A multi-capability intelligent question generation method and system

Abstract

The present application relates to a kind of multi-capability intelligent question generation method and system, belong to intelligent natural language processing technical field.Wherein, the method includes: obtaining core competency dimension and industry general competency element, constructs initial multi-capability evaluation graph, then constructs the multi-capability evaluation graph containing competency node, priority weight and association by competency dimension mapping algorithm;Based on graph construction dual-mode dynamic combination scheduling mechanism, select batch or progressive mode to generate initial question;Build competency verification model, calculate the deviation of question in coverage integrity, depth adaptability and distribution balance;According to the deviation adjustment question, according to priority and depth rule collation output question package, while reverse optimization algorithm parameter and model threshold value.The present application realizes multi-scene flexible adaptation, improves the precision of question and evaluation requirement Adaptation degree, guarantees the quality of question, meets the demand of efficient generation standardized question in diversified multi-capability evaluation scene.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent natural language processing, and particularly relates to a multi-competency intelligent questioning generation method and system. BACKGROUND

[0002] In intelligent evaluation, talent selection, skill training, knowledge examination and many other fields, obtaining competency-related information of evaluation objects through accurate questioning is the core prerequisite for efficient evaluation. At present, intelligent questioning generation technology has been gradually applied to the above-mentioned scenarios. Its core idea is to generate questioning content that adapts to evaluation needs based on preset competency dimensions, so as to replace the traditional manual design of questioning, improve the efficiency and standardization of questioning design, reduce the influence of human subjective factors on the quality of questioning, and adapt to large-scale and multi-scenario evaluation needs.

[0003] Existing intelligent questioning generation technology is mostly developed around a single competency dimension, or uses fixed competency combination templates to generate questions. Although this reduces labor costs and improves the efficiency of question generation to some extent, there are still many defects to be solved in actual application:

[0004] Firstly, the multi-competency combination mode is rigid, and most technologies only support a single mode of "determining the complete combination first and then generating the question". This mode cannot adapt to the needs of dynamic adjustment of evaluation targets and flexible changes of evaluation scenarios, making it difficult to meet the batch questioning needs in standardized evaluation scenarios and unable to adapt to the dynamic adjustment demands in open evaluation scenarios, resulting in a lack of flexibility in question generation;

[0005] Secondly, the initial question generation has strong dependence on pre-information. Existing solutions mostly rely on the detailed information of evaluation objects such as resumes and past evaluation records to generate targeted questions through keyword matching. If there is a lack of pre-information such as resumes, only general fixed template questions can be output, which is difficult to match the personalized evaluation needs of different industries and different positions, and the scope of application is limited;

[0006] Thirdly, the initial question generation lacks balanced coverage of competencies. Existing technologies do not systematically check the completeness of coverage of each competency dimension and the distribution of question depth when generating questions, which may lead to problems such as over-coverage of some dimensions, lack of some core dimensions, or single-depth questions that cannot match the evaluation accuracy, affecting the comprehensiveness and accuracy of subsequent evaluation;

[0007] Fourthly, there is a lack of complete optimization loop from demand analysis to question generation. Existing technologies mostly stay in the one-way process of "demand input - question output", and do not establish a mechanism for reverse optimization of competency analysis logic, combination rules and verification standards based on the results of question generation, making it difficult to continuously improve the adaptability of question generation and adapt to long-term and changing evaluation needs.

[0008] Meanwhile, the prior art lacks standardized optimization logic in the competency dimension analysis link, and directly uses industry general dimensions without optimizing the competency elements and priorities in combination with specific evaluation targets, further leading to a disconnection between the generated questions and the core demands of the evaluation. SUMMARY

[0009] To solve the above problems in the prior art, the present application provides a multi-competency intelligent question generation method and system, and the object of the present application can be achieved by the following technical solutions:

[0010] Comprising:

[0011] S1: Obtain core competency dimensions and industry general competency elements, and construct an initial multi-competency evaluation graph; according to the evaluation target parameters, construct a multi-competency evaluation graph containing competency nodes, node priority weights, and node association relationships through a multi-competency dimension mapping algorithm;

[0012] S2: Based on the evaluation target parameters and the multi-competency evaluation graph, extract state variables representing the state of the evaluation scene; according to the state variables and the preset generation strategy determination rules, construct a double-mode dynamic combination scheduling mechanism, including a batch combination generation mode and a progressive combination generation mode; in the batch combination generation mode, the initial question set is generated in batches according to the association logic, and in the progressive combination generation mode, the questions are generated in order according to the priority and dynamically supplemented with the uncovered dimensions;

[0013] S3: Build a competency verification model to perform competency verification operations on the initial question set, extract the competency coverage label and question depth label corresponding to each question; based on the preset competency coverage threshold, question depth matching threshold, and question distribution balance threshold, calculate the deviation of the question set in terms of coverage completeness, depth adaptability, and distribution balance;

[0014] S4: According to the deviation, automatically adjust the competency node weight parameters and the competency verification model threshold; sort the initial question package according to the competency priority and question depth rules; feed back the adjusted parameters to the subsequent question generation process, and optimize the multi-competency dimension mapping algorithm parameters and the competency verification model threshold.

