A large model-based personnel online auditing method and system

Abstract

The application relates to the technical field of personnel online auditing, and discloses a personnel online auditing method and system based on a large model, which comprises the following steps: performing feature extraction on pre-acquired multi-modal data of a target person, and constructing a multi-modal feature vector; performing analysis on the target ability and behavior intention of the target person through an execution sandbox and a track analysis sandbox in a pre-constructed double-layer sandbox architecture, to obtain an ability analysis result; configuring a dynamic question and answer mechanism, dynamically adjusting the question and answer content of the target person, and obtaining a dynamic question and answer result; combining the ability analysis result and the dynamic question and answer result, evaluating the ability of the target person, and generating an ability evaluation report; the application can comprehensively evaluate the ability of the target person from the two dimensions of result and process by constructing the double-layer sandbox architecture, can formulate individual evaluation content for different persons, can realize online real-time auditing of the ability of the target person, and can improve the comprehensiveness and accuracy of the ability evaluation result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of personnel online auditing, and more particularly to a personnel online auditing method and system based on a large model. BACKGROUND

[0002] Traditional personnel ability assessment and auditing methods rely on manual assessment, paper material auditing and standardized testing methods. The auditing process is complex, the auditing data of the auditing personnel is large, it is difficult to grasp the auditing standards, it is difficult to implement different auditing methods for different personnel, dynamic auditing, and there are problems such as low auditing efficiency, strong subjectivity, limited auditing range, etc.

[0003] The prior art has the following problems: based on a single type of data, the comprehensive ability of the evaluated person cannot be fully captured; using a fixed evaluation process and standardized questions, the evaluation content and difficulty cannot be adjusted according to the real-time performance of the evaluated person, and it is difficult to effectively evaluate the generalization ability and problem solving ability of the personnel when facing unknown problems; the ability evaluation is based on the final result, and the analysis of the execution process is ignored, which leads to the inability to accurately identify the thinking mode, coding style and potential ability short board of the personnel, and the ability analysis result is inaccurate and incomplete; in order to solve at least one of the above problems, the present application provides a personnel online auditing method and system based on a large model. SUMMARY

[0004] In view of the deficiencies of the prior art, the purpose of the present application is to provide a personnel online auditing method and system based on a large model, which can effectively solve the problems in the background art. The specific technical solution of the present application is as follows:

[0005] A personnel online auditing method based on a large model, comprising:

[0006] performing feature extraction on pre-acquired multi-modal data of a target person to construct a multi-modal feature vector;

[0007] based on the multi-modal feature vector, analyzing the target ability and behavior intention of the target person through the execution sandbox and the trajectory analysis sandbox in the pre-constructed double-layer sandbox architecture respectively to obtain an ability analysis result;

[0008] configuring a dynamic question and answer mechanism according to the ability analysis result to dynamically adjust the question and answer content of the target person to obtain a dynamic question and answer result;

[0009] combining the ability analysis result and the dynamic question and answer result, evaluating the ability of the target person through a pre-set ability evaluation model to generate an ability evaluation report for online auditing of the personnel.

[0010] Specifically, the feature extraction on the pre-acquired multi-modal data of the target personnel is performed to construct a multi-modal feature vector, including:

[0011] Acquiring multi-modal data of a target personnel, wherein the multi-modal data includes text data, code data and audio data;

[0012] Respectively performing feature extraction on the text data, the code data and the audio data through a preset feature extraction model to obtain a first feature set, a second feature set and a third feature set;

[0013] Combining the first feature set, the second feature set and the third feature set, a multi-modal feature vector is constructed.

[0014] Specifically, based on the multi-modal feature vector, the target ability and the behavior intention of the target personnel are analyzed through an execution sandbox and a trajectory analysis sandbox in a pre-constructed double-layer sandbox architecture, and an ability analysis result is obtained, including:

[0015] Based on the multi-modal feature vector, a double-layer sandbox architecture including an execution sandbox and a trajectory analysis sandbox is constructed;

[0016] The code data in the multi-modal data is input into the execution sandbox, the code is simulated and run, and the target ability of the target personnel is analyzed to obtain a first target ability;

[0017] The code data is input into the trajectory analysis sandbox, the execution process trajectory of the code is collected, and the behavior intention of the target personnel is analyzed to obtain a first behavior intention;

[0018] The first target ability and the first behavior intention are associated and compared to analyze, the ability advantage and the ability short board of the target personnel are positioned, and an ability analysis result is obtained.

[0019] Specifically, the code data in the multi-modal data is input into the execution sandbox, the code is simulated and run, and the target ability of the target personnel is analyzed to obtain a first target ability, including:

[0020] The code data in the multi-modal data is input into the execution sandbox, the code is simulated and run, and in the running process, the code running index is collected;

[0021] According to the code running index, the time complexity and the space complexity are calculated, the target ability of the target personnel is analyzed, and a first target ability is obtained.

[0022] Specifically, the code data is input into the trajectory analysis sandbox, the execution process trajectory of the code is collected, and the behavior intention of the target personnel is analyzed to obtain a first behavior intention, including:

[0023] Input the code data into the track analysis sandbox for execution, and collect track data during code execution;

[0024] According to the track data, identify the coding style and logical correlation degree through a preset track analysis model, and construct a track index vector;

[0025] Based on the track index vector, analyze the behavior intention of the target person, and obtain a first behavior intention.

[0026] Specifically, according to the ability analysis result, a dynamic question and answer mechanism is configured to dynamically adjust the question and answer content of the target person, and a dynamic question and answer result is obtained, including:

[0027] According to the ability analysis result, the positions of the ability advantages and the ability short boards are screened out, and scene switching points and difficulty adjustment points are respectively set at the corresponding positions;

[0028] At the scene switching point, the scene switching content in the question and answer content is identified, and a replacement scene is generated to replace the scene switching content, and a first adjustment result is obtained;

[0029] At the difficulty adjustment point, the ability matching result of the question and answer content and the target person is analyzed, a corresponding difficulty adjustment instruction is generated to perform real-time difficulty adjustment, and a second adjustment result is obtained;

[0030] The question and answer content of the target person is dynamically adjusted in combination with the first adjustment result and the second adjustment result, and a dynamic question and answer result is obtained.

