questions and the initial question bank is used, and natural language processing technology is used to extract questions that meet the requirements of test paper composition from the initial question bank to construct a candidate test question bank, specifically including: according to the test point requirement information, knowledge point requirement information, question type requirement information, question difficulty requirement information and total test paper difficulty requirement information corresponding to the content of the second item when stored in columns in the two-way detailed table of questions, respectively, and the field mapping relationship of the initial question bank, natural language processing technology is used to perform semantic understanding and question extraction in the initial question bank, match the questions that meet the requirements of test paper composition, and construct a candidate test question bank.
[0014] In some optional implementations of some embodiments, the random question drawing and test paper composition algorithm and natural language processing technology are used to randomly draw test questions from the candidate test question bank to construct a candidate test paper, so as to meet the examination dimension information, test point requirement information, question type requirement information, question quantity requirement information and question difficulty requirement information of the two-way detailed table of questions as constraints, and specifically include: using natural language processing technology to match a first test question set classified based on the question type requirement information in the candidate test question bank; classifying the first test question set under each question type requirement information classification according to the examination dimension information, and obtaining the first test question set under each question type requirement information classification. A second test question set corresponding to each examination dimension information; classify the second test question set under each examination dimension information category according to the test point requirement information to obtain a third test question set under each examination dimension information category corresponding to each test point requirement information; classify and arrange the third test question set under each test point requirement information category according to the question difficulty requirement information and in ascending order to obtain a fourth test question set corresponding to each question difficulty requirement information under each test point requirement information category; randomly select test questions that meet the question quantity requirement information from the fourth test question set according to the hierarchy of question type requirement information-examination dimension information-test point requirement information-question difficulty requirement information to generate candidate test papers.
[0015] In some optional implementations of some embodiments, the calculation of the total difficulty value of the candidate test paper specifically includes: taking the historical test questions contained in the historical test question bank as the text to be trained, and adding corresponding test points, knowledge points and difficulty coefficient labels at the end of each piece of data to be trained, constructing a data set, dividing the data set into a training set and a verification set, and building an original neural network model, inputting the training set into the original neural network model for training to obtain a trained neural network model, inputting the verification set into the trained neural network model and training in combination with the loss function until the neural network model converges to obtain a trained neural network model; inputting the test questions contained in the candidate test paper, the test points corresponding to the test questions, the knowledge points corresponding to the test questions, and the difficulty coefficients corresponding to the test questions into the trained neural network model to obtain the difficulty coefficients of each test question in the candidate test paper, summing and averaging them to obtain the total difficulty value of the candidate test paper.
[0016] According to a second aspect of the present invention, a system for setting questions and forming papers based on a paper composition algorithm is provided, the system comprising: an initial test question bank construction module, used to obtain a historical test question bank and a newly added test question bank, and construct an initial test question bank; a competency model configuration module, used to obtain target position recruitment data and new examination syllabus information in enterprises and institutions, and configure a corresponding competency model; a question two-way detailed table construction module, used to construct a question two-way detailed table based on the configured competency model, the question two-way detailed table including information on the assessment dimensions of the ability and quality required for the target position, information on the assessment dimensions of the knowledge required for the target position, test point requirement information, knowledge point requirement information, question type requirement information, question quantity requirement information, question difficulty requirement information, total test paper difficulty requirement information, total score requirement information for each question type, score requirement information for each question type, and question arrangement order requirement information; a candidate test question bank module, used to generate a candidate test question bank based on the fields of the question two-way detailed table and the initial test question bank. The mapping relationship formed is used to extract test questions that meet the test paper composition requirements from the initial test question bank using natural language processing technology to construct a candidate test question bank; the candidate test paper module is used to use a random test paper composition algorithm and natural language processing technology to meet the examination dimension information, test point requirement information, question type requirement information, question quantity requirement information and question difficulty requirement information of the two-way detailed table of questions as constraints, randomly extract test questions from the candidate test question bank to construct a candidate test paper; the test paper generation module is used to calculate the total difficulty value of the candidate test paper, and determine whether the total difficulty value of the candidate test paper meets the total difficulty requirement information of the test paper. If so, adjust and generate the corresponding test paper according to the question type requirement information, question difficulty requirement information, total score requirement information of each question type, score requirement information of the corresponding question of each question type and question arrangement order requirement information; if not, adjust some questions in the candidate test paper until the total difficulty requirement information of the test paper is met.
[0017] The advantages and beneficial effects of the present invention are as follows: the present invention provides a method and system for setting questions and assembling questions based on an examination paper assembly algorithm. The method combines the knowledge graph with the R-GNN model to improve the accuracy of the model link prediction, and configures the corresponding competency model through the test points corresponding to the ability and quality required for the target position and the test points corresponding to the knowledge required for the target position, so as to construct a corresponding two-way detailed table of questions, and can accurately and specifically determine the required test points according to the degree of fit between the knowledge and ability and quality requirements that the target position should have, and realize the mapping relationship formed by the fields of the two-way detailed table of questions and the initial question bank. A candidate test question bank that meets the requirements for paper composition is constructed, and then a random question drawing algorithm is used to randomly draw questions from the candidate test question bank that meet the target position's required ability and quality assessment dimension information, target position's required knowledge assessment dimension information, test point requirement information, question type requirement information, question quantity requirement information and question difficulty requirement information in the two-way detailed list of questions, and obtain candidate test papers. Based on the total difficulty value of the candidate test papers and the total difficulty requirement information of the papers, a comparison can be made, which can not only ensure the preset paper composition difficulty requirements, but also flexibly adjust the required test papers according to actual needs, thereby improving the adaptability and practicality of the evaluation method. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 The present invention is a flowchart of a method for setting questions and composing examination papers based on an examination paper composition algorithm in an embodiment of the present invention.
[0019] Figure 2 It is a structural schematic diagram of a system for setting questions and composing examination papers based on an examination paper composition algorithm in an embodiment of the present invention.
[0020] Figure 3 Schematic diagram of the structure of the R-GNN model in an embodiment of the present invention.