[0015] Specifically, the specific process of constructing the initial multi-competency evaluation graph is:

[0016] Integrate the screened core competency dimensions and industry general competency elements to obtain the corresponding core evaluation direction of each element;

[0017] According to the industry general evaluation standard and the basic evaluation demand, preset the evaluation boundary of each competency dimension, and remove redundant and irrelevant evaluation elements;

[0018] Set the initial association rules between each competency dimension and the general elements, and allocate the initial weight proportion of each element.

[0019] Specifically, the specific process of constructing the multi-competency evaluation graph containing competency nodes, node priority weights and node association relationships is:

[0020] Input the evaluation target difference related parameters, and build the dimension evaluation index system;

[0021] Calculate the fit degree of each competency component element and the evaluation target based on the evaluation index system;

[0022] Adjust the association logic of competency component elements combined with the evaluation target, and redistribute the priority order and weight proportion of each element using hierarchical analysis method;

[0023] Verify the optimization effect through adaptability check, determine the optimized multi-competency evaluation mechanism after the check meets the standard, and repeat the adjustment of elements and weights until it meets the requirements if it does not meet the standard.

[0024] Specifically, the specific process of extracting state variables representing the state of the evaluation scene is: presetting the evaluation scale, evaluation rhythm, evaluation form and preposed data availability core dimensions; obtaining the original data corresponding to each dimension; standardizing the original data and completing the missing values; generating scale, rhythm, form and data availability state variables by weighted summation, forming a variable set.

[0025] Specifically, the specific process of constructing the dual-mode dynamic combination scheduling mechanism is: preset the triggering conditions of batch combination and progressive combination two modes; extract the association characteristics of each competency dimension; establish the matching relationship between the two modes and the corresponding question generation strategy; execute the question generation logic according to the selected mode.

[0026] Specifically, the specific process of generating an initial question set is: combine the dimension association logic and competency priority in the multi-competency evaluation graph, and execute intelligent dynamic generation operation through the selected generation mode; in batch combination generation mode, based on competency element semantic fusion algorithm, extract each competency core evaluation point according to the dimension association order, generate basic questions through cross-dimension element semantic deep cross-fusion, and then output questions covering core competency and industry general competency full-dimension through scene-based semantic optimization, finally form an initial question set containing multi-competency questions.

[0027] Specifically, the specific process of generating questions in the progressive combination generation mode according to priority and dynamically supplementing uncovered dimensions is as follows: the optimized priority of competencies is sorted; the covered dimensions corresponding to the generated questions are counted in real time, and compared with all dimensions in the multi-competency evaluation graph; after the uncovered dimensions are identified, the supplement generation order is determined according to the association logic, and corresponding questions are generated in sequence until all dimensions are covered.

[0028] Specifically, the specific process of building the competency verification model is as follows:

[0029] It includes a coverage identification layer, a depth matching layer and a balance evaluation layer.

[0030] The coverage identification layer extracts the core semantic features of each initial question and converts them into competency labels, matches the labels with the dimensions in the multi-competency evaluation graph, and clearly marks the covered dimensions and the uncovered dimensions.

[0031] The depth matching layer presets the question depth standard corresponding to each competency dimension, and determines the depth level corresponding to each question.

[0032] The balance evaluation layer builds a balance calculation logic, and sets a threshold interval for the number of questions for each dimension in combination with the evaluation requirements.

[0033] Specifically, the specific process of extracting the competency coverage label and the question depth label corresponding to each question is as follows:

[0034] The core semantic information of the question is captured by a semantic feature extraction algorithm, and the semantic similarity between the competency nodes in the multi-competency evaluation graph is matched to determine the corresponding covered competency dimensions and generate the competency coverage label.

[0035] Based on the semantic depth and evaluation direction of the question, the depth level is automatically determined by a semantic level analysis algorithm to generate the question depth label.

[0036] Specifically, the specific process of calculating the deviation amount of the question set in coverage integrity, depth adaptability and distribution balance is as follows: the coverage integrity deviation amount is calculated by the relative deviation of the actual coverage dimension proportion and the preset threshold; the depth adaptability deviation amount is the weighted sum of the absolute difference value of the actual proportion of each depth level and the threshold; the distribution balance deviation amount is the deviation rate mean value of the number of questions for each dimension and the preset interval median value; and the integrated comprehensive deviation amount set is output.

[0037] Specifically, the specific process of sorting the initial question set according to the competence priority and question depth rule is as follows: the competence priority and question depth level corresponding to each question in the optimized question set are extracted; the questions are grouped and classified according to the competence priority from high to low, and the questions of the same priority competence are collected into the same group; the questions in each group are sequentially sorted according to the question depth; and all the sorted questions are integrated according to the group priority.