[0031] Specifically, the scene switching point, the scene switching content in the question and answer content is identified, and a replacement scene is generated to replace the scene switching content, and a first adjustment result is obtained, including:

[0032] At the scene switching point, the scene switching content in the question and answer content is identified by analyzing the content corresponding to the ability advantages and the ability short boards;

[0033] In combination with the context association of the scene switching content, a corresponding replacement scene is generated;

[0034] According to the replacement scene, the scene switching content is dynamically replaced, the corresponding scene content is reconstructed, and a first adjustment result is obtained.

[0035] Specifically, the difficulty adjustment point, the ability matching result of the question and answer content and the target person is analyzed, a corresponding difficulty adjustment instruction is generated to perform real-time difficulty adjustment, and a second adjustment result is obtained, including:

[0036] At the difficulty adjustment point, the ability matching result of the question and answer content and the target person is analyzed, and a first matching result is obtained;

[0037] match corresponding difficulty adjustment instructions in a pre-constructed difficulty instruction library based on the first matching result;

[0038] In response to the difficulty adjustment instructions, the question and answer content is rewritten in real time to adjust the difficulty, and a second adjustment result is obtained.

[0039] Specifically, in combination with the ability analysis result and the dynamic question and answer result, the ability of the target person is evaluated through a pre-set ability evaluation model to generate an ability evaluation report, including:

[0040] The ability analysis result and the dynamic question and answer result are analyzed in context to extract a corresponding set of ability indicators.

[0041] According to the set of ability indicators, the ability of the target person is evaluated through a pre-set ability evaluation model to generate an ability evaluation report.

[0042] A personnel online auditing system based on a large model is used to implement the personnel online auditing method based on a large model, and includes:

[0043] The data analysis module extracts features from the pre-acquired multi-modal data of the target person to construct a multi-modal feature vector.

[0044] The ability analysis module analyzes the target ability and behavior intention of the target person through the execution sandbox and the trajectory analysis sandbox in the pre-constructed double-layer sandbox architecture based on the multi-modal feature vector to obtain an ability analysis result.

[0045] The question and answer adjustment module configures a dynamic question and answer mechanism to dynamically adjust the question and answer content of the target person according to the ability analysis result to obtain a dynamic question and answer result.

[0046] The personnel auditing module evaluates the ability of the target person through a pre-set ability evaluation model in combination with the ability analysis result and the dynamic question and answer result to generate an ability evaluation report for online auditing of the personnel.

[0047] The beneficial effects of the present application: based on multi-modal data for feature extraction and fusion, construct a multi-modal feature vector, construct a double-layer sandbox architecture containing an execution sandbox and a trajectory analysis sandbox, analyze the target ability and behavior intention of the target personnel, evaluate the comprehensive ability of the target personnel, set the scene switching point and difficulty adjustment point according to the ability analysis result, dynamically adjust the question and answer content, at the scene switching point, generate a replacement scene by analyzing the context association, dynamically reconstruct the scene content, at the difficulty adjustment point, adjust the question and answer difficulty in real time based on the ability matching result, combine the ability analysis result and the dynamic question and answer result to comprehensively analyze the comprehensive ability of the target personnel and generate an ability evaluation report. By fusing multi-modal data and constructing a double-layer sandbox architecture, the ability of the target personnel can be comprehensively evaluated from the result and process dimensions, the evaluation content and difficulty can be dynamically adjusted according to the real-time performance of the target personnel, personalized evaluation content can be developed for different personnel, the real ability level of the evaluated person can be more effectively mined, misjudgment caused by too difficult or too easy evaluation content can be avoided, online real-time auditing of the target personnel's ability is realized, and the comprehensiveness and accuracy of the ability evaluation result are improved. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A workflow diagram of a personnel online auditing method based on a large model in an embodiment of the present application;

[0049] Figure 2 A schematic diagram of a double-layer sandbox architecture in an embodiment of the present application;

[0050] Figure 3 A schematic diagram of a scene switching point and a difficulty adjustment point in an embodiment of the present application;

[0051] Figure 4 A structural schematic diagram of a personnel online auditing system based on a large model in an embodiment of the present application. DETAILED DESCRIPTION

[0052] The present application will be further described in detail below in conjunction with the drawings and embodiments. Identical parts are denoted by identical reference numerals in the following description. It should be noted that the words "front", "back", "left", "right", "up" and "down" used in the following description refer to the directions in the drawings, and the words "bottom surface" and "top surface", "inner" and "outer" refer to the directions towards or away from the geometric center of a particular part.

[0053] In the embodiments of the present application, the words "exemplary" or "for example" are used to mean serving as an example, instance, or illustration. Any embodiment or design presented as "exemplary" or "for example" in the embodiments of the present application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Rather, the use of the words "exemplary" or "for example" is intended to present concepts in a concrete manner.

[0054] Hereinafter, the terms "first", "second", and the like are generic terms, and are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first" and "second" can explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more.

[0055] Reference Figure 1 As shown, the personnel online auditing method based on a large model provided by the present application comprises:

[0056] S101, feature extraction is performed on the pre-acquired multi-modal data of the target personnel, and a multi-modal feature vector is constructed;

[0057] S102, based on the multi-modal feature vector, the target ability and behavior intention of the target personnel are analyzed by executing the sandbox and trajectory analysis sandbox in the pre-constructed double-layer sandbox architecture respectively, and an ability analysis result is obtained;

[0058] S103, according to the ability analysis result, a dynamic question and answer mechanism is configured, the question and answer content of the target personnel is dynamically adjusted, and a dynamic question and answer result is obtained;

[0059] S104, combining the ability analysis result and the dynamic question and answer result, the ability of the target personnel is evaluated through a pre-set ability evaluation model, an ability evaluation report is generated, and the personnel is audited online.

[0060] With the development of remote office and online recruitment, the demand of enterprises for talent ability evaluation is increasingly diversified and accurate, and the traditional auditing method has been difficult to meet the requirements of modern talent evaluation. The present application provides a multi-modal data fusion and dynamic ability evaluation mechanism, which can effectively solve the problems of low efficiency and limited coverage of traditional personnel auditing methods, and meet the demand of large-scale and high-quality online auditing.