[0021] Figure 4 Schematic diagram of the knowledge graph in an embodiment of the present invention. DETAILED DESCRIPTION
[0022] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below through specific implementation methods in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0023] Embodiment 1
[0024] This application proposes a method for setting questions and assembling examination papers based on an examination paper assembly algorithm, such as Figure 1 Shown: Includes the following steps.
[0025] S1. Obtain the historical question bank and the newly added question bank to build the initial question bank.
[0026] In this embodiment, step S1 specifically includes: obtaining a historical question bank and a newly added question bank, and constructing an initial question bank, specifically including: obtaining a historical question bank and a newly added question bank, marking the questions contained in the historical questions and the newly added questions, respectively obtaining historical question-related information and newly added question-related information, and constructing an initial question bank based on the historical question-related information and newly added question-related information; wherein, the historical question-related information includes historical question type information, historical question title information, historical question answer information, test point information corresponding to the historical questions, knowledge point information corresponding to the historical questions, and historical question difficulty coefficient information, and the newly added question-related information includes newly added question type information, newly added question title information, newly added question answer information, test point information corresponding to the newly added questions, knowledge point information corresponding to the newly added questions, and newly added question difficulty coefficient information.
[0027] Furthermore, in the labeling process, historical test questions and newly added test questions are labeled one by one respectively, and relevant information of historical test questions and relevant information of newly added test questions are obtained respectively, and the labeled data in the form of labels are obtained. Among them, the initial test question bank is constructed with the labeled data in the form of labels, which is not only convenient for the display and search of subsequent project contents, but also convenient for the training of neural network models.
[0028] Furthermore, history questions and newly added questions are marked separately. For example, if A is a history question, it is marked according to the history question type information, the history question answer information, the test point information corresponding to the history question, the knowledge point information corresponding to the history question and the history question difficulty coefficient information, that is, the history question A title label, the history question A answer label, the history question A type label, the test point label corresponding to the history question A, the knowledge point label corresponding to the history question A and the difficulty coefficient label corresponding to the history question A are obtained; if B is a newly added question, it is marked according to the new question type information, the new question question information, the new question answer information, the test point information corresponding to the new question, the knowledge point information corresponding to the new question and the difficulty coefficient information of the new question, that is, the new question B title label, the new question B answer label, the new question B type label, the test point label corresponding to the new question B, the knowledge point label corresponding to the new question B and the difficulty coefficient label corresponding to the new question B are obtained.
[0029] In this embodiment, the process of determining the difficulty coefficient information of the newly added test questions specifically includes: according to the field mapping relationship between the test point information corresponding to the newly added test questions and the knowledge point information corresponding to the newly added test questions and the corresponding test points and corresponding knowledge points in the history test question bank, respectively, using a natural language processing model, extracting historical test questions that meet the similarity with the test points of the newly added test questions to reach a preset test point threshold from the history test question bank and generating a historical test question test point set, and several historical test questions that meet the similarity with the knowledge points of the newly added test questions to reach a preset knowledge point threshold and generating a historical test question knowledge point set; assigning values to the historical test question test point set and the historical test question knowledge point set in turn according to preset weights, and calculating the comprehensive similarity value of the historical test question test point set and the historical test question knowledge point set; sorting the comprehensive similarity values from high to low, and selecting the historical test question difficulty coefficient information corresponding to the historical test question with the highest comprehensive similarity value as the difficulty coefficient information of the newly added test questions.
[0030] Furthermore, in order to determine the difficulty coefficient information of the newly added test questions so as to mark the difficulty coefficients of the newly added test questions later, according to the field mapping relationship between the test point information corresponding to the newly added test questions and the corresponding test points in the historical test question bank, as well as the field mapping relationship between the knowledge point information corresponding to the newly added test questions and the corresponding knowledge point information in the historical test question bank, the historical test questions that meet the similarity with the test points of the newly added test questions up to the preset test point threshold and generate a historical test question test point set, and several historical test questions that meet the similarity with the knowledge points of the newly added test questions up to the preset knowledge point threshold and generate a historical test question knowledge point set are searched from the historical test question bank respectively; then, the historical test question test point set and the historical test question knowledge point set are assigned values according to preset weights, and the comprehensive similarity values of the historical test question test point set and the historical test question knowledge point set are obtained by summing up; and the obtained comprehensive similarity values are sorted from high to low, and the difficulty coefficient corresponding to the historical test question with the highest comprehensive similarity value is taken as the difficulty coefficient information of the newly added test questions, so as to improve the accuracy of the difficulty coefficient information of the new test questions.
[0031] Furthermore, the similarity is calculated by using one of the following methods: correlation comparison, Chi-Square comparison, Intersection comparison, and Bhattacharyya distance comparison.
[0032] Furthermore, the difficulty coefficient of the test questions can also be divided into the difficulty level of the test questions. The difficulty level of the test questions can be divided into 5 levels. When the difficulty coefficient is greater than or equal to 0.9, the corresponding difficulty level is classified as a simple level; when the difficulty coefficient is greater than or equal to 0.7 and less than 0.9, the corresponding difficulty level is classified as an easy level; when the difficulty coefficient is greater than or equal to 0.5 and less than 0.7, the corresponding difficulty level is classified as a medium level; when the difficulty coefficient is greater than or equal to 0.2 and less than 0.5, the corresponding difficulty level is classified as a difficult level; when the difficulty coefficient is less than 0.2, the corresponding difficulty level is classified as a difficult level.
[0033] S2. Obtain the target position recruitment data and new examination syllabus information in enterprises and institutions, and configure the corresponding competency model.
[0034] In this embodiment, in order to comprehensively grasp the degree of fit between the newly compiled examination syllabus information and the talents required for the target recruitment positions, the present invention combines the knowledge graph and the R-GNN model to improve the accuracy of the prediction of the target position, the ability and quality required for the target position, the knowledge required for the target position, and the links between the test points.