[0038] Specifically, a multi-competence intelligent question generation system comprises:

[0039] A multi-competence evaluation mechanism construction module: acquires core competence dimensions and industry general competence elements, and constructs an initial multi-competence evaluation graph; according to an evaluation target parameter, a multi-competence evaluation graph comprising competence nodes, node priority weights and node correlation relationships is constructed through a multi-competence dimension mapping algorithm;

[0040] A dual-mode combination and initial question generation module: based on the evaluation target parameter and the multi-competence evaluation graph, state variables representing the state of the evaluation scene are extracted; according to the state variables and a preset generation strategy determination rule, a dual-mode dynamic combination scheduling mechanism is constructed, comprising a batch combination generation mode and a progressive combination generation mode; in the batch combination generation mode, an initial question set is batch generated according to the correlation logic, and in the progressive combination generation mode, questions are sequentially generated according to the priority and dynamically supplemented with uncovered dimensions;

[0041] An initial question competence verification module: a competence verification model is built, and a competence verification operation is performed on the initial question set, and the competence coverage label and the question depth label corresponding to each question are extracted; based on a preset competence coverage threshold, a question depth matching threshold and a question distribution balance threshold, the deviation of the question set in terms of coverage integrity, depth adaptability and distribution balance is calculated;

[0042] An initial question set output module: according to the deviation, the competence node weight parameter and the competence verification model threshold are automatically adjusted; the initial question set is sorted according to the competence priority and the question depth rule; the adjusted parameters are fed back to the subsequent question generation process, and the multi-competence dimension mapping algorithm parameter and the competence verification model threshold are optimized.

[0043] The beneficial effects of the present application are as follows:

[0044] (1) By setting the dual-mode dynamic combination scheduling mechanism, the triggering conditions of batch combination and progressive combination two modes are preset first, then the matching relationship between the mode and the generation strategy is established by extracting the correlation characteristics of each competence dimension, and the generation mode can be flexibly switched according to the evaluation scene: the batch combination generation mode is selected in the standardized evaluation scene, the initial questions of full-dimension are generated in batch relying on the correlation logic graph, and the efficiency of question generation is improved; The progressive combination generation mode is selected in the dynamic adjustment evaluation scene, the questions are generated in turn according to the priority, and the covered dimensions are compared in real time, the questions of the uncovered dimensions are dynamically supplemented, and the full-dimension coverage is ensured. At the same time, the correlation logic generated by combining semantic similarity calculation and evaluation target synergy analysis can effectively avoid the problems of rigid and fragmented questions caused by single combination mode;

[0045] (2) Through the reverse optimization logic of the competence verification model and the question generation and evaluation mechanism, on the one hand, the coverage identification layer of the competence verification model can accurately match the questions and the competence dimensions, mark the uncovered dimensions, the depth matching layer can determine the depth level of the questions, and the balanced evaluation layer can calculate the balanced degree of the questions and output the optimization direction, realizing the systematic verification of the initial question coverage integrity, depth adaptability and distribution balance, supplementing the missing dimension questions, adjusting the imbalance of the question ratio, and from the root cause, guaranteeing the initial question quality, and avoiding the influence of dimension missing or imbalance on the evaluation comprehensiveness; On the other hand, through the reverse optimization of the multi-competence dimension mapping algorithm parameters and the competence verification model threshold based on the question generation result, a closed loop iteration of "generation-verification-optimization" is formed, which continuously improves the accuracy and scene adaptability of subsequent question generation, and cooperates with the preset multi-dimension sorting rule and output specification, so that the final output question package is clear in classification and coherent in logic, and is more convenient for practical evaluation scene application. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to facilitate the understanding of those skilled in the art, the present application will be further described below with reference to the accompanying drawings.

[0047] Fig. 1 The flowchart of a multi-competence intelligent question generation method and system of the present application;

[0048] Fig. 2 The data flow diagram of a multi-competence intelligent question generation method and system of the present application. DETAILED DESCRIPTION

[0049] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined invention purpose, the specific embodiments, structures, features and effects according to the present application will be described in detail below with reference to the accompanying drawings and preferred embodiments.

[0050] Please refer to Figs. 1-2 A multi-competence intelligent question generation method and system;

[0051] comprise:

[0052] S1: Obtain core competence dimensions and industry general competence elements, and construct an initial multi-competence evaluation graph; according to the evaluation target parameters, a multi-competence evaluation graph containing competence nodes, node priority weights and node association relationships is constructed through a multi-competence dimension mapping algorithm;

[0053] S2: Based on the evaluation target parameters and the multi-competence evaluation graph, state variables representing the state of the evaluation scene are extracted; according to the state variables and the preset generation strategy judgment rule, a double-mode dynamic combination scheduling mechanism is constructed, including a batch combination generation mode and a progressive combination generation mode; the initial question set is generated in batches according to the association logic in the batch combination generation mode, and the questions are generated in order according to the priority and dynamically supplemented with the uncovered dimensions in the progressive combination generation mode;

[0054] S3: Build a competence verification model to perform competence verification operation on the initial question set, extract the competence coverage label and question depth label corresponding to each question; based on the preset competence coverage threshold, question depth matching threshold and question distribution balance threshold, calculate the deviation of the question set in terms of coverage integrity, depth adaptability and distribution balance;

[0055] S4: According to the deviation, automatically adjust the competence node weight parameter and the competence verification model threshold; arrange the initial question package according to the competence priority and question depth rule; feed back the adjusted parameters to the subsequent question generation process, and optimize the multi-competence dimension mapping algorithm parameters and the competence verification model threshold.