[0061] In the embodiment, feature extraction is performed on pre-acquired multi-modal data of the target personnel to construct a multi-modal feature vector; the multi-modal data includes text data, code data and audio data, for different types of data, corresponding preset feature extraction models are used for feature extraction, the extracted features are subjected to feature alignment processing, feature weights are assigned according to the contribution of different features to personnel capability evaluation, the features and corresponding weights are weighted and fused to obtain the multi-modal feature vector. By combining multi-modal data for feature extraction and fusion, multi-dimensional performance of personnel capability can be analyzed, and one-sidedness of evaluation caused by single data can be avoided; by weighted fusion of different features, complementary information of multiple features can be effectively integrated, feature redundancy or information loss can be avoided, and the constructed multi-modal feature vector can comprehensively and accurately reflect the comprehensive capability characteristics of the target personnel, providing a high-quality data basis for capability analysis of the target personnel.

[0062] Specifically, based on the multi-modal feature vector, the execution sandbox and the trajectory analysis sandbox in the pre-constructed double-layer sandbox architecture are used to analyze the target capability and the behavior intention of the target personnel respectively to obtain the capability analysis result; according to the multi-modal feature vector, a double-layer sandbox architecture is constructed, including an execution sandbox and a trajectory analysis sandbox which are independent of each other and data can be interacted, the execution sandbox is used to analyze the target capability of the target personnel, simulates the running of the code and monitors the running index in real time, and analyzes the target capability of the target personnel in combination with the index; the trajectory analysis sandbox is used to analyze the behavior intention of the target personnel, collects complete trajectory data in the code execution process, identifies the coding style and logical correlation degree of the target personnel, and analyzes the behavior intention of the target personnel; the target capability and the behavior intention are combined to analyze the capability advantages and short boards of the target personnel to obtain the capability analysis result. The double-layer sandbox architecture realizes the combination of result evaluation and process analysis, can capture the behavior intention characteristics ignored in the traditional review process, improves the comprehensiveness of the personnel capability evaluation process, accurately locates the core advantages and short boards of the personnel capability, and obtains accurate capability analysis result.

[0063] After analyzing the ability of the target person to obtain the ability analysis result, a dynamic question and answer mechanism is configured, the question and answer content of the target person is dynamically adjusted, and a dynamic question and answer result is obtained; the ability advantages and short boards are screened based on the ability analysis result, and the scene switching points are set in the question and answer links corresponding to the advantage dimensions, and the difficulty adjustment points are set in the question and answer links corresponding to the short board dimensions; the scene switching content is generated in combination with the context association at the scene switching points, the scene is replaced for the question and answer content, and the difficulty of the question and answer content is adjusted in real time at the difficulty adjustment points; the question and answer content of the target person is dynamically adjusted in combination with the scene switching and the difficulty adjustment, and a dynamic question and answer result is obtained. By setting the scene switching points and the difficulty adjustment points to dynamically adjust the scene and the difficulty of the question and answer content respectively, the potential ability of the target person can be deeply analyzed in the ability advantage dimension, and the problem can be accurately located in the short board dimension, so that the dynamic question and answer result is more targeted and comprehensive.

[0064] Specifically, in combination with the ability analysis result and the dynamic question and answer result, the ability of the target person is evaluated through a preset ability evaluation model to generate an ability evaluation report; the ability analysis result and the dynamic question and answer result are analyzed in combination with the context association, the corresponding ability indicators are extracted, the ability of the target person is evaluated through the preset ability evaluation model, and an ability evaluation report is generated to perform online auditing on the personnel. Through the context association analysis, the correlation between the ability analysis result and the dynamic question and answer result can be analyzed, and the ability of the target person can be analyzed in combination with the ability analysis result and the dynamic question and answer result, so that the accuracy and comprehensiveness of the analysis result can be improved.

[0065] Based on the feature extraction and fusion of the multi-modal data, the multi-modal feature vector is constructed, a double-layer sandbox architecture including an execution sandbox and a trajectory analysis sandbox is constructed, the target ability and behavior intention of the target person are analyzed, the comprehensive ability of the target person is evaluated, the scene switching points and the difficulty adjustment points are set according to the ability analysis result, the question and answer content is dynamically adjusted, the scene is replaced by analyzing the context association at the scene switching points, the scene content is dynamically reconstructed, the question and answer difficulty is adjusted in real time based on the ability matching result at the difficulty adjustment points, the comprehensive ability of the target person is analyzed in combination with the ability analysis result and the dynamic question and answer result, and an ability evaluation report is generated. By fusing the multi-modal data and constructing the double-layer sandbox architecture, the ability of the target person can be comprehensively evaluated from the result and the process dimensions, the evaluation content and the difficulty can be dynamically adjusted according to the real-time performance of the target person, the personalized evaluation content can be formulated for different personnel, the real ability level of the evaluated person can be more effectively mined, the misjudgment caused by the over-difficult or over-easy evaluation content can be avoided, the online real-time auditing of the ability of the target person is realized, and the comprehensiveness and accuracy of the ability evaluation result are improved.

[0066] Further, the multi-modal data of the target person is pre-acquired, features are extracted, and a multi-modal feature vector is constructed, including:

[0067] S201, obtaining multi-modal data of a target person, wherein the multi-modal data comprises text data, code data and audio data;

[0068] S202, performing feature extraction on the text data, the code data and the audio data respectively by a preset feature extraction model to obtain a first feature set, a second feature set and a third feature set;

[0069] S203, constructing a multi-modal feature vector in combination with the first feature set, the second feature set and the third feature set.

[0070] In the embodiment, multi-modal data of a target person is obtained, and the multi-modal data comprises text data, code data and audio data. The text data includes but is not limited to materials submitted by the target person actively, the code data includes but is not limited to codes input by the target person in a real-time manner during an examination process, and the audio data includes but is not limited to a recording of a technical interview participated by the target person. It should be noted that personal privacy data involved in the present application has been agreed, and the obtained data is preprocessed, such as cleaning and normalization, to obtain preprocessed multi-modal data. By obtaining and preprocessing the multi-modal data, the evaluation one-sidedness caused by single data can be avoided, and rich and complete data can be provided for the analysis and evaluation of the ability of the target person.