[0035] In this embodiment, step S2 specifically includes: obtaining target job recruitment data and new examination syllabus information in enterprises and institutions, and configuring corresponding competency models, specifically including: obtaining historical job recruitment data and historical examination syllabus information in enterprises and institutions, and performing information extraction processing on the historical job recruitment data and historical examination syllabus information, wherein the content of the information extraction processing includes: extracting relevant entities and relationship information between relevant entities, wherein the relevant entities include positions, required abilities and qualities for positions, required knowledge for positions, and historical examination point requirement information; constructing a knowledge graph of the relationship between positions, required abilities and qualities for positions, required knowledge for positions, and historical examination point requirement information (refer to Appendix 1). Figure 4); Input the knowledge graph into the preset initial R-GNN model, encode the knowledge graph through the graph encoder in the initial R-GNN model, output the related entity type feature representation, the related entity type feature representation is normalized by the first softmax layer to generate the classification prediction probability of the related entity, and the loss value of the related entity classification is obtained according to the classification prediction probability of the related entity; Based on the related entity type feature representation, construct the related entity type feature representation matrix, input the related entity type feature representation matrix into the graph decoder in the initial R-GNN model for link prediction, and normalize it by the second softmax layer to output the link prediction probability between related entities, and calculate the loss value of the link prediction between related entities; Based on the loss value of the related entity classification and the loss value of the link prediction between related entities, calculate the total target loss value, and then update the graph encoder network parameters and the graph decoder network parameters in the initial R-GNN model through the gradient descent algorithm, optimize the total target loss value, and output the optimal network of the graph encoder in the R-GNN model. parameters and optimal network parameters of the graph decoder, and obtain the trained R-GNN model; obtain the target position recruitment data and the newly compiled examination syllabus information in enterprises and institutions, and perform information extraction processing on the target position recruitment data and the newly compiled examination syllabus information, wherein the content of the information extraction processing includes: extracting the target position, the ability and quality required for the target position, the knowledge required for the target position and the newly compiled test point requirement information contained in the target position recruitment data, and inputting the target position, the ability and quality required for the target position, the knowledge required for the target position and the newly compiled test point requirement information into the trained R-GNN model to obtain the test point requirement information corresponding to the ability and quality required for the target position and the test point requirement information corresponding to the knowledge required for the target position; based on the test point requirement information corresponding to the ability and quality required for the target position and the test point requirement information corresponding to the knowledge required for the target position, create the first item content of the key fields of the ability and quality required for the target position and the key fields of the knowledge required for the target position based on the columns in the table, configure the key fields of the test point requirement information corresponding to the first item content, and generate the corresponding competency model.
[0036] Furthermore, the knowledge graph constructed is specifically as follows: obtaining information on positions, abilities and qualities required for positions, knowledge required for positions, and historical test point requirements, constructing a first triplet and a second triplet according to the position type, and constructing a knowledge graph based on the first triplet and the second triplet, wherein the first triplet includes a first entity, a first relationship, and a second entity, and the second triplet includes a second entity, a second relationship, and a third entity.
[0037] It should be understood that the reference Figure 4,The knowledge graph is essentially a semantic network with multiple nodes in the network, which are called entities. For example, “position”, “ability and quality”, “knowledge”, “exam point requirement information”, etc. The edges (i.e. link relationships) emitted from these entities represent the information related to this entity. For example, the knowledge required for a human resources position is human resources management. In the knowledge graph, "human resources position" and "human resources management" are both entities. There is an edge between the two entities, and the name of the edge is "knowledge required for the position", that is, an edge is used to establish an association relationship between the first entity and the second entity. For the first triple, when there is a relationship between "position" and "ability and quality", or "position" and "knowledge", an edge can be used to connect "position" and "ability and quality" or "position" and "knowledge". It should be noted here that the first entity involved in the present invention represents the position, the second entity represents the ability and quality or knowledge, and the first relationship represents the ability and quality required for the corresponding position or the knowledge required for the corresponding position; for the second triple, when there is a relationship between "ability and quality" and "historical test point requirement information", or "knowledge" and "test point requirement information", an edge can be used to connect "ability and quality" and "test point requirement information" or "knowledge" and "historical test point requirement information". It should be noted here that the third entity involved in the present invention represents the test point requirement information, and the second relationship represents the test point requirement information required for the corresponding ability and quality or the test point requirement information required for the corresponding knowledge.
[0038] Among them, refer to the attached Figure 3The R-GNN model involved in the present invention is composed of a graph encoder and a graph decoder connected in sequence, and after the knowledge graph is input into the preset initial R-GNN model, the knowledge graph is encoded by the graph encoder, and the related entity type feature representation is output, which is then connected to the first softmax layer for normalization processing to realize entity classification prediction, wherein the predicted related entity classification includes position type classification, knowledge type classification, ability and quality type classification and test point requirement information classification, and the loss value of the related entity classification is predicted according to the classification probability of the related entity; for link prediction, the related entity type feature representation is used to construct a related entity type feature representation matrix, the related entity type feature representation matrix is input into the graph decoder for link prediction, and the observation label of the link prediction between related entities is output, which is then connected to the second softmax layer for normalization processing. , which is used to realize the prediction of the link between entities, wherein the predicted link relationship between entities includes the link relationship between the position and the ability and quality required for the corresponding position, the link relationship between the position and the knowledge required for the corresponding position, the link relationship between the ability and quality and the test point requirement information required to be examined for the corresponding ability and quality, and the link relationship between the knowledge and the test point requirement information required to be examined for the corresponding knowledge, and calculates the loss value of the link prediction between related entities; based on the loss value of the related entity classification and the loss value of the link prediction between related entities, calculates the total target loss value, and updates the model parameters (i.e., the graph encoder network parameters and the graph decoder network parameters) based on the gradient descent (Adam) algorithm, optimizes the total target loss value, outputs the optimal network parameters of the graph encoder and decoder in the R-GNN model, and uses the optimal network parameters to adjust the R-GNN model to obtain the trained R-GNN model.
