An intelligent deep interrogation method and system

Abstract

This invention relates to an intelligent deep probing method and system, belonging to the field of natural language processing and intelligent assessment technology. The method includes: acquiring candidate responses and converting them into standardized text; parsing the text using multi-dimensional monitoring data to locate competency deficiencies and generate a set of missing identifiers; marking and ranking probing points; constructing a dynamic adaptive scoring matrix linking competencies and probing points; integrating monitoring data to calibrate scores and confidence levels; generating a probing intervention plan; pre-setting a virtual interviewer role library; dynamically matching roles and generating personalized probing content based on intervention logic and the set of missing identifiers; when core meta-competency elements are insufficiently represented or confidence levels are not met, generating targeted metacognitive probing questions; and outputting multi-dimensional competency scores and confidence reports after element extraction and fusion. This invention upgrades probing from mechanical questioning to intelligent technical intervention, improving the targeting and accuracy of probing, and providing reliable support for precise talent selection.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing and intelligent evaluation technology, specifically relating to an intelligent deep questioning method and system. Background Technology

[0002] In talent selection, job competition, and skills assessment, in-depth follow-up questioning is a core element in uncovering candidates' true abilities and accurately assessing their competence; its quality directly determines the reliability and effectiveness of the assessment results. With the application of intelligent technologies in the human resources field, traditional manual follow-up questioning is gradually being replaced by intelligent follow-up questioning systems. However, existing technologies still have many key shortcomings, making it difficult to meet the needs of precise and intelligent assessment.

[0003] Existing intelligent follow-up questioning technologies largely rely on pre-set question databases for mechanical pushes, triggering fixed follow-up questions based only on the candidate's surface-level answers. They lack the ability to deeply analyze and dynamically adapt the answer information. In terms of follow-up questioning logic, there is no clear intermediate technical state support, making it impossible to accurately locate the missing elements of core competencies. This results in scattered and untargeted follow-up questions, easily bypassing key assessment points and making it difficult to achieve in-depth exploration of the candidate's abilities.

[0004] Meanwhile, existing technologies equate follow-up questions with mere verbal inquiries, failing to elevate them to the level of technological intervention. They cannot adjust follow-up strategies based on the candidate's real-time state (such as emotional fluctuations, completeness of responses, and logical coherence) and changes in the environment (such as interfering factors and data transmission status). This rigid follow-up questioning model not only makes it difficult to guide candidates to fully express themselves but may also lead to evaluation bias due to inappropriate timing and methods of follow-up questions, reducing the objectivity of the evaluation results.

[0005] In terms of assessing metacognitive abilities, existing systems lack standardized technical means of implementation. For in-depth questioning types such as process reflection, hypothesis subversion, self-assessment of knowledge boundaries, and value weighing, they rely heavily on manual design and lack systematic question generation logic and technical support. This results in highly subjective and inconsistent questioning content, making it impossible to form a comprehensive and accurate assessment of the candidate's metacognitive abilities.

[0006] Furthermore, existing intelligent probing systems fail to fully integrate multi-dimensional monitoring data, and cannot utilize key data such as response completeness, semantic relevance, and environmental interference levels to optimize the probing process. This results in a lack of targeted probing, failing to adapt to individual candidate differences or handle dynamic changes in scenarios, further exacerbating biases in assessment results and failing to meet the actual needs of enterprises for precise selection of high-caliber talent. In summary, existing intelligent probing technologies suffer from insufficient targeting, a lack of technical intervention logic, a lack of standardized technical support for metacognitive probing, and ineffective integration of monitoring data. Summary of the Invention

[0007] To address the aforementioned problems in the existing technology, this invention provides an intelligent deep questioning method and system. The objective of this invention can be achieved through the following technical solutions:

[0008] include:

[0009] S1: Acquire multi-dimensional monitoring status data and candidates' voice or text responses to the initial questions, and convert the voice responses into standardized text; perform targeted information completeness analysis on the standardized text; construct a result element identification model and generate a competency element missing identifier set;

[0010] S2: Construct a dynamic adaptive scoring matrix that links competence and follow-up questions, integrate the multi-dimensional monitoring status data to calibrate the matrix score and confidence level, and preset the fluctuation threshold of each scoring dimension; based on the matrix features and the competence element missing identifier set, generate a follow-up questioning technology intervention scheme.