[0071] Specifically, feature extraction is performed on the text data, the code data and the audio data respectively by a preset feature extraction model to obtain a first feature set, a second feature set and a third feature set. For the text data, the feature extraction model includes but is not limited to a BERT model based on a Transformer architecture. The BERT model is trained by using a large amount of historical text data to obtain a pre-trained BERT model. The text data is input into the pre-trained BERT model, the model analyzes the semantics of the text data and extracts corresponding semantic features to obtain the first feature set. For the code data, the feature extraction model includes but is not limited to a CodeBERT model adapted to code semantic understanding. The CodeBERT model is trained by using a large amount of historical code data to obtain a pre-trained CodeBERT model. The code data is input into the pre-trained CodeBERT model, the model analyzes the syntax logic and semantic association of the code data and extracts corresponding features to obtain the second feature set. For the audio data, the feature extraction model includes but is not limited to an LSTM model. The LSTM model is trained by using a large amount of historical audio data to obtain a pre-trained LSTM model. The audio data is input into the pre-trained LSTM model, the model analyzes the acoustic characteristics of the audio data and extracts corresponding features to obtain the third feature set.

[0072] It should be emphasized that the feature extraction of the text data, the code data and the audio data through the corresponding feature extraction model respectively obtains the corresponding feature set, which can avoid the limitation that a single feature cannot reflect the comprehensive ability of the target personnel, and the three types of feature sets correspond to the written expression ability, the programming ability and the language expression ability of the target personnel, thereby providing rich feature data for the ability analysis of the target personnel and improving the accuracy of the analysis process.

[0073] Specifically, the features in the first feature set, the second feature set and the third feature set are subjected to feature alignment processing, the weights of the corresponding features are set according to the importance of the features in the personnel ability evaluation, the features are weighted and summed according to the corresponding weights, and a multi-modal feature vector is obtained; through the fusion of the features extracted from the multi-modal data, the complementary information between the written, programming and language ability features of the target personnel can be integrated, the comprehensive ability of the target personnel can be reflected, and the accuracy of the ability analysis process of the target personnel can be improved.

[0074] Further, based on the multi-modal feature vector, the target ability and the behavior intention of the target personnel are analyzed through the execution sandbox and the trajectory analysis sandbox in the pre-constructed double-layer sandbox architecture, and an ability analysis result is obtained, including:

[0075] S301, based on the multi-modal feature vector, a double-layer sandbox architecture including an execution sandbox and a trajectory analysis sandbox is constructed;

[0076] S302, the code data in the multi-modal data is input into the execution sandbox, the code is simulated and run, and the target ability of the target personnel is analyzed, and a first target ability is obtained;

[0077] S303, the code data is input into the trajectory analysis sandbox, the execution process trajectory of the code is collected, the behavior intention of the target personnel is analyzed, and a first behavior intention is obtained;

[0078] S304, the first target ability and the first behavior intention are associated and compared, the ability advantage and the ability short board of the target personnel are positioned, and an ability analysis result is obtained.

[0079] As Figure 2As shown, based on the multi-modal feature vector, a double-layer sandbox architecture including an execution sandbox and a trajectory analysis sandbox is constructed; a virtual machine is used to build a technical foundation of the double-layer sandbox architecture to ensure resource independence between the two sandboxes, and the execution sandbox environment is configured according to the programming language type of the code data extracted from the multi-modal feature vector, to ensure that the code can run smoothly; the execution sandbox includes a compilation tool library, a test case library and a resource monitoring tool, the compilation tool ensures that the code can be normally compiled, the test case performs process testing on the code, and the resource monitoring tool collects running indicators in real time during code running; the trajectory analysis sandbox is configured, which includes a trajectory recording tool, a log analysis tool and a trajectory analysis tool, the trajectory recording tool implants a tracking probe during code compilation or running to record the complete trajectory of code execution, the log analysis tool extracts and analyzes key events in the log, and the trajectory analysis tool analyzes the code execution trajectory in real time.

[0080] Meanwhile, a connection is established between the execution sandbox and the trajectory analysis sandbox through an encrypted API interface, bidirectional data synchronization between the execution sandbox and the trajectory analysis sandbox is realized, and code data and running states can be shared in real time. The double-layer sandbox architecture can guarantee the isolation between the two sandboxes, the execution sandbox analyzes the running results of the code, and the trajectory analysis sandbox analyzes the trajectory during code execution, to provide data support for the ability and intention analysis of the target personnel, and the comprehensive ability of the target personnel can be analyzed in combination with the code ability and intention, to improve the comprehensiveness and accuracy of the analysis process.

[0081] Specifically, the code data in the multi-modal data is input into the execution sandbox, the code is simulated to run, and the target ability of the target personnel is analyzed to obtain a first target ability; the code running results and compilation error information can be obtained through the code simulation, to realize quantitative analysis of the target ability of the target personnel, calculate and analyze the code ability of the target personnel, and obtain an accurate target ability evaluation result.

[0082] Specifically, the code data is input into the trajectory analysis sandbox, the execution process trajectory of the code is collected, and the behavior intention of the target personnel is analyzed to obtain a first behavior intention; the logic correlation during code running can be captured through the collection of the execution process trajectory of the code to obtain an accurate behavior intention analysis result.

[0083] The first target ability and the first behavior intention are compared and analyzed in association, the first target ability is matched with the first behavior intention according to the corresponding relationship, if the target ability performance is good and the corresponding behavior intention performance is high, it is determined that the ability is strong, if the target ability performance is weak and the corresponding behavior intention performance is low, it is determined that the ability is short; for the ability dimension determined as the ability strong, the specific performance of the strong is specified, for the ability dimension determined as the ability short, the specific problem and influence of the short are determined; the analysis results of the ability strong and the ability short are integrated to obtain a structured ability analysis result. By associating and analyzing the first target ability and the first behavior intention, the result performance of the ability and the process characteristics of the intention are combined, the limitation of the traditional evaluation of only looking at the result is broken through, the strong ability and the short ability can be more accurately positioned, and the accuracy and comprehensiveness of the ability analysis result are improved.