[0039] Furthermore, the operation process of the R-GNN model specifically includes the following steps.
[0040] 1. Initialize the graph encoder and Graph Decoder , where G represents the input knowledge graph, represents the graph encoder network parameters, represents the matrix composed of all output related entity type feature representations of the last layer of the graph encoder (i.e., the related entity type feature representation matrix), Represents the network parameters of the graph decoder. The R-GNN model consists of a graph encoder and Graph Decoder Together they form.
[0041] 2. Encode the knowledge graph through the graph encoder to generate the feature representation of related entity types. The calculation formula is: ,in, Representing Entities The graph encoder The output features of the layer, Representing Entities The graph encoder The output features of the layer, represents the maximum activation function of each element, Representation and Entity Related The entity index, represents the normalization constant, represents the sample size (number of entities), Indicates The unknown weight parameter matrix of the layer, , Represents the number of hidden layers of the graph encoder.
[0042] 3. Calculate the output of the first softmax layer for node classification, that is, calculate the classification prediction probability of the relevant entity. The calculation formula is , where the vector No. The calculation formula for the classification prediction probability of an element is: , Indicates the type number of the related entity.
[0043] 4. Calculate the loss value of the relevant entity classification. The calculation formula is ,in, represents the set of indicators for labeled entities, express Label No. elements.
[0044] 5. Perform link prediction on the feature representation matrix of related entity types through the graph decoder, and calculate the output of the second softmax layer of link prediction in the graph decoder, that is, calculate the link prediction probability between related entities. The calculation formula is: , ,in, Representing relationships The observation vector, Represents the relationship The parameter matrix of .
[0045] 6. The loss value of link prediction between related entities is calculated through the cross entropy loss function. The calculation formula is: ,in, Representing relationships The observation label of and Relationship Type When ,otherwise .
[0046] 7. Sum the loss values of related entity classification and the loss values of link prediction between related entities to calculate the total target loss value. The calculation formula is: ,in, is a hyperparameter.
[0047] 8. Use the gradient descent (Adam) algorithm to update the graph encoder network parameters in the initial R-GNN model and graph decoder network parameters , optimize the total target loss value.
[0048] 9. Output the optimal network parameters of the graph encoder in the R-GNN model and the optimal network parameters of the graph decoder , and obtain the trained R-GNN model, where the trained R-GNN model consists of the graph encoder and Graph Decoder Together they form.
[0049] S3. Based on the configuration-based competency model, construct a two-way detailed list of propositions.
[0050] Among them, the two-way detailed table of questions includes information on the dimensions of ability and quality examinations required for the target position, information on the dimensions of knowledge examinations required for the target position, information on test point requirements, information on knowledge point requirements, information on question type requirements, information on the number of questions required, information on question difficulty requirements, information on the total difficulty of the test paper, information on the total score requirements for each question type, information on the score requirements for the corresponding questions of each question type, and information on the requirements for the order of question arrangement.
[0051] In this embodiment, step S3 specifically includes: based on the competency model, creating a second project content of the key fields of ability and quality required for the target position and the key fields of knowledge required for the target position based on columns in the table, and configuring the test point requirement information, knowledge point requirement information, question type requirement information, question quantity requirement information, question difficulty requirement information, total test paper difficulty requirement information, total score requirement information for each question type, score requirement information for each question type and test question arrangement order requirement information corresponding to the second project content, and generating a corresponding two-way detailed table of questions.
[0052] Furthermore, the corresponding competency model is configured as shown in Table 1 below.
[0053] Table 1 Competency model
[0054]
[0055] Furthermore, an example of a two-way detailed list of propositions is shown in Table 2 below.
[0056] Table 2 Bidirectional details of propositions
[0057]
[0058] S4. Based on the mapping relationship formed by the fields of the two-way detailed table of questions and the initial question bank, natural language processing technology is used to extract questions that meet the requirements of test paper composition from the initial question bank to construct a candidate test question bank.
[0059] In this embodiment, step S4 specifically includes: according to the test point requirement information, knowledge point requirement information, question type requirement information, question difficulty requirement information and total test paper difficulty requirement information corresponding to the second item content when stored in columns in the question bidirectional detailed table, respectively mapping the fields of the initial test question bank, using natural language processing technology to perform semantic understanding and test question extraction in the initial test question bank, matching the test questions that meet the test paper composition requirements, and constructing a candidate test question bank.
[0060] S5. Use random question selection algorithm and natural language processing technology to meet the examination dimension information, test point requirement information, question type requirement information, question quantity requirement information and question difficulty requirement information of the two-way detailed table of questions, randomly select questions from the candidate test question bank to construct the candidate test paper.
[0061] In this embodiment, step S5 specifically includes: using natural language processing technology to match a first test question set based on the question type requirement information classification in the candidate test question library; classifying the first test question set under each question type requirement information classification according to the examination dimension information to obtain a second test question set corresponding to each examination dimension information under each question type requirement information classification; classifying the second test question set under each examination dimension information classification according to the test point requirement information to obtain a third test question set corresponding to each test point requirement information under each examination dimension information classification; classifying and arranging the third test question set under each test point requirement information classification according to the question difficulty requirement information and in ascending order to obtain a fourth test question set corresponding to each question difficulty requirement information under each test point requirement information classification; and randomly selecting test questions that meet the question quantity requirement information from the fourth test question set according to the hierarchy of question type requirement information-examination dimension information-test point requirement information-question difficulty requirement information to generate a candidate test paper.
[0062] Furthermore, the natural language processing technology involved in the present invention adopts one of the methods including NLP technology, MT-DNN technology, and TAPAS to perform semantic understanding and information extraction.
[0063] Furthermore, by following the hierarchy of question type requirement information - examination dimension information - test point requirement information - question difficulty requirement information, recursive extraction can be performed in the fourth question set, which can not only quickly eliminate test questions that do not meet the requirements, but also quickly extract test questions that meet the question type requirement information, examination dimension information, test point requirement information and question difficulty. The search efficiency and accuracy can be improved by combining the examination dimensions, test points, question types, difficulty coefficients and the degree of fit with the basic knowledge and ability requirements of the recruitment position.