[0056] In this embodiment, the core competence dimension refers to the core ability dimension that fits the specific evaluation scene and is directly related to the evaluation core demand, specifically including communication and coordination, project management, decision-making, stress resistance, team management, and professional empowerment; the core direction needs to be determined in combination with the evaluation object and the scene; the industry general competence element refers to the ability element that is general in each industry and each post and is suitable for basic evaluation needs, specifically including responsibility, learning, execution, problem solving, and compliance awareness; it is classified and organized according to industry attributes, and is suitable for different industry basic evaluation standards, and together with the core competence dimension, it constitutes a basic ability system for evaluation.

[0057] In this embodiment, the evaluation target parameter refers to the difference in core evaluation appeal in different evaluation scenarios, which is the core basis for optimizing the multi-competency evaluation mechanism, and specifically includes four categories: evaluation scenario attribute difference, evaluation object level difference, core evaluation appeal difference, and evaluation cycle difference. The evaluation scenario attribute difference covers differences in scenarios such as standardized recruitment, open training, and skill assessment. The evaluation object level difference covers the differences in ability requirements of evaluation objects at different levels, such as basic, intermediate, and high levels. The core evaluation appeal difference covers the differences in appeal, such as focusing on ability screening, deepening ability, and filling short boards. The evaluation cycle difference covers the differences in evaluation period, such as short-term rapid evaluation and long-term deep evaluation. The above differences together constitute the evaluation target difference parameter, which serves as the input basis for algorithm optimization.

[0058] In this embodiment, the competency composition elements are the detailed and disassembled contents of the core competency dimensions and industry general competency elements, and specifically include: the first-level main dimensions and second-level sub-dimensions disassembled from each core competency dimension, and specific ability points corresponding to the industry general competency elements. The priority order refers to the importance level division of the above-mentioned competency composition elements based on the evaluation target difference. The specific sorting logic is to determine the sorting by using hierarchical analysis method in combination with the fit degree of elements and evaluation targets, the contribution degree of elements to evaluation results, and the evaluability of elements. The elements with high priority are preferentially included in the question generation range, preferentially matched with more questions, and the generated questions are ensured to be in line with the core evaluation appeal.

[0059] In this embodiment, the question depth information refers to the level of initial questions corresponding to competency evaluation, and specifically includes three categories: basic cognitive layer, advanced application layer, and deep analysis layer, which are one-to-one corresponding to each competency dimension. The basic cognitive layer corresponds to the question information about basic concepts and basic cognition related to competency, focusing on "whether to understand, whether to master". The advanced application layer corresponds to the question information about the application of competency in actual scenarios, focusing on "how to apply, how to operate". The deep analysis layer corresponds to the question information about the deep analysis and decision optimization of scenarios related to competency, focusing on "why do it this way, how to optimize". The semantic depth is determined to determine the depth level of each question, ensuring that the question depth meets the evaluation accuracy requirements.

[0060] In this embodiment, the multi-competency dimension mapping algorithm parameters are the core configurations for constructing and optimizing the multi-competency evaluation atlas, specifically including the dimension fit degree calculation weight and the correlation logic adjustment coefficient, which can directly output the competency node weight parameters (i.e. the importance weight of each competency node in the evaluation atlas) as algorithm input, and the two constitute an "input configuration-output result" relationship; the competency verification model threshold is a complete benchmark threshold system of the verification model, which mainly includes the competency coverage threshold, the question depth matching threshold, the question distribution balance threshold, and also covers auxiliary thresholds such as semantic similarity matching threshold, and the "competency verification model threshold" is a common name for them, and the two are synonymous; the competency node weight parameters adjusted in the S4 step need to be supported by algorithm parameter optimization, and the adjustment of the competency verification model threshold is to modify the competency verification model threshold, and the adjusted two types of parameters will optimize the multi-competency dimension mapping algorithm parameters and the competency verification model threshold in reverse.

[0061] Specifically, the specific process of constructing the initial multi-competency evaluation atlas is:

[0062] Integrate the screened core competency dimensions and industry general competency elements, and obtain the corresponding core evaluation directions of each element;

[0063] According to the industry general evaluation standard and the basic evaluation demand, preset the evaluation boundary of each competency dimension, and remove redundant and irrelevant evaluation elements;

[0064] Set the initial correlation rules between each competency dimension and the general elements, and allocate the initial weight proportion of each element.

[0065] Specifically, the specific process of constructing the multi-competency evaluation atlas containing competency nodes, node priority weight and node correlation is:

[0066] Input the evaluation target difference related parameters, and build a dimension evaluation index system;

[0067] Calculate the fit degree of each competency constituent element and the evaluation target based on the evaluation index system;

[0068] Adjust the correlation logic of the competency constituent elements in combination with the evaluation target, and redistribute the priority order and weight proportion of each element by using hierarchical analysis method;

[0069] Verify the optimization effect through adaptability verification, determine the optimized multi-competency evaluation mechanism after the verification meets the standard, and repeat the adjustment of elements and weights until the requirements are met if the standard is not met.