[0084] Further, the code data in the multi-modal data is input into the execution sandbox, the code is simulated and run, and the target ability of the target person is analyzed to obtain the first target ability, including:

[0085] S401, input the code data in the multi-modal data into the execution sandbox, simulate and run the code, and collect code running indexes during the running process;

[0086] S402, calculate the time complexity and the space complexity according to the code running indexes, analyze the target ability of the target person, and obtain the first target ability.

[0087] In this embodiment, the code data in the multi-modal data is input into the execution sandbox, the code is simulated and run, and code running indexes are collected during the running process; when the code data is extracted from the multi-modal data, the code segment directly related to the target ability evaluation is located through the code type, the code segment is input into the execution sandbox, the execution sandbox automatically activates the corresponding compilation environment according to the code language to obtain the compilation result, if the compilation is error, the error type and data are recorded; the function, the boundary condition and the performance pressure of the code are tested according to the test case library in the execution sandbox, the code running indexes are collected through the monitoring tool in the execution sandbox during the testing process, the code running indexes include but are not limited to the total running time, the code block level time consumption, the memory peak value and the resource occupation. By simulating and running the code and collecting the running indexes, comprehensive data support can be provided for the target ability analysis process of the target person.

[0088] Specifically, the time complexity and the space complexity are calculated according to the code running index, the target ability of the target person is analyzed, and the first target ability is obtained; the time complexity is calculated by analyzing the change trend of the code execution time consumption with the data volume and combining the hierarchical structure of the code; for example, if the data volume increases to 2n, the time consumption increases to 4 times, and there is a double nested loop, it is determined that the time complexity is O(n2); according to the memory occupation of the variables and the data structure in the code, the static memory occupation benchmark value is determined according to the number of bytes, the memory occupation is analyzed with the data volume according to the code running index, the space complexity is calculated, for example, if the data volume increases from n to 2n, the memory peak value increases to approximately 2 times, and there is no memory leakage, it is determined that the space complexity is O(n); the time complexity and the space complexity are combined to determine the target ability level of the target person, if the time complexity is O(n2) or the space complexity is O(n2), it is determined to be intermediate; if the time complexity is greater than or equal to O(2 n ) or there is memory leakage, it is determined to be weak; if the time complexity is less than or equal to O(n) and the space complexity is less than or equal to O(n), it is determined to be strong, and the first target ability is obtained.

[0089] It should be emphasized that the time complexity is combined with the code structure and the running index, which can avoid the singleness of only relying on the code structure, and the calculated time complexity is more consistent with the actual running performance of the code; the space complexity is combined with the static and dynamic memory, which avoids the defects of the traditional calculation of only static memory, and more comprehensively reflects the utilization ability of the code to resources; the target ability is analyzed by combining the time complexity and the space complexity of the code, which can improve the accuracy of the target ability analysis result.

[0090] Further, the code data is input into the trajectory analysis sandbox, the trajectory data in the code execution process is collected, and the behavior intention of the target person is analyzed, and the first behavior intention is obtained, including:

[0091] S501, inputting code data into a trajectory analysis sandbox for execution, and collecting trajectory data in the code execution process;

[0092] S502, according to the trajectory data, the coding style and the logical correlation degree are identified through a preset trajectory analysis model, and a trajectory index vector is constructed;

[0093] S503, based on the trajectory index vector, the behavior intention of the target person is analyzed, and the first behavior intention is obtained.

[0094] In the embodiment, the code data track is analyzed in the sandbox, the track data in the code execution process is collected, the code data in the execution sandbox which has passed the compilation is input into the track analysis sandbox, the corresponding dynamic instrumenting tool is loaded according to the code programming language type, the compatibility of the instrumenting tool and the code running environment is ensured, the track events to be collected are determined, including function call events, variable operation events, branch jump events and exception handling events, the dynamic instrumenting tool starts collecting the track data according to the track events, the complete track data collected is cleaned, and the preprocessed track data is obtained; the instrumenting tool is selected to reduce external interference and realize fine-grained track collection of language adaptation, avoid the problem that the traditional log printing only records the key nodes and the track is incomplete, and complete and accurate track data is obtained.

[0095] Specifically, according to the track data, the coding style and the logical correlation degree are identified through a preset track analysis model to construct a track index vector; coding style correlation features and logical correlation degree correlation features are extracted from the track data, the coding style correlation features include but are not limited to variable naming features and annotation features, and the logical correlation degree correlation features include but are not limited to function call features and exception handling features; the extracted features are input into the preset track analysis model to identify the coding style and the logical correlation degree, the track analysis model includes but is not limited to a graph neural network model, a large amount of historical track data is used to train the graph neural network model to obtain a pre-trained graph neural network model, the coding style correlation features and the logical correlation degree correlation features are input into the pre-trained graph neural network model, the model analyzes the coding style and the logical correlation degree, and outputs coding style scores and logical correlation degree scores; the coding style scores and the logical correlation degree scores are weighted and summed to construct the track index vector; by analyzing the coding style and the logical correlation degree, the scores output by the model can be converted into an index vector by combining the logical dependence between events, and clear data input is provided for behavior intention analysis.

[0096] Specifically, based on the track index vector, the behavior intention of the target person is analyzed through a preset intention analysis model to obtain a first behavior intention; the intention analysis model includes but is not limited to a neural network model, a large amount of historical track index vector data is used to train the neural network model to obtain a pre-trained neural network model, the track index vector is input into the pre-trained neural network model, the model analyzes and calculates the behavior intention of the target person to obtain a corresponding index dimension intention score and judges an intention level, the intention level includes high, medium and low, and the first behavior intention is obtained; by calculating the intention score, the behavior intention of the target person is analyzed, and the accuracy and comprehensiveness of the behavior intention analysis result are improved.

[0097] Further, according to the ability analysis result, a dynamic question and answer mechanism is configured to dynamically adjust the question and answer content of the target person to obtain a dynamic question and answer result, including:

[0098] S601. Based on the capability analysis results, identify the capability strengths and weaknesses, and set scene switching points and difficulty adjustment points at the corresponding locations.