[0064] S6. Calculate the total difficulty value of the candidate test paper, and determine whether the total difficulty value of the candidate test paper meets the total difficulty requirement information of the test paper. If so, adjust and generate the corresponding test paper according to the question type requirement information, question difficulty requirement information, total score requirement information of each question type, score requirement information of the corresponding questions of each question type, and question arrangement order requirement information; if not, adjust some questions in the candidate test paper until the total difficulty requirement information of the test paper is met.
[0065] In this embodiment, the total difficulty value of the candidate test paper is calculated, specifically including: taking the historical test questions contained in the historical test question bank as the text to be trained, and adding the corresponding test point, knowledge point and difficulty coefficient label at the end of each piece of data to be trained, constructing a data set, dividing the data set into a training set and a verification set, and building an original neural network model, inputting the training set into the original neural network model for training to obtain a trained neural network model, inputting the verification set into the trained neural network model and training in combination with the loss function until the neural network model converges to obtain a trained neural network model; inputting the test questions contained in the candidate test paper, the test points corresponding to the test questions, the knowledge points corresponding to the test questions, and the difficulty coefficients corresponding to the test questions into the trained neural network model to obtain the difficulty coefficients of each test question in the candidate test paper, summing and averaging to obtain the total difficulty value of the candidate test paper.
[0066] Furthermore, in the process of constructing a data set involved in the present invention, when adding labels to the corresponding test points, knowledge points and difficulty coefficients of each piece of data to be trained, the historical test questions and the newly added test questions involved in step S1 can be used to directly construct a data set of historical test questions after the test questions have been annotated, that is, there is no need to repeat the process of adding labels to the historical test questions.
[0067] Furthermore, the neural network model involved in the present invention is a fully convolutional network (FCN) model, and the loss function is a cross entropy loss function.
[0068] In a specific application scenario, taking a scenario of setting and composing test papers as an example, the situation of adjusting and generating the corresponding test paper is explained. The preset test paper structure and preset arrangement order of the test questions are shown in Table 2, wherein the preset total test paper difficulty requirement information (i.e., the total test paper difficulty coefficient) is 0.5-0.6. If the calculated total difficulty value of the candidate test paper is 0.55, that is, the calculated total difficulty value of the candidate test paper meets the preset total test paper difficulty requirement information, and the test paper structure and arrangement order of the test questions in the candidate test paper are adjusted based on the preset test paper structure and arrangement order to generate the corresponding test paper; if the calculated total difficulty value of the candidate test paper is 0.65, that is, the calculated total difficulty value of the candidate test paper does not meet the preset total test paper difficulty requirement information. In order to meet the preset total difficulty requirement of the test paper, xx single-choice questions with a medium difficulty level in the candidate test paper can be selected and replaced with xxx single-choice questions with a relatively difficult difficulty level. Alternatively, xxxx multiple-choice questions with an easy difficulty level in the candidate test paper can be selected and replaced with xxxxx multiple-choice questions with a difficult difficulty level. The selection can be flexibly made according to the actual situation and is not limited here. The purpose is to replace the difficulty level (or difficulty coefficient) of some questions in the candidate test paper to ensure that the total difficulty value of the candidate test paper meets the total difficulty requirement of the test paper, and then adjust the test paper structure and arrangement order of the questions in the candidate test paper according to the preset test paper structure and arrangement order, so as to generate the corresponding test paper.
[0069] In this embodiment, the present invention uses historical test questions, corresponding test points, knowledge points and difficulty coefficients in the process of training the neural network model, which can improve the training effect of the neural network model. At the same time, the test questions contained in the candidate test paper, the test points corresponding to the test questions, the knowledge points corresponding to the test questions, and the difficulty coefficients corresponding to the test questions are input into the trained neural network model, which can improve the accuracy of the prediction of the difficulty coefficient of each test question in the candidate test paper, and further calculate the total difficulty value of the candidate test paper through the summation and averaging calculation process, and then compare the total difficulty value of the candidate test paper with the total difficulty requirement information of the test paper, and then obtain a judgment result of whether the total difficulty value of the candidate test paper meets the total difficulty requirement information of the test paper, and according to the judgment result, in order to flexibly adjust each test question in the test paper, the required test paper can be flexibly adjusted according to actual needs, thereby improving the adaptability and practicality of the evaluation method.
[0070] Embodiment 2
[0071] This embodiment provides a test paper setting system 200 based on the test paper setting algorithm on the basis of the above-mentioned embodiment 1. Figure 2, used to implement the steps of a method for setting questions and tests based on a test paper setting algorithm as described in the first embodiment above, the system mainly includes: an initial test question bank construction module 210, a competency model configuration module 220, a question two-way detailed table construction module 230, a candidate test question bank module 240, a candidate test test paper module 250 and a test paper generation module 260.
[0072] Among them, the initial question bank construction module 210 is used to obtain the historical question bank and the newly added question bank to construct the initial question bank; the competency model configuration module 220 is used to obtain the target position recruitment data and the new examination outline information in enterprises and institutions, and configure the corresponding competency model; the proposition two-way detailed table construction module 230 is used to construct the proposition two-way detailed table based on the configured competency model, and the proposition two-way detailed table includes the target position required ability and quality assessment dimension information, the target position required knowledge assessment dimension information, test point requirement information, knowledge point requirement information, question type requirement information, question quantity requirement information, question difficulty requirement information, test paper total difficulty requirement information, each question type total score requirement information, each question type corresponding question score requirement information and question arrangement order requirement information; the candidate test question bank module 240 is used to form a mapping relationship between the two-way detailed table of questions and the fields of the initial question bank, using natural language Processing technology, extracting test questions that meet the test paper composition requirements from the initial test question bank to construct a candidate test question bank; the candidate test paper module 250 is used to adopt a random test paper composition algorithm and natural language processing technology to meet the examination dimension information, test point requirement information, question type requirement information, question quantity requirement information and question difficulty requirement information of the two-way detailed table of questions as constraints, randomly extract test questions from the candidate test question bank to construct a candidate test paper; the test paper generation module 260 is used to calculate the total difficulty value of the candidate test paper, and determine whether the total difficulty value of the candidate test paper meets the total difficulty requirement information of the test paper. If so, adjust and generate the corresponding test paper according to the question type requirement information, question difficulty requirement information, total score requirement information of each question type, score requirement information of the corresponding question of each question type and question arrangement order requirement information; if not, adjust some questions in the candidate test paper until the total difficulty requirement information of the test paper is met.