[0070] Specifically, the specific process of the extraction characterization evaluates the state variable of the scene state is: presetting the evaluation scale, the evaluation rhythm, the evaluation form and the preposed data availability core dimension; obtaining the original data corresponding to each dimension; standardizing the original data and completing the missing values; generating the scale, rhythm, form and data availability state variable by weighted summation, forming the variable set.

[0071] Specifically, the specific process of constructing the dual-mode dynamic combination scheduling mechanism is: presetting the trigger conditions of the two modes of batch combination and progressive combination; extracting the associated features of each competence dimension; establishing the matching relationship between the two modes and the corresponding question generation strategy; executing the logic of question generation according to the selected mode.

[0072] Specifically, the specific process of generating the initial question set is: combining the dimension association logic and the priority of the multi-competence evaluation graph, and executing the intelligent dynamic generation operation through the selected generation mode; in the batch combination generation mode, based on the competence element semantic fusion algorithm, the core evaluation points of each competence are extracted in the dimension association order, the basic questions are generated through cross-dimensional element semantic deep cross-fusion, and then the questions covering the core competence and the industry general competence are batched out through scene semantic optimization; in the progressive combination generation mode, the deep questions focusing on the core competence are generated in order according to the priority, the covered dimensions are compared with the full dimensions of the evaluation graph in real time, and the non-repetitive and strongly associated supplementary questions are generated for the uncovered dimensions through the association derivation algorithm, finally forming the initial question set containing multi-competence questions.

[0073] Specifically, the specific process of generating questions in the progressive combination generation mode according to the priority and dynamically supplementing the uncovered dimensions is: sorting according to the optimized competence priority; real-time statistics of the covered dimensions corresponding to the generated questions are compared with the full dimensions in the multi-competence evaluation graph; after identifying the uncovered dimensions, the generation order is determined according to the association logic, and the corresponding questions are generated in turn until the full dimensions are covered.

[0074] Specifically, the specific process of building the competence verification model is:

[0075] It includes: coverage identification layer, deep matching layer and balance evaluation layer;

[0076] The coverage identification layer extracts the core semantic features of each initial question and converts them into competence labels, matches the labels with the dimensions in the multi-competence evaluation graph, and clearly marks the covered dimensions and the uncovered dimensions;

[0077] The deep matching layer presets the question depth standard corresponding to each competence dimension, and determines the depth level corresponding to each question;

[0078] The balance evaluation layer builds balance calculation logic and sets threshold intervals for the number of questions in each dimension according to evaluation requirements.

[0079] Specifically, the specific process of extracting the corresponding competence coverage label and question depth label of each question is:

[0080] Through a semantic feature extraction algorithm, the core semantic information of the question is captured, and a semantic similarity matching is performed with the competence nodes in the multi-competence evaluation graph to determine the corresponding covered competence dimension and generate a competence coverage label;

[0081] Based on the semantic depth and evaluation direction of the question, a semantic hierarchy analysis algorithm is used to automatically determine the depth level and generate a question depth label.

[0082] Specifically, the specific process of calculating the deviation amount of the question set in terms of coverage completeness, depth adaptability, and distribution balance is as follows: the coverage completeness deviation amount is calculated by calculating the relative deviation of the actual coverage dimension proportion from the preset threshold; the depth adaptability deviation amount is calculated by weighted summing the absolute difference between the actual proportion of each depth level and the threshold; and the distribution balance deviation amount is calculated by taking the average of the deviation rate of the number of questions in each dimension from the median value in the preset interval; and the integrated deviation amount set is output.

[0083] Specifically, the specific process of arranging the initial question package according to the competence priority and question depth rules is as follows: the competence priority and question depth level corresponding to each question in the optimized question package are extracted; the questions are grouped and classified according to the competence priority from high to low, and the questions of the same priority competence are grouped into the same group; the questions in each group are sorted according to the question depth; and all sorted questions are integrated according to the group priority.

[0084] Specifically, a multi-competence intelligent question generation system comprises:

[0085] A multi-competence evaluation mechanism construction module: acquires core competence dimensions and industry general competence elements, and constructs an initial multi-competence evaluation graph; according to evaluation target parameters, a multi-competence evaluation graph containing competence nodes, node priority weights, and node association relationships is constructed through a multi-competence dimension mapping algorithm;

[0086] A dual-mode combination and initial question generation module: based on the evaluation target parameters and the multi-competence evaluation graph, state variables representing the state of the evaluation scene are extracted; according to the state variables and a preset generation strategy determination rule, a dual-mode dynamic combination scheduling mechanism is constructed, including a batch combination generation mode and a progressive combination generation mode; in the batch combination generation mode, an initial question set is generated in batches according to the association logic, and in the progressive combination generation mode, questions are generated in order according to the priority and dynamically supplemented with uncovered dimensions.