[0099] S602. At the scene switching point, identify the scene switching content in the question and answer content, and generate a replacement scene to replace the scene switching content to obtain the first adjustment result;

[0100] S603. At the difficulty adjustment point, analyze the matching result between the question and answer content and the target personnel's ability, generate the corresponding difficulty adjustment instruction to adjust the difficulty in real time, and obtain the second adjustment result.

[0101] S604. Combine the first adjustment result and the second adjustment result to dynamically adjust the question and answer content of the target personnel to obtain dynamic question and answer results.

[0102] like Figure 3 As shown, based on the ability analysis results, strengths and weaknesses are identified, and scenario switching points and difficulty adjustment points are set at corresponding locations. Based on the strengths and weaknesses identified in the ability analysis, the corresponding strengths and weaknesses are located. For each strength, a scenario switching point is set in the corresponding assessment stage of the question-and-answer process, allowing for a more complex scenario and in-depth assessment of the target personnel's abilities. For each weakness, a difficulty adjustment point is set in the corresponding assessment stage of the question-and-answer process, dynamically adjusting the difficulty to avoid either being too difficult (resulting in an inability to answer) or too easy (failing to accurately analyze the personnel's abilities). By setting scenario switching points and difficulty adjustment points, the assessment scenario and difficulty can be adjusted in real time, ensuring that the question-and-answer content matches the target personnel's abilities, thus improving the accuracy and effectiveness of the assessment process.

[0103] Specifically, at scene transition points, the system identifies scene transition content within the question-and-answer content and generates a replacement scene to replace the transition content, resulting in the first adjustment result. At difficulty adjustment points, the system analyzes the matching results between the question-and-answer content and the target personnel's abilities, generates corresponding difficulty adjustment instructions, and adjusts the difficulty in real time, resulting in the second adjustment result. By switching scenes and adjusting difficulty in real time, the system can combine the real-time ability analysis results of the target personnel to adjust the question-and-answer content in real time, ensuring that the question-and-answer content matches the target personnel's abilities and improving the assessment effect.

[0104] Specifically, the question and answer content of the target person is dynamically adjusted according to the first adjustment result and the second adjustment result to obtain a dynamic question and answer result; logical consistency between the first adjustment result and the second adjustment result is analyzed and judged, if the adjustment correlation dimensions of the first adjustment result and the second adjustment result are different, there is no conflict, and the first adjustment result and the second adjustment result can be directly integrated; if the adjustment correlation dimensions are the same, at this time, the scene switching makes the difficulty increase, and the difficulty adjustment makes the difficulty decrease, and it is determined that there is a conflict, the difficulty adjustment priority of the ability short board is greater than the scene switching priority of the ability advantage, and the difficulty adjustment is preferentially performed; the first adjustment result and the second adjustment result after adjustment are integrated, the question and answer content of the target person is dynamically adjusted, and a dynamic question and answer result is obtained. Through logical consistency verification and priority setting, the content confusion problem caused by adjustment result conflict is avoided, the logicality and coherence of the question and answer content can be ensured, the adjustment effect of the question and answer content is improved, and therefore the accuracy of the examination process is improved.

[0105] Further, at the scene switching point, the scene switching content in the question and answer content is identified, and a replacement scene is generated to replace the scene switching content, to obtain a first adjustment result, comprising:

[0106] S701, at the scene switching point, the scene switching content in the question and answer content is identified by analyzing the content corresponding to the ability advantage and the ability short board;

[0107] S702, a corresponding replacement scene is generated in combination with the context association of the scene switching content;

[0108] S703, the scene switching content is dynamically replaced according to the replacement scene, and the corresponding scene content is reconstructed to obtain a first adjustment result.

[0109] In the embodiment, at the scene switching point, the scene switching content in the question and answer content is identified by analyzing the content corresponding to the ability advantage and the ability short board; the content corresponding to the ability advantage and the ability short board is extracted as the scene switching content in the question and answer content by analyzing the content at the scene switching point; the scene switching content is identified to provide data basis for scene switching and improve the scene switching efficiency.

[0110] Specifically, the corresponding replacement scene is generated in combination with the context association of the scene switching content; the context association information of the scene switching content is extracted, including the pre-context, technical background and interaction history, the scene switching template is generated according to the historical scene data, the replacement scene resource library is constructed, the scene switching template consistent with the industry field of the context association information, matching the technical background and higher than the original scene switching content in difficulty is screened from the replacement scene resource library, the corresponding replacement scene is generated according to the scene switching template, and the semantic coherence of the replacement scene to the context is analyzed and verified. By extracting the context association information, the problem that the replacement scene is separated from the actual question and answer context can be avoided, and the replacement scene is highly adapted to the context. By constructing the scene resource library, the corresponding scene switching template can be quickly matched, and the scene switching efficiency is improved.

[0111] Specifically, the scene switching content is dynamically replaced according to the replacement scene, the corresponding scene content is reconstructed, and a first adjustment result is obtained; the information in the replacement scene is contextually matched with the content of the scene switching content, the core information of the replacement scene is replaced in the scene switching content, and the original question and answer content is integrated, the core information including background, demand and difficulty; after integrating the replacement scene information, the semantic connection is checked, transition sentences are added to avoid logical break in the information fusion and splicing process; the replacement content is integrated to obtain the first adjustment result. By checking and adjusting the logical consistency of the replacement content, the unnatural problem caused by direct replacement is avoided, and the replacement logic of different types of question and answer content is adapted; by reconstructing the scene content, scene upgrading is realized while the interaction logic of the question and answer content is maintained, the question and answer are avoided due to replacement, and the quality of the question and answer content is improved.

[0112] Further, at the difficulty adjustment point, the matching result of the question and answer content and the target personnel's ability is analyzed, the corresponding difficulty adjustment instruction is generated to adjust the difficulty in real time, and a second adjustment result is obtained, including:

[0113] S801, at the difficulty adjustment point, analyzing the matching result of the question and answer content and the target personnel's ability, and obtaining a first matching result;

[0114] S802, matching the corresponding difficulty adjustment instruction in the pre-constructed difficulty instruction library based on the first matching result;

[0115] S803, in response to the difficulty adjustment instruction, the question and answer content is adjusted in real time, and a second adjustment result is obtained.