[0073] Obviously, it should be understood by those skilled in the art that the above-mentioned various steps of the present invention can be performed in a manner different from the present invention, and the simulation method and experimental equipment include but are not limited to the above description. The above-mentioned various steps of the present invention can be performed in an order different from that here in some cases, and the steps shown or described above can be performed separately. Therefore, the present invention is not limited to any specific combination of hardware and software.
[0074] The above contents are further detailed descriptions of the present invention in combination with specific implementation methods, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.
Claims
1. A method for setting questions and composing examination papers based on an examination paper composition algorithm. It is characterized in that include: Obtain the historical question bank and the newly added question bank to build the initial question bank; Obtaining target position recruitment data and newly compiled examination syllabus information in enterprises and institutions, configuring corresponding competency models, and constructing a two-way detailed table of questions based on the configured competency models, wherein the two-way detailed table of questions includes information on the dimensions of ability and quality examinations required for the target position, information on the dimensions of knowledge examinations required for the target position, information on examination point requirements, information on knowledge point requirements, information on question type requirements, information on question quantity requirements, information on question difficulty requirements, information on total difficulty requirements for the test paper, information on total score requirements for each question type, information on score requirements for the corresponding questions for each question type, and information on question arrangement order requirements; According to the mapping relationship formed by the fields of the two-way detailed table of questions and the initial question bank, natural language processing technology is used to extract questions that meet the requirements of test paper composition from the initial question bank to build a candidate test question bank; Adopting random question selection and composition algorithm and natural language processing technology, randomly selecting questions from the candidate test question bank to construct the candidate test paper, with the information of examination dimension, test point requirement, question type requirement, question quantity requirement and question difficulty requirement in the two-way detailed table of questions as constraints; Calculate the total difficulty value of the candidate test paper, and determine whether the total difficulty value of the candidate test paper meets the total difficulty requirement information of the test paper. If so, adjust and generate the corresponding test paper according to the question type requirement information, the question difficulty requirement information, the total score requirement information of each question type, the score requirement information of the corresponding questions of each question type, and the question arrangement order requirement information; if not, adjust some questions in the candidate test paper until the total difficulty requirement information of the test paper is met; The acquisition of target position recruitment data and newly compiled examination syllabus information in enterprises and institutions and configuration of corresponding competency models specifically include: Obtaining historical job recruitment data and historical examination syllabus information in enterprises and institutions, and performing information extraction processing on the historical job recruitment data and historical examination syllabus information, wherein the content of the information extraction processing includes: extracting relevant entities and relationship information between relevant entities, wherein the relevant entities include positions, required abilities and qualities for positions, required knowledge for positions, and historical examination point requirement information; constructing a knowledge graph of the relationship between positions, required abilities and qualities for positions, required knowledge for positions, and historical examination point requirement information; The knowledge graph is input into the preset initial R-GNN model, and the knowledge graph is encoded by the graph encoder in the initial R-GNN model to output the related entity type feature representation, and the related entity type feature representation is normalized by the first softmax layer to generate the classification prediction probability of the related entity, and the loss value of the related entity classification is obtained according to the classification prediction probability of the related entity; based on the related entity type feature representation, a related entity type feature representation matrix is constructed, and the related entity type feature representation matrix is input into the graph decoder in the initial R-GNN model for link prediction, and the link prediction probability between related entities is output after normalization by the second softmax layer, and the loss value of the link prediction between related entities is calculated; based on the loss value of the related entity classification and the loss value of the link prediction between related entities, the total target loss value is calculated, and then the graph encoder network parameters and the graph decoder network parameters in the initial R-GNN model are updated by the gradient descent algorithm, the total target loss value is optimized, the optimal network parameters of the graph encoder and the optimal network parameters of the graph decoder in the R-GNN model are output, and the trained R-GNN model is obtained; Obtaining target position recruitment data and newly compiled examination syllabus information from enterprises and institutions, and performing information extraction processing on the target position recruitment data and the newly compiled examination syllabus information, wherein the information extraction processing includes: extracting the target position, the ability and quality required for the target position, the knowledge required for the target position, and the newly compiled examination point requirement information contained in the target position recruitment data, and inputting the target position, the ability and quality required for the target position, the knowledge required for the target position, and the newly compiled examination point requirement information into the trained R-GNN model to obtain the examination point requirement information corresponding to the ability and quality required for the target position and the examination point requirement information corresponding to the knowledge required for the target position; Based on the required test points corresponding to the required competence and quality of the target position and the test point requirement information corresponding to the required knowledge of the target position, create the first item content of the key fields of the required competence and quality of the target position and the key fields of the required knowledge of the target position based on the columns in the table, configure the key fields of the test point requirement information corresponding to the first item content, and generate the corresponding competency model; The configuration-based competency model constructs a two-way detailed list of propositions, specifically including: Based on the competency model, create the second item content of the key fields of the ability and quality required for the target position and the key fields of the knowledge required for the target position based on the columns in the table, and configure the test point requirement information, knowledge point requirement information, question type requirement information, question quantity requirement information, question