[0087] The initial question competence verification module: a competence verification model is built, a competence verification operation is performed on the initial question set, and the competence coverage label and the question depth label corresponding to each question are extracted; based on the preset competence coverage threshold, the question depth matching threshold and the question distribution balance threshold, the deviation amount of the question set in the coverage integrity, the depth adaptability and the distribution balance is calculated;

[0088] The initial question package output module: according to the deviation amount, the competence node weight parameter and the competence verification model threshold are automatically adjusted; the initial question package is arranged according to the competence priority and the question depth rule; the adjusted parameters are fed back to the subsequent question generation process, and the multi-competence dimension mapping algorithm parameter and the competence verification model threshold are optimized.

[0089] In this embodiment, the construction of the initial multi-competence evaluation graph specifically includes the following steps:

[0090] Multi-source data acquisition and standardization preprocessing: the network crawler technology is used to crawl the job competence requirements of the industry recruitment platform, combined with the national occupation skill standard database, the enterprise internal competence manual and other authoritative data sources, the core competence dimension and the industry general competence element original data are collected; the original data is processed by the Python jieba word segmentation tool, the stop word removal is completed by using the NLTK library, the text data is converted into calculable vector features by using the Word2Vec model, and the original data standardization preprocessing is realized, which lays a foundation for subsequent screening and integration.

[0091] Core dimension and general element screening and integration: based on the preprocessed vector feature data, the TF-IDF algorithm is used to calculate the weight of each ability element, the core ability with high weight ranking is selected as the core competence dimension, and the core evaluation demand of different posts is adapted; the remaining general ability elements are classified and collected by using the K-means clustering algorithm, a multi-class industry general competence element cluster is formed, the core dimension and the general element cluster are integrated, and an initial dimension-element knowledge base is built.

[0092] Evaluation boundary definition and association rule setting: the natural language understanding technology is used to analyze the core semantics of each dimension and element, a reasonable semantic similarity threshold is preset, the redundant elements with a semantic matching degree not meeting the standard are removed, and the evaluation boundaries of each dimension and element are defined; the association rules between the dimensions and the elements are set by using the production rule engine, the binding relationship between the core dimension and the corresponding general element is defined, the association rules are stored in the database, and a callable rule library is formed.

[0093] Initial weight ratio calculation: The initial weight calculation is carried out by combining the analytic hierarchy process and the entropy weight method. First, industry experts construct a pairwise comparison judgment matrix by the analytic hierarchy process to obtain subjective weights. Then, the historical evaluation data is processed by the entropy weight method to obtain objective weights. The subjective weights and objective weights are integrated by a weighted fusion method to determine the initial weight ratio of each core dimension and general element, and finally an initial multi-competency evaluation map with clear elements and explicit rules is output.

[0094] In this embodiment, the construction of the competency verification model specifically includes the following steps:

[0095] A three-layer architecture model covering the recognition layer, the deep matching layer, and the balance evaluation layer is built using the TensorFlow framework. Real-time data intercommunication between the three layers is achieved through RESTful API to ensure efficient and smooth data flow. The model input is the initial question text and the multi-competency evaluation map, and the output is the list of missing dimensions, the list of questions with depth mismatch, and the balance score and optimization direction, meeting the full-dimensional verification needs of the questions. The TF-IDF algorithm is used to extract the keyword features of the initial question text, and the BM25 algorithm is used to calculate the matching degree of the keywords and each dimension in the multi-competency evaluation map. A reasonable matching threshold is preset. If the matching degree meets the standard, it is marked as a covered dimension, otherwise it is marked as an uncovered dimension. F1-score is introduced as an evaluation index for coverage recognition. Model training and parameter optimization are carried out based on labeled samples to ensure the accuracy and reliability of the coverage recognition results. The depth of the question is divided into three categories: basic cognition, advanced application, and deep analysis, which are matched one by one with each competency dimension. A BERT text classification model is trained based on labeled samples. After the input question text is encoded by BERT, the classification probability of the corresponding depth category is output through the full connection layer. The category with the highest probability is the depth level to which the question belongs. Cross-entropy loss function is used to iteratively optimize the model to improve the accuracy of question depth classification and adapt to different evaluation accuracy needs. The coefficient of variation method is used to build the balance calculation logic. The coefficient of variation is the ratio of the standard deviation to the mean of the number of questions for each dimension. The smaller the coefficient of variation, the more balanced the question distribution. A reasonable interval for the number of questions for each dimension is preset based on evaluation needs. The deviation of the actual number of questions for each dimension from the reasonable interval is calculated. A balance determination standard is preset. The balance score and targeted optimization direction are output based on the deviation and the coefficient of variation, and the overall construction and deployment of the competency verification model is finally completed.