[0116] In the embodiment, at the difficulty adjustment node, the matching result of the question and answer content and the ability of the target person is analyzed to obtain a first matching result; the ability short board content of the position is analyzed at the difficulty adjustment node, the corresponding ability feature vector is extracted, the content difficulty of the position is analyzed, the features reflecting the content difficulty are extracted, including but not limited to technical depth and logical complexity, the content difficulty feature vector is constructed, the cosine similarity between the ability feature vector and the content difficulty feature vector is calculated, the similarity threshold is set according to the matching accuracy requirement, and if the cosine similarity is less than the similarity threshold, it indicates that the content difficulty of the position is not matched with the ability of the target person, and the first matching result is obtained. By analyzing the personnel ability features and the content difficulty features, the accuracy of the matching result can be improved, and it is ensured that the matching result meets the real ability level of the target person.

[0117] Specifically, the corresponding difficulty adjustment instruction is matched in the pre-constructed difficulty instruction library based on the first matching result; the difficulty adjustment instruction is generated according to the matching result, and for example, for the adjustment requirement of the operation rule of simplifying the problem condition and retaining the core examination point, a first difficulty adjustment instruction is generated to slightly reduce the difficulty; for the adjustment requirement of the operation rule of splitting the complex problem and reducing the operation requirement, a second difficulty adjustment instruction is generated to significantly reduce the difficulty; the difficulty instruction library is constructed based on the generated difficulty adjustment instruction, the operation rule is analyzed according to the first matching result, and the corresponding difficulty adjustment instruction is matched in the difficulty instruction library; by constructing the difficulty instruction library, reference is provided for instruction matching, the corresponding difficulty adjustment instruction is quickly matched in the difficulty instruction library based on the first matching result, and the difficulty adjustment efficiency is improved.

[0118] Specifically, in response to the difficulty adjustment instruction, the difficulty of the question and answer content is adjusted in real time, and a second adjustment result is obtained; the question and answer content is rewritten in real time according to the matched difficulty adjustment instruction, in the rewriting process, the expression style of the original question and answer content is retained, and the context coherence is analyzed in real time, after rewriting, the difficulty feature vector similarity before and after rewriting is calculated to check whether the rewritten content meets the difficulty level required by the instruction, and the difficulty of the rewritten content is verified; the rewritten content is integrated to obtain the second adjustment result. By rewriting the content in real time according to the difficulty adjustment instruction, it can be ensured that the rewritten content is unified with the original content style, and adapts to the difficulty requirement of the instruction, so that the rewritten content is logically coherent, and the quality of the rewritten content is improved.

[0119] Further, in combination with the ability analysis result and the dynamic question and answer result, the ability of the target person is evaluated through a pre-set ability evaluation model to generate an ability evaluation report, including:

[0120] S901, context association analysis is performed on the ability analysis result and the dynamic question and answer result, and a corresponding ability index set is extracted;

[0121] S902, according to the ability index set, the ability of the target personnel is evaluated by a preset ability evaluation model, and an ability evaluation report is generated.

[0122] In the embodiment, the context correlation analysis is performed on the ability analysis result and the dynamic question and answer result, and the corresponding ability index set is extracted; the context correlation analysis model based on the Transformer is constructed, the context correlation analysis model is trained by using a large number of historical ability analysis results and dynamic question and answer results, the pre-trained context correlation analysis model is obtained, the ability analysis result and the dynamic question and answer result are input into the pre-trained context correlation analysis model, the correlation between the analysis results of the model is analyzed, and the ability index is extracted, the ability index includes but is not limited to technical depth, logical integrity and practical adaptability; the ability index set is obtained by integrating the ability index. The correlation between the results can be quickly analyzed by the model, and the corresponding ability index is extracted, which provides a data basis for the ability analysis of the target personnel. The ability of the target personnel can be analyzed by combining the ability analysis result and the dynamic question and answer result, and the accuracy and comprehensiveness of the analysis result can be improved.

[0123] Specifically, according to the ability index set, the ability of the target personnel is evaluated by a preset ability evaluation model, and an ability evaluation report is generated; the ability evaluation model includes but is not limited to a multi-layer fully connected neural network model, a large number of ability index data are used to train the multi-layer fully connected neural network model, a pre-trained multi-layer fully connected neural network model is obtained, the ability index set is input into the pre-trained multi-layer fully connected neural network model, the model combines multiple ability indexes to calculate the comprehensive ability score of the target personnel, and according to the comprehensive ability score, a detailed ability evaluation report is generated. The ability indexes in the ability index set are integrated by the ability evaluation model, the ability is evaluated and analyzed, the accuracy and comprehensiveness of the ability analysis result can be improved, and the generated ability evaluation report can comprehensively reflect the ability of the target personnel.

[0124] As shown in Figure 4 Fig. 1, a personnel online auditing system based on a large model is used to implement a personnel online auditing method based on a large model, which includes:

[0125] The data analysis module extracts features from the pre-acquired multi-modal data of the target personnel, and constructs a multi-modal feature vector;

[0126] The ability analysis module analyzes the target ability and behavior intention of the target personnel by executing the sandbox and the trajectory analysis sandbox in the pre-constructed double-layer sandbox architecture based on the multi-modal feature vector, and obtains an ability analysis result;

[0127] The question and answer adjustment module configures a dynamic question and answer mechanism according to the ability analysis result, dynamically adjusts the question and answer content of the target personnel, and obtains a dynamic question and answer result.

[0128] The personnel auditing module combines the capability analysis result and the dynamic question and answer result, evaluates the capability of the target personnel through a preset capability evaluation model, and generates a capability evaluation report to perform online auditing on the personnel.