difficulty requirement information, total test paper difficulty requirement information, total score requirement information of each question type, score requirement information of the corresponding question of each question type and test question arrangement order requirement information corresponding to the second item content, and generate the corresponding two-way detailed table of questions; According to the mapping relationship formed by the fields of the question bidirectional detailed table and the initial question bank, natural language processing technology is used to extract questions that meet the test paper composition requirements from the initial question bank to construct a candidate test question bank, which specifically includes: According to the test point requirement information, knowledge point requirement information, question type requirement information, question difficulty requirement information and total test paper difficulty requirement information corresponding to the second item content when stored in columns in the question proposition bidirectional detail table, the field mapping relationship with the initial test question bank is respectively established, and natural language processing technology is used to perform semantic understanding and test question extraction in the initial test question bank, match the test questions that meet the test paper composition requirements, and build a candidate test question bank; The random selection and composition algorithm and natural language processing technology are used to randomly select test questions from the candidate test question bank to construct the candidate test paper, with the examination dimension information, test point requirement information, question type requirement information, question quantity requirement information and question difficulty requirement information of the two-way detailed table of questions as constraints, and specifically include: Using natural language processing technology, matching a first test question set classified based on question type requirement information in a candidate test question bank; Classify the first test question sets under each question type requirement information classification according to the examination dimension information to obtain the second test question sets corresponding to each examination dimension information under each question type requirement information classification; Classify the second test question sets under each examination dimension information classification according to the test point requirement information to obtain the third test question sets under each examination dimension information classification corresponding to each test point requirement information; The third test question sets under each test point requirement information classification are classified and arranged in ascending order according to the question difficulty requirement information, to obtain the fourth test question sets corresponding to the question difficulty requirement information under each test point requirement information classification; Randomly extract test questions that meet the question quantity requirement information from the fourth test question set according to the hierarchy of question type requirement information - examination dimension information - test point requirement information - question difficulty requirement information to generate candidate test papers; The calculating of the total difficulty value of the candidate test paper specifically includes: The historical test questions contained in the historical test question bank are used as texts to be trained, and corresponding test points, knowledge points and difficulty coefficient labels are added to the end of each piece of data to be trained to construct a data set. The data set is divided into a training set and a validation set, and an original neural network model is constructed. The training set is input into the original neural network model for training to obtain a trained neural network model, and the validation set is input into the trained neural network model and trained in combination with a loss function until the neural network model converges to obtain a trained neural network model; the test questions contained in the candidate test paper, the test points corresponding to the test questions, the knowledge points corresponding to the test questions, and the difficulty coefficients corresponding to the test questions are input into the trained neural network model to obtain the difficulty coefficients of each test question in the candidate test paper, and the total difficulty value of the candidate test paper is obtained by summing and averaging.
2. A method for setting questions and composing examination papers based on an examination paper composition algorithm according to claim 1, It is characterized in that The obtaining of the historical question bank and the newly added question bank and the construction of the initial question bank specifically includes: A historical question bank and a newly added question bank are obtained, questions contained in the historical questions and the newly added questions are annotated, relevant information about the historical questions and relevant information about the newly added questions are obtained respectively, and an initial question bank is constructed based on the relevant information about the historical questions and the relevant information about the newly added questions.
3. A method for setting questions and composing examination papers based on an examination paper composition algorithm according to claim 2, It is characterized in that The information related to history test questions includes information on history test question types, information on history test question titles, information on history test question answers, information on test points corresponding to history test questions, information on knowledge points corresponding to history test questions, and information on difficulty coefficients of history test questions; the information related to newly added test questions includes information on new test question types, information on new test question titles, information on new test question answers, information on test points corresponding to new test questions, information on knowledge points corresponding to new test questions, and information on difficulty coefficients of new test questions.
4. A method for setting questions and composing examination papers based on an examination paper composition algorithm according to claim 3, It is characterized in that The process of determining the difficulty coefficient information of the newly added test questions specifically includes: According to the field mapping relationship between the test point information corresponding to the newly added test questions and the corresponding test point information in the history test question bank, as well as the field mapping relationship between the knowledge point information corresponding to the newly added test questions and the corresponding knowledge point information in the history test question bank, natural language processing technology is used to extract historical test questions that meet the preset test point threshold similarity with the test points of the newly added test questions from the history test question bank and generate a historical test question test point set, and several historical test questions that meet the preset knowledge point threshold similarity with the knowledge points of the newly added test questions and generate a historical test question knowledge point set; assign values to the historical test question test point set and the historical test question knowledge point set in turn according to preset weights, and calculate the comprehensive similarity value of the historical test question test point set and the historical test question knowledge point set; sort the comprehensive similarity values from high to low, and select the historical test question difficulty coefficient information corresponding to the historical test question with the highest comprehensive similarity value as the difficulty coefficient information of the newly added test questions.