[0096] In this embodiment, in the context of a technical research and development position in a campus recruitment of an Internet company, the candidate's "technical professional ability, learning ability, problem solving ability, team cooperation ability" four core competencies and "responsibility, execution" two general competencies need to be evaluated. The scene is a large-scale standardized screening (200+ people are evaluated within 1 day), there is no candidate pre-resume, the requirement is to cover all dimensions, and it is suitable for the cognition of college graduates. The specific process is as follows:

[0097] An optimized evaluation system is constructed to collect industry campus recruitment competency requirements, enterprise internal standards, etc. Core and general competency elements are selected. Redundant content such as "project management experience" is excluded, and element association rules (such as "technical professional ability" associated with "problem solving ability") are set. Combined with the campus recruitment scene, the priority is adjusted (technical professional ability > problem solving ability > others), and a multi-competency evaluation map is formed. The initial question is generated by double-mode scheduling. Since the scene is a large-scale standardized evaluation, the "batch combination generation mode" is triggered. Analyze the association relationship of each competency dimension (such as "code writing" and "debug optimization" progression, "technical ability" and "problem solving" complement each other), and generate basic questions in batches according to the association logic + priority. Optimize the expression to adapt to college graduates, eliminate out-of-class problems, and form an initial question set. Check the question coverage, find that the "technical professional ability - architecture design" dimension is missing, check the question depth, find that "team cooperation ability" only has basic cognitive questions (such as "how to understand team cooperation"), and lacks application questions. Evaluate the proportion of questions in each dimension, find that there are too many technical questions, and output optimization requirements: supplement architecture design and team cooperation application questions, and reduce the proportion of technical questions. Optimize the output and iterate in reverse according to the optimization requirements. Supplement and adjust the questions, group them according to the "core competency priority", sort them within the same group according to "basic cognition → advanced application", and mark the corresponding dimension of each question to form a standardized initial question package (20 questions in total). The generated results are fed back to the optimization evaluation system parameters to improve the adaptability of subsequent similar scenes;

[0098] Core excerpts of the initial question package: technical professional ability - basic cognition: "Please explain the core syntax of the programming language you are most familiar with"; technical professional ability - advanced application: "Please design a simple user login module architecture"; problem solving ability - advanced application: "If the code debugging fails multiple times, how will you handle it"; team cooperation ability - advanced application: "If the team has different opinions on the technical solution, how will you promote consensus".

[0099] The above merely describes the preferred embodiments of the present application, and is not intended to limit the present application in any form. Although the present application has been disclosed with the preferred embodiments as above, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content to obtain equivalent embodiments with equivalent changes, as long as the changes or modifications do not deviate from the technical solution of the present application. Any modification, change, equivalent change and modification of the above embodiments made according to the technical essence of the present application still belong to the scope of the technical solution of the present application.

Claims

1. A multi-capable intelligent question generation method, characterized in that, The application relates to a multi-competency evaluation method and system. S1: acquiring core competency dimensions and industry general competency elements, and constructing an initial multi-competency evaluation graph; according to an evaluation target parameter, a multi-competency evaluation graph containing competency nodes, node priority weights and node correlation relationships is constructed through a multi-competency dimension mapping algorithm; S2: based on the evaluation target parameter and the multi-competency evaluation graph, state variables representing the state of an evaluation scene are extracted; according to the state variables and a preset generation strategy judgment rule, a double-mode dynamic combination scheduling mechanism is constructed, including a batch combination generation mode and a progressive combination generation mode; in the batch combination generation mode, an initial question set is generated in batches according to correlation logic, and in the progressive combination generation mode, questions are generated in sequence according to priority and dynamically supplemented with uncovered dimensions; S3: a competency verification model is built, and a competency verification operation is performed on the initial question set, competency coverage labels and question depth labels corresponding to each question are extracted; Based on preset competency coverage thresholds, question depth matching thresholds and question distribution balance thresholds, the deviation of the question set in terms of coverage integrity, depth adaptability and distribution balance is calculated; S4: according to the deviation, the competency node weight parameter and the competency verification model threshold are automatically adjusted; the initial question package is sorted according to the competency priority and the question depth rule; The adjusted parameters are fed back to the subsequent question generation process, and the multi-competency dimension mapping algorithm parameters and the competency verification model threshold are optimized.

2. The method of claim 1, wherein, The specific process of constructing the initial multi-competency evaluation graph is as follows: Integrate the screened core competency dimensions and industry general competency elements, obtain the core evaluation direction corresponding to each element; According to the industry general evaluation standard and the basic evaluation demand, the evaluation boundary of each competency dimension is preset, and the redundant and irrelevant evaluation elements are removed; Set the initial correlation rule between each competency dimension and the general element, and distribute the initial weight proportion of each element.

3. The method of claim 1, wherein, The specific process of constructing the multi-competency evaluation graph containing competency nodes, node priority weights and node correlation relationships is as follows: Input the evaluation target difference related parameters, and build a dimension evaluation index system; Based on the evaluation index system, the fitness degree of each competency constituent element and the evaluation target is calculated; Adjust the correlation logic of the competency constituent elements in combination with the evaluation target, and redistribute the priority order and weight proportion of each element by using the hierarchical analysis method; Optimization effect is verified through adaptability check, and the optimized multi-competency evaluation mechanism is determined after the check meets the requirements, otherwise the elements and weights are repeatedly adjusted until the requirements are met.