[0011] S3: Pre-set a differentiated virtual interviewer role library, with each role associated with technical intervention logic; combine the candidate's competency shortcomings from the previous round of responses, multi-dimensional monitoring status data, and follow-up questioning technical intervention schemes to dynamically match virtual roles; through targeted matching of the technical intervention logic corresponding to the role with the competency element missing identifier set, form personalized follow-up question content;

[0012] S4: When the candidate's core meta-competency elements are not sufficiently represented or the score confidence level does not reach the preset threshold, the final output is a targeted metacognitive follow-up question through element extraction and correlation fusion.

[0013] Specifically, the process of performing targeted information completeness analysis on standardized text is as follows: preset core analysis dimensions and corresponding element templates; extract information fragments from each dimension from the standardized text, and align the extraction results with the field requirements of the element templates one by one; locate the uncovered core elements and ambiguous content fragments through field coverage statistics and semantic consistency verification.

[0014] Specifically, the process of constructing the result element recognition model is as follows:

[0015] It is divided into a data preprocessing layer, a feature extraction layer, and an element matching layer;

[0016] Collect competency data from multiple industry positions to construct a training set;

[0017] The preprocessing layer performs noise reduction, word segmentation, and part-of-speech tagging on the parsed text; the feature extraction layer generates fixed-dimensional semantic feature vectors; and the element matching layer compares the feature vectors with the competency element feature library using a semantic similarity calculation method, sets a matching threshold, and establishes rules for determining missing elements.

[0018] Specifically, the process of generating the competency element missing identifier set is as follows: based on the adjusted parsing weights, standard text segmentation, entity recognition, and semantic association analysis are performed to extract effective information fragments related to competency elements; the information is judged against the preset information completeness judgment criteria; element items that do not match effective information or whose information does not meet the standards are marked as missing elements; the missing elements are summarized and sorted according to their degree of influence on job competency to form the competency element missing identifier set.

[0019] Specifically, the process of constructing the dynamic adaptive scoring matrix that links competence with follow-up questions is as follows:

[0020] A matrix framework is constructed with core competency dimensions as rows and detailed assessment indicators as columns.

[0021] The initial score is calculated based on element completeness and semantic fit.

[0022] Establish a mapping relationship between follow-up questions and matrix cells. For each follow-up question with valid information, automatically update the corresponding cell score and dimension total score. Integrate multi-dimensional monitoring status data to calibrate score deviations and simultaneously calculate confidence values.

[0023] Specifically, the process of generating follow-up questioning intervention scheme is as follows: the core of the preset intervention is to supplement key missing elements and correct score deviations, which is divided into three types of intervention methods: targeted follow-up questioning, logical guidance, and rhythm adaptation; for targeted follow-up questioning, set question generation rules and upper limit of follow-up questioning rounds, and preset questioning interval and speech speed adjustment parameters.

[0024] Specifically, the process of the preset differentiated virtual interviewer role library is as follows: design role types according to the needs of intervention scenarios, write language style rules, questioning logic chains and intervention trigger thresholds for each role; preset the competency dimensions of each role and associate them with corresponding follow-up questioning strategy templates; and store the role attributes, triggering conditions and strategy templates in a structured manner to establish a role index library.

[0025] Specifically, the process of dynamically matching virtual roles is as follows: extracting the candidate's historical response competency deficit vector, emotional state feature value, and current intervention target parameters, and integrating the three to generate a comprehensive demand vector; calculating the fit between the comprehensive demand vector and the adaptation feature vector of each role using a vector distance calculation method, and selecting roles with a fit within a preset range as candidate roles.

[0026] Specifically, the process of targeted matching between the role-corresponding technical intervention logic and the competency element missing identifier set is as follows:

[0027] Deconstruct the questioning direction, rhetoric style, and in-depth mining rules in the role's technical intervention logic, and extract the missing type, related dimensions, and priority information from the competency element missing marker set;

[0028] Establish a core element mapping table, match the corresponding follow-up questions according to the missing type, determine the depth of exploration according to priority, and adapt the dialogue style according to the relevance dimension to form a logical chain of follow-up questions.

[0029] Specifically, the process of element extraction and association fusion is as follows: core meta-competency elements and related arguments are extracted from candidate responses using keyword weight calculation methods and semantic role labeling technology; the extraction results are cross-dimensionally correlated and verified with previous response data and competency element missing identifier sets; weights are assigned according to element importance and data credibility, and multi-source data are integrated into a unified feature vector through feature integration technology.

[0030] Specifically, the process of outputting targeted metacognitive follow-up questions is as follows: the metacognitive follow-up question type is determined by a preset type mapping rule; a question generation framework containing sentence structure and semantic expansion rules is constructed for each type; after inputting candidate response fragments, the core expression logic and key information nodes are extracted through semantic analysis, and the missing element information, the rhetorical style parameters in the role intervention logic and key nodes are matched and integrated, and the sentence structure is constructed and semantic expansion is completed according to the framework rules; finally, after grammatical rule verification and logical consistency detection, metacognitive follow-up questions targeting missing elements are output.

[0031] Specifically, an intelligent deep questioning system includes:

[0032] Information parsing and missing information identification module: acquires the candidate's voice or text answer to the initial question, converts the voice answer into standardized text; performs targeted information completeness parsing on the standardized text; constructs a result element identification model, and generates a set of competency element missing information identifiers;

[0033] The scoring matrix construction and intervention scheme module constructs a dynamic adaptive scoring matrix that links competence and follow-up questions, integrates the multi-dimensional monitoring status data to calibrate the matrix score and confidence level, and presets the fluctuation threshold for each scoring dimension; based on the matrix features and the competence element missing identifier set, it generates a follow-up questioning technology intervention scheme.

[0034] Role matching and follow-up question generation module: Pre-set differentiated virtual interviewer role library, each role is associated with specific technical intervention logic; Combine the candidate’s competency shortcomings in the previous round of responses, multi-dimensional monitoring status data and follow-up question technical intervention scheme, dynamically match virtual roles; Targeted matching between the technical intervention logic corresponding to the role and the competency element missing identifier set is used to form personalized follow-up question content;

[0035] Targeted metacognitive question output module: When the candidate's core meta-competency elements are not sufficiently represented or the score confidence level does not reach the preset threshold, the module extracts and integrates elements to finally output targeted metacognitive follow-up questions.

[0036] The beneficial effects of this invention are as follows:

[0037] (1) By setting up a set of missing competency elements, a dynamic adaptive scoring matrix and a multi-dimensional monitoring data fusion mechanism, the missing competency elements are first accurately located through the directional information completeness analysis and result element identification model. Then, the follow-up questions are sorted by weight calculation. At the same time, the follow-up questions and evaluation scores are linked and calibrated by relying on the scoring matrix. This provides an indispensable intermediate technical support for follow-up questions, avoids blind follow-up questions, and improves the accuracy and confidence of competency assessment through multi-dimensional data calibration.

[0038] (2) By setting up a differentiated virtual interviewer role library, a targeted metacognitive probing question generation logic and a technical intervention scheme, roles are dynamically matched based on the candidate's competency shortcomings and real-time status. Characteristic probing questions are generated by targeted matching with role intervention logic and missing identifier set. At the same time, the systematic generation of multiple types of metacognitive probing questions is achieved through standardized technical means. This not only upgrades probing questions from simple language behavior to intelligent technical intervention, but also deeply explores the candidate's meta-ability and outputs comprehensive and credible evaluation results to meet the needs of precise talent selection. Attached Figure Description

[0039] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0040] Figure 1 This is a flowchart illustrating an intelligent deep questioning method according to the present invention.

[0041] Figure 2 This is a data flow diagram of an intelligent deep interrogation method according to the present invention. Detailed Implementation

[0042] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0043] Please see Figure 1-2 A method for intelligent, in-depth questioning;

[0044] include:

[0045] S1: Acquire multi-dimensional monitoring status data and candidates' voice or text responses to the initial questions, and convert the voice responses into standardized text; perform targeted information completeness analysis on the standardized text; construct a result element identification model and generate a competency element missing identifier set;

[0046] S2: Construct a dynamic adaptive scoring matrix that links competence and follow-up questions, integrate the multi-dimensional monitoring status data to calibrate the matrix score and confidence level, and preset the fluctuation threshold of each scoring dimension; based on the matrix features and the competence element missing identifier set, generate a follow-up questioning technology intervention scheme.

[0047] S3: Pre-set a differentiated virtual interviewer role library, with each role associated with technical intervention logic; combine the candidate's competency shortcomings from the previous round of responses, multi-dimensional monitoring status data, and follow-up questioning technical intervention schemes to dynamically match virtual roles; through targeted matching of the technical intervention logic corresponding to the role with the competency element missing identifier set, form personalized follow-up question content;

[0048] S4: When the candidate's core meta-competency elements are not sufficiently represented or the score confidence level does not reach the preset threshold, the final output is a targeted metacognitive follow-up question through element extraction and correlation fusion.

[0049] In this embodiment, the multi-dimensional monitoring status data refers to a real-time data set covering the three core dimensions of "response-environment-operation" to support the accuracy of follow-up questions and the objectivity of evaluation. When using the multi-dimensional monitoring status data, it can dynamically adapt to changes in candidate status and scenario, improving the accuracy of follow-up questions, the accuracy of score calibration, and the adaptability of intervention. Without using it, it is impossible to perceive fluctuations in response status, environment, and operation status in real time, resulting in a lack of focus in follow-up questions and an expansion of evaluation bias. Therefore, the monitoring status data is forcibly integrated to form a closed loop through technical design. Specifically, it includes candidate response completeness, semantic fit, and emotional fluctuation parameters in the response status dimension; interference type and interference intensity quantification data in the follow-up questioning interaction scenario in the environment status dimension; and system data transmission latency and program failure frequency in the operation status dimension. This type of data is mainly used to adjust text parsing weights, calibrate the confidence of the score matrix, dynamically match virtual roles, and optimize follow-up questioning strategies.

[0050] In this embodiment, the specific technical intervention logic refers to a set of structured rules bound to the differentiated virtual interviewer role and used to guide the execution of precise follow-up questions. Specifically, it includes targeted follow-up questioning direction rules for different types of competency deficiencies, quantifiable execution parameter rules such as questioning interval duration and follow-up questioning round limit, role-specific expression logic and rhetoric style rules, and intervention initiation and adjustment trigger adaptation rules based on data such as candidate emotional state and response quality. Its core function is to transform follow-up questions from simple language behavior into implementable technical intervention actions, ensuring that follow-up questions are aligned with the deficient elements and the candidate's state.

[0051] In this embodiment, the core analytical dimensions refer to the core competency dimensions determined based on the job competency model and used for targeted analysis of candidate response texts. Specifically, these include hard skill dimensions that are strongly related to job requirements (such as the degree of mastery of professional knowledge, proficiency in practical skills, and ability to use tools), soft skill dimensions (such as communication and expression skills, logical thinking skills, problem-solving skills, and teamwork awareness), and metacognitive dimensions (such as the ability to reflect on processes, the ability to verify hypotheses, the awareness of knowledge boundaries, and the ability to weigh values). These dimensions are the core basis for the completeness analysis of text information and are used to decompose element templates and locate missing information.

[0052] In this embodiment, the core expression logic and key information nodes are the core content extracted from the candidate's response fragments, which together support the targeted generation of follow-up questions. The core expression logic refers to the supporting relationships, reasoning chains, stances and argumentation logic in the candidate's response, while the key information nodes refer to the keywords, core arguments, key evidence, and ambiguous statements that are not fully explained in the response that carry the core viewpoints. Both are extracted through semantic analysis technology and used to adapt to missing element information to ensure that the generated follow-up questions can accurately connect with the core context of the candidate's response.

[0053] In this embodiment, the core meta-competency elements and related arguments refer to the core content used to evaluate the candidate's deep-level abilities and are the key basis for competency assessment. The core meta-competency elements focus on the metacognitive level and specifically include process reflection ability, hypothesis overturning ability, knowledge boundary self-assessment ability, and value weighing ability. Related arguments refer to the specific content in the candidate's response that supports the above meta-competency elements, such as self-deficiencies mentioned during reflection, reasoning given when overturning hypotheses, specific explanations when defining knowledge boundaries, and judgment basis when weighing values. After extraction and correlation fusion, this type of content is used to update the competency score matrix and support the final evaluation result output.

[0054] In this embodiment, the technical function of adding the virtual interviewer role is specifically as follows: Adapting to differentiated cognitive characteristics: For candidates with different personalities and answering styles, matching virtual roles with corresponding communication traits reduces candidate psychological resistance, improves answering cooperation, and ensures the effectiveness of obtaining follow-up information; Achieving precise and targeted follow-up questioning: The follow-up questioning strategies bound to each role are deeply related to the job competency elements, allowing for targeted efforts based on the type of missing elements, avoiding generalized follow-up questioning, and improving the efficiency of filling in missing information; Ensuring the logical coherence of follow-up questioning: The virtual role has a preset communication logic chain, which can dynamically adjust the follow-up questioning pace based on the candidate's previous round of response feedback, forming a closed loop of "identifying gaps - targeted questioning - feedback follow-up," avoiding gaps in follow-up questioning; Its irreplaceable nature lies in: The virtual role "adapts to communication traits." Deeply integrated with the "competency-targeted follow-up questioning strategy," a dedicated technical chain is formed. This integration mechanism cannot be replaced by a single follow-up question generation model. A single model can only output questions and cannot take into account the candidate's communication adaptability and the coherence of the follow-up questioning logic. If the virtual role is removed and follow-up questions are generated solely through a uniform template, the lack of adaptability to different candidate characteristics will lead to low follow-up cooperation and insufficient completion of missing information, thereby affecting the effectiveness of subsequent in-depth follow-up questions and failing to achieve the "precise in-depth follow-up questioning" goal set by the solution. The virtual role matching mechanism is the core intermediate link connecting the "follow-up questioning technology intervention solution" and "personalized follow-up questioning content generation." Its functions of "strategy adaptation-communication coordination-logical connection" are necessary guarantees for the smooth operation of the entire intelligent in-depth follow-up questioning process and cannot be replaced by other technical modules.

[0055] Specifically, the process of performing targeted information completeness analysis on standardized text is as follows: preset core analysis dimensions and corresponding element templates; extract information fragments from each dimension from the standardized text, and align the extraction results with the field requirements of the element templates one by one; locate the uncovered core elements and ambiguous content fragments through field coverage statistics and semantic consistency verification.

[0056] Specifically, the process of constructing the result element recognition model is as follows:

[0057] It is divided into a data preprocessing layer, a feature extraction layer, and an element matching layer;

[0058] Collect competency data from multiple industry positions to construct a training set;

[0059] The preprocessing layer performs noise reduction, word segmentation, and part-of-speech tagging on the parsed text; the feature extraction layer generates fixed-dimensional semantic feature vectors; and the element matching layer compares the feature vectors with the competency element feature library using a semantic similarity calculation method, sets a matching threshold, and establishes rules for determining missing elements.

[0060] Specifically, the process of marking follow-up questions and prioritizing them according to their impact is as follows:

[0061] Points of missing information, semantic ambiguity, and logical contradiction are marked as inquiry points, and each inquiry point is associated with the corresponding competency dimension and evaluation indicator.

[0062] Three weighted indicators are set: relevance, supplementary value, and job fit. The weight ratio of each indicator is preset. The comprehensive score of each follow-up question is calculated and the follow-up questions are arranged in descending order of the scores to form a follow-up question sequence.

[0063] At the same time, a dynamic adjustment interface is reserved, which can adjust the sorting according to the real-time response situation.

[0064] Specifically, the process of constructing the dynamic adaptive scoring matrix that links competence with follow-up questions is as follows:

[0065] A matrix framework is constructed with core competency dimensions as rows and detailed assessment indicators as columns.

[0066] The initial score is calculated based on element completeness and semantic fit.

[0067] Establish a mapping relationship between follow-up questions and matrix cells. For each follow-up question with valid information, automatically update the corresponding cell score and dimension total score. Integrate multi-dimensional monitoring status data to calibrate score deviations and simultaneously calculate confidence values.

[0068] Specifically, the process of generating follow-up questioning intervention scheme is as follows: the core of the preset intervention is to supplement key missing elements and correct score deviations, which is divided into three types of intervention methods: targeted follow-up questioning, logical guidance, and rhythm adaptation; for targeted follow-up questioning, set question generation rules and upper limit of follow-up questioning rounds, and preset questioning interval and speech speed adjustment parameters.

[0069] Specifically, the process of the preset differentiated virtual interviewer role library is as follows: design role types according to the needs of intervention scenarios, write language style rules, questioning logic chains and intervention trigger thresholds for each role; preset the competency dimensions of each role and associate them with corresponding follow-up questioning strategy templates; and store the role attributes, triggering conditions and strategy templates in a structured manner to establish a role index library.

[0070] Specifically, the process of dynamically matching virtual roles is as follows: extracting the candidate's historical response competency deficit vector, emotional state feature value, and current intervention target parameters, and integrating the three to generate a comprehensive demand vector; calculating the fit between the comprehensive demand vector and the adaptation feature vector of each role using a vector distance calculation method, and selecting roles with a fit within a preset range as candidate roles.

[0071] Specifically, the process of targeted matching between the role-corresponding technical intervention logic and the competency element missing identifier set is as follows:

[0072] Deconstruct the questioning direction, rhetoric style, and in-depth mining rules in the role's technical intervention logic, and extract the missing type, related dimensions, and priority information from the competency element missing marker set;

[0073] Establish a core element mapping table, match the corresponding follow-up questions according to the missing type, determine the depth of exploration according to priority, and adapt the dialogue style according to the relevance dimension to form a logical chain of follow-up questions.

[0074] Specifically, the process of element extraction and association fusion is as follows: core meta-competency elements and related arguments are extracted from candidate responses using keyword weight calculation methods and semantic role labeling technology; the extraction results are cross-dimensionally correlated and verified with previous response data and competency element missing identifier sets; weights are assigned according to element importance and data credibility, and multi-source data are integrated into a unified feature vector through feature integration technology.

[0075] Specifically, the process of outputting targeted metacognitive follow-up questions is as follows: the metacognitive follow-up question type is determined by a preset type mapping rule; a question generation framework containing sentence structure and semantic expansion rules is constructed for each type; after inputting candidate response fragments, the core expression logic and key information nodes are extracted through semantic analysis, and the missing element information, the rhetorical style parameters in the role intervention logic and key nodes are matched and integrated, and the sentence structure is constructed and semantic expansion is completed according to the framework rules; finally, after grammatical rule verification and logical consistency detection, metacognitive follow-up questions targeting missing elements are output.

[0076] In this embodiment, the specific process of constructing the result element recognition model based on the parsing results is as follows: The model is divided into a data preprocessing layer, a feature extraction layer, and an element matching layer. First, competency element data of multiple industry positions (covering core job types such as technical positions and management positions) are collected to construct a labeled training set; the preprocessing layer uses a word segmentation tool to segment the parsed text, and combines it with a part-of-speech tagging tool to complete part-of-speech tagging. At the same time, special characters and redundant interjections are removed using regular expressions to achieve text denoising; the feature extraction layer uses a fine-tuned pre-trained semantic understanding model to convert the processed text into a semantic feature vector of fixed dimension (such as a dimension); the element matching layer compares the feature vector with the competency element feature library through a semantic similarity calculation method. The matching threshold is calibrated according to the industry average recognition accuracy (usually set to x), and a rule of "if it is below the threshold, it is judged as missing" is established. The model parameters are optimized through more than n rounds of cross-validation to ensure that the accuracy of missing element recognition is not lower than y.

[0077] In this embodiment, the specific process of dynamically matching virtual roles is as follows: Emotional state feature values ​​(such as tension and confidence metrics) of candidates' historical responses are extracted using a voice emotion recognition model. These features are then combined with competency gap vectors (generated from the low-dimension partition of the scoring matrix) and current intervention target parameters (such as "supplementing professional skill deficiencies" and "calibrating logical thinking scores"). A weighted summation method (weights set according to job requirements, with emotional state accounting for p%, gap vectors accounting for q%, and intervention targets accounting for r%) is used to generate a comprehensive requirement vector. The fit between this vector and the matching feature vectors of each role in the role library is calculated using a vector distance calculation method. A preset fit threshold of m is used to select roles with a fit higher than the threshold as candidate roles. If no suitable role is found, the core strategies of k roles are fused according to the intervention target weights (such as fusion of guiding conversational styles and in-depth questioning logic) to generate a customized virtual role.

[0078] In this embodiment, the specific process of generating various targeted metacognitive follow-up questions is as follows: Preset type mapping rules (e.g., "logical contradiction missing" corresponds to process reflection follow-up questions, "fuzzy professional knowledge boundary" corresponds to knowledge boundary self-assessment follow-up questions); determine metacognitive follow-up question types based on competency element missing identifier sets; construct a question generation framework for each type containing basic sentence structures and semantic expansion rules (according to the hierarchical expansion of "core question + missing element guidance + logical extension"); after inputting candidate response fragments, extract core expression logic (e.g., the logical chain of "problem-cause-solution") and key information nodes (e.g., unclear causes, contradictory solutions) through dependency parsing technology; adapt and integrate missing element information, rhetorical style parameters in role intervention logic (e.g., using gentle sentences for guidance, and rhetorical questions for challenge) with key nodes through semantic splicing technology, completing sentence construction and semantic expansion according to the framework rules; finally, use regular expressions for syntax verification, detect the internal logical fluency of the questions through a logical reasoning engine, eliminate generated results with semantic conflicts and ambiguous expressions, and output metacognitive follow-up questions targeting missing elements.

[0079] Specifically, an intelligent deep questioning system includes:

[0080] Information parsing and missing information identification module: acquires the candidate's voice or text answer to the