5. A system for setting questions and assembling examination papers based on an examination paper assembly algorithm. It is characterized in that include: An initial question bank construction module is used to obtain a historical question bank and a newly added question bank to construct an initial question bank; The competency model configuration module is used to obtain the target position recruitment data and the newly compiled examination syllabus information in enterprises and institutions, and configure the corresponding competency model; A two-way detailed list construction module for questions is used to construct a two-way detailed list of questions based on the configured competency model, wherein the two-way detailed list of questions includes information on the dimensions of ability and quality examination required for the target position, information on the dimensions of knowledge examination required for the target position, information on test point requirements, information on knowledge point requirements, information on question type requirements, information on the number of questions required, information on the difficulty of questions required, information on the total difficulty of the test paper required, information on the total score requirement for each question type, information on the score requirement for each question type corresponding to the question, and information on the order of question arrangement required; The candidate test question bank module is used to extract questions that meet the test paper composition requirements from the initial question bank based on the mapping relationship formed by the question bidirectional detailed table and the fields of the initial question bank, using natural language processing technology to construct the candidate test question bank; The candidate test paper module is used to randomly select test questions from the candidate test question bank and construct the candidate test paper by using a random question selection algorithm and natural language processing technology to meet the examination dimension information, test point requirement information, question type requirement information, question quantity requirement information and question difficulty requirement information of the two-way detailed table of questions as constraints; The test paper generation module is used to calculate the total difficulty value of the candidate test paper and determine whether the total difficulty value of the candidate test paper meets the total difficulty requirement information of the test paper. If it meets the requirement information, the total score requirement information of each question type, the score requirement information of the corresponding questions of each question type and the test question arrangement order requirement information are adjusted and generated accordingly; if it does not meet the requirement information, some questions in the candidate test paper are adjusted until the total difficulty requirement information of the test paper is met; The acquisition of target position recruitment data and newly compiled examination syllabus information in enterprises and institutions and configuration of corresponding competency models specifically include: Obtaining historical job recruitment data and historical examination syllabus information in enterprises and institutions, and performing information extraction processing on the historical job recruitment data and historical examination syllabus information, wherein the content of the information extraction processing includes: extracting relevant entities and relationship information between relevant entities, wherein the relevant entities include positions, required abilities and qualities for positions, required knowledge for positions, and historical examination point requirement information; constructing a knowledge graph of the relationship between positions, required abilities and qualities for positions, required knowledge for positions, and historical examination point requirement information; The knowledge graph is input into the preset initial R-GNN model, and the knowledge graph is encoded by the graph encoder in the initial R-GNN model to output the related entity type feature representation, and the related entity type feature representation is normalized by the first softmax layer to generate the classification prediction probability of the related entity, and the loss value of the related entity classification is obtained according to the classification prediction probability of the related entity; based on the related entity type feature representation, a related entity type feature representation matrix is constructed, and the related entity type feature representation matrix is input into the graph decoder in the initial R-GNN model for link prediction, and the link prediction probability between related entities is output after normalization by the second softmax layer, and the loss value of the link prediction between related entities is calculated; based on the loss value of the related entity classification and the loss value of the link prediction between related entities, the total target loss value is calculated, and then the graph encoder network parameters and the graph decoder network parameters in the initial R-GNN model are updated by the gradient descent algorithm, the total target loss value is optimized, the optimal network parameters of the graph encoder and the optimal network parameters of the graph decoder in the R-GNN model are output, and the trained R-GNN model is obtained; Obtaining target position recruitment data and newly compiled examination syllabus information from enterprises and institutions, and performing information extraction processing on the target position recruitment data and the newly compiled examination syllabus information, wherein the information extraction processing includes: extracting the target position, the ability and quality required for the target position, the knowledge required for the target position, and the newly compiled examination point requirement information contained in the target position recruitment data, and inputting the target position, the ability and quality required for the target position, the knowledge required for the target position, and the newly compiled examination point requirement information into the trained R-GNN model to obtain the examination point requirement information corresponding to the ability and quality required for the target position and the examination point requirement information corresponding to the knowledge required for the target position; Based on the required test points corresponding to the required competence and quality of the target position and the test point requirement information corresponding to the required knowledge of the target position, create the first item content of the key fields of the required competence and quality of the target position and the key fields of the required knowledge of the target position based on the columns in the table, configure the key fields of the test point requirement information corresponding to the first item content, and generate the corresponding competency model; The configuration-based competency model constructs a two-way detailed list of propositions, specifically including: Based on the competency model, create the second item content of the key fields of the ability and quality required for the target position and the key fields of the knowledge required for the target position based on the columns in the table, and configure the test point requirement information, knowledge point requirement information, question type requirement information, question quantity requirement information, question difficulty requirement information, total test paper difficulty requirement information, total score requirement information of each question type, score requirement information of the corresponding question of each question type and test question arrangement order requirement information corresponding to the second item content, and generate the corresponding two-way detailed table of questions; According to the mapping relationship formed by the fields of the question bidirectional detailed table and the initial question bank, natural language processing technology is used to extract questions that meet the test paper composition requirements from the initial question bank to construct a candidate test question bank, which specifically includes: According to the test point requirement information, knowledge point requirement information, question type requirement information, question difficulty requirement information and total test paper difficulty requirement information corresponding to the second item content when stored in columns in the question proposition bidirectional detail table, the field mapping relationship with the initial test question bank is respectively established, and natural language processing technology is used to perform semantic understanding and test question extraction in the initial test question bank, match the test questions that meet the test paper composition requirements, and build a candidate test question bank; The random selection and composition algorithm and natural language processing technology are used to randomly select test questions from the candidate test question bank to construct the candidate test paper, with the examination dimension information, test point requirement information, question type requirement information, question quantity requirement information and question difficulty requirement information of the two-way detailed table of questions as constraints, and specifically include: Using natural language processing technology, matching a first test question set classified based on question type requirement information in a candidate test question bank; Classify the first test question sets under each question type requirement information classification according to the examination dimension information to obtain the second test question sets corresponding to each examination dimension information under each question type requirement information classification; Classify the second test question sets under each examination dimension information classification according to the test point requirement information to obtain the third test question sets under each examination dimension information classification corresponding to each test point requirement information; The third test question sets under each test point requirement information classification are classified and arranged in ascending order according to the question difficulty requirement information, to obtain the fourth test question sets corresponding to the question difficulty requirement information under each test point requirement information classification; Randomly extract test questions that meet the question quantity requirement information from the fourth test question set according to the hierarchy of question type requirement information - examination dimension information - test point requirement information - question difficulty requirement information to generate candidate test papers; The calculating of the total difficulty value of the candidate test paper specifically includes: The historical test questions contained in the historical test question bank are used as texts to be trained, and corresponding test points, knowledge points and difficulty coefficient labels are added to the end of each piece of data to be trained to construct a data set. The data set is divided into a training set and a validation set, and an original neural network model is constructed. The training set is input into the original neural network model for training to obtain a trained neural network model, and the validation set is input into the trained neural network model and trained in combination with a loss function until the neural network model converges to obtain a trained neural network model; the test questions contained in the candidate test paper, the test points corresponding to the test questions, the knowledge points corresponding to the test questions, and the difficulty coefficients corresponding to the test questions are input into the trained neural network model to obtain the difficulty coefficients of each test question in the candidate test paper, and the total difficulty value of the candidate test paper is obtained by summing and averaging.