[0129] In this embodiment, the data analysis module analyzes and extracts features from the obtained multi-modal raw data of the target personnel, extracts features from multi-modal data such as text, code and audio through a feature extraction model, and fuses feature vectors of different modalities into a multi-modal feature vector. Through feature fusion, the limitations of single modal features can be avoided, and multi-source data complementary fusion can improve the comprehensiveness and accuracy of the description of the target personnel, thereby improving the accuracy and comprehensiveness of the personnel auditing process. The capability analysis module is based on the output of the data analysis module and analyzes in parallel in a double-layer sandbox architecture. The execution sandbox is responsible for verifying the actual running performance of the code in a secure isolation environment and evaluating the technical implementation capability of the personnel. The trajectory analysis sandbox analyzes the corresponding coding style and potential behavior intention by monitoring the internal logic path of the code execution process. Through the double-layer sandbox architecture, the capability and intention of the target personnel are verified twice, the analysis process of the comprehensive capability of the target personnel is deepened, and the comprehensiveness and accuracy of the personnel capability auditing process are improved.

[0130] Specifically, the question and answer adjustment module configures corresponding dynamic question and answer adjustment strategies according to the output of the capability analysis module, locates the scene switching point and difficulty adjustment point in the question and answer process, and generates new question and answer content matching the capability of the target personnel in real time in combination with the context. Through dynamic adjustment of the personnel examination process, the strengths and weaknesses of the target personnel can be further examined and verified, thereby improving the accuracy and efficiency of the personnel capability auditing process. The personnel auditing module receives the capability analysis result and the dynamic question and answer result, fuses, weights and calculates the capability auditing indicators through a preset trained capability evaluation model, and generates a structured capability evaluation report. By combining the capability analysis result and the dynamic question and answer result, a capability evaluation report reflecting the comprehensive capability of the target personnel can be obtained, thereby improving the rationality, fairness and reliability of the personnel auditing decision.

[0131] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solutions falling within the scope of the present application shall be considered as falling within the protection scope of the present application. It should be noted that, for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application shall also be considered as falling within the protection scope of the present application.

Claims

1. A large model-based online personnel auditing method, characterized in that, The method comprises the following steps: acquiring multi-modal data of a target person, wherein the multi-modal data comprises text data, code data and audio data; performing feature extraction on the text data, the code data and the audio data respectively through a preset feature extraction model to obtain a first feature set, a second feature set and a third feature set; combining the first feature set, the second feature set and the third feature set to construct a multi-modal feature vector; constructing a double-layer sandbox architecture comprising an execution sandbox and a trajectory analysis sandbox based on the multi-modal feature vector; inputting the code data in the multi-modal data into the execution sandbox, simulating running of the code and analyzing a target ability of the target person to obtain a first target ability; inputting the code data into the trajectory analysis sandbox, collecting a code execution process trajectory and analyzing a behavior intention of the target person to obtain a first behavior intention; performing correlation and comparison analysis on the first target ability and the first behavior intention to locate an ability advantage and an ability short board of the target person to obtain an ability analysis result; filtering out the ability advantage and the ability short board position according to the ability analysis result, and setting a scene switching point and a difficulty adjusting point at the corresponding positions; at the scene switching point, identifying scene switching content in the question and answer content, generating a replacement scene to replace the scene switching content, and obtaining a first adjusting result; at the difficulty adjusting point, analyzing a matching result of the question and answer content and the target person, generating a corresponding difficulty adjusting instruction to perform real-time difficulty adjustment, and obtaining a second adjusting result; performing dynamic adjustment on the question and answer content of the target person in combination with the first adjusting result and the second adjusting result to obtain a dynamic question and answer result; performing context correlation analysis on the ability analysis result and the dynamic question and answer result to extract a corresponding ability index set; performing ability evaluation on the target person through a preset ability evaluation model according to the ability index set to generate an ability evaluation report for online auditing of the person.

2. The personnel online auditing method based on a large model according to claim 1, characterized in that, The method comprises the following steps: inputting the code data in the multi-modal data into the execution sandbox, simulating running of the code, and collecting code running indexes during the running process; calculating time complexity and space complexity according to the code running indexes to analyze the target ability of the target person to obtain the first target ability.

3. The personnel online auditing method based on a large model according to claim 1, characterized in that, The method comprises the following steps: inputting the code data into the trajectory analysis sandbox to execute, collecting trajectory data in the code execution process; identifying coding style and logical correlation degree through a preset trajectory analysis model according to the trajectory data to construct a trajectory index vector; analyzing the behavior intention of the target person based on the trajectory index vector to obtain the first behavior intention.

4. The personnel online auditing method based on a large model according to claim 1, characterized in that, The method comprises the following steps: At the scene switching point, the scene switching content in the question and answer content is identified by analyzing the content corresponding to the capability advantages and capability short boards; The corresponding replacement scene is generated in combination with the context association of the scene switching content; The scene switching content is dynamically replaced according to the replacement scene, the corresponding scene content is reconstructed, and a first adjustment result is obtained.

5. The personnel online auditing method based on a large model according to claim 1, characterized in that, At the difficulty adjustment point, the capability matching result of the question and answer content and the target personnel is analyzed, and a corresponding difficulty adjustment instruction is generated to perform real-time difficulty adjustment, and a second adjustment result is obtained, including: At the difficulty adjustment point, the capability matching result of the question and answer content and the target personnel is analyzed, and a first matching result is obtained. Based on the first matching result, the corresponding difficulty adjustment instruction is matched in the pre-constructed difficulty instruction library. In response to the difficulty adjustment instruction, the question and answer content is real-time rewritten and adjusted in difficulty, and a second adjustment result is obtained.

6. A large model-based personnel online auditing system, characterized in that, A personnel online auditing method based on a large model is implemented, including: A data analysis module extracts features from pre-acquired multi-modal data of a target personnel, and constructs a multi-modal feature vector; A capability analysis module analyzes the target capability and behavior intention of the target personnel through an execution sandbox and a trajectory analysis sandbox in a pre-constructed double-layer sandbox architecture based on the multi-modal feature vector, and obtains a capability analysis result; A question and answer adjustment module configures a dynamic question and answer mechanism according to the capability analysis result, dynamically adjusts the question and answer content of the target personnel, and obtains a dynamic question and answer result; A personnel auditing module evaluates the capability of the target personnel through a pre-set capability evaluation model in combination with the capability analysis result and the dynamic question and answer result, generates a capability evaluation report, and performs online auditing on the personnel.

Images