4. The method of claim 1, wherein, The specific process of extracting state variables representing the state of an evaluation scene is as follows: presetting the evaluation scale, evaluation rhythm, evaluation form and pre-data availability core dimension; obtaining the original data corresponding to each dimension; standardizing the original data and completing the missing values; generating scale, rhythm, form and data availability state variables through weighted summation to form a variable set.

5. The method of claim 1, wherein, The specific process of constructing the dual-mode dynamic combination scheduling mechanism is: presetting the trigger conditions of the two modes of batch combination and progressive combination; extracting the associated characteristics of each competency dimension; establishing the matching relationship between the two modes and the corresponding question generation strategy; and executing the question generation logic after selecting the mode.

6. The method of claim 1, wherein, The specific process of generating the initial question set is: combining the dimension association logic and competency priority in the multi-competency evaluation graph, and performing intelligent dynamic generation operation through the selected generation mode; in the batch combination generation mode, based on the competency element semantic fusion algorithm, the core evaluation points of each competency are extracted in the dimension association order, the basic questions are generated through cross-dimensional element semantic deep cross-fusion, and then the questions are batched and output covering the core competency and the industry general competency full-dimension questions after scene-based semantic optimization, and finally the initial question set containing multi-competency questions is formed.

7. The method of claim 1, wherein, The specific process of generating questions in the progressive combination generation mode according to the priority and dynamically supplementing the uncovered dimensions is: sorting according to the optimized competency priority; real-time statistics of the covered dimensions corresponding to the generated questions are compared with the full dimensions in the multi-competency evaluation graph; after identifying the uncovered dimensions, the supplement generation order is determined according to the association logic, and the corresponding questions are generated in turn until the full dimensions are covered.

8. The method of claim 1, wherein, The specific process of building the competency verification model is: It includes: coverage identification layer, deep matching layer and balance evaluation layer; The coverage identification layer extracts the core semantic features of each initial question and converts them into competency labels, matches the labels with the dimensions in the multi-competency evaluation graph, and clearly marks the covered dimensions and the uncovered dimensions; The deep matching layer presets the question depth standard corresponding to each competency dimension, and determines the depth level corresponding to each question; The balance evaluation layer builds a balance calculation logic, and sets a threshold interval for the number of questions for each dimension according to the evaluation requirements.

9. The method of claim 1, wherein, The specific process of extracting the competency coverage label and the question depth label corresponding to each question is: Through the semantic feature extraction algorithm, the core semantic information of the question is captured, and the semantic similarity matching with the competency nodes in the multi-competency evaluation graph is performed to determine the corresponding covered competency dimensions and generate the competency coverage label; Based on the semantic depth and evaluation direction of the question, the depth level is automatically determined by the semantic level analysis algorithm, and the question depth label is generated.

10. The method of claim 1, wherein, The specific process of calculating the deviation amount of the question set in coverage integrity, depth adaptability and distribution balance is: the coverage integrity deviation amount is calculated by the relative deviation of the actual coverage dimension proportion and the preset threshold; the depth adaptability deviation amount is the weighted sum of the absolute difference value of the actual proportion of each depth level and the threshold; the distribution balance deviation amount is the average deviation rate of the number of questions of each dimension and the median value in the preset interval; and the comprehensive deviation amount set is output by integration.

11. The method of claim 1, wherein, The specific process of sorting the initial question package according to the competence priority and question depth rules is: extracting the competence priority and question depth level corresponding to each question in the optimized question package; grouping and classifying the questions according to the competence priority from high to low, and grouping the questions of the same priority competence into the same group; sorting the questions in each group according to the question depth in turn; and integrating all sorted questions according to the group priority.

12. A multi-capable intelligent question generation system, characterized in that, Comprise: A multi-competence evaluation mechanism construction module: obtain core competence dimensions and industry general competence elements, and construct an initial multi-competence evaluation graph; according to the evaluation target parameters, a multi-competence evaluation graph containing competence nodes, node priority weights and node association relationships is constructed through a multi-competence dimension mapping algorithm; A dual-mode combination and initial question generation module: based on the evaluation target parameters and the multi-competence evaluation graph, state variables representing the state of the evaluation scene are extracted; according to the state variables and the preset generation strategy determination rules, a dual-mode dynamic combination scheduling mechanism is constructed, including a batch combination generation mode and a progressive combination generation mode; the initial question set is generated in batches according to the association logic in the batch combination generation mode, and the questions are generated in turn according to the priority and dynamically supplemented with the uncovered dimensions in the progressive combination generation mode; An initial question competence verification module: a competence verification model is built to perform competence verification operation on the initial question set, and the competence coverage label and question depth label corresponding to each question are extracted; Based on the preset competence coverage threshold, question depth matching threshold and question distribution balance threshold, the deviation of the question set in coverage integrity, depth adaptability and distribution balance is calculated; An initial question package output module: according to the deviation, automatically adjust the competence node weight parameter and the competence verification model threshold; Sort the initial question package according to the competence priority and question depth rules; The adjusted parameters are fed back to the subsequent question generation process to optimize the multi-competence dimension mapping algorithm parameters and the competence verification model threshold.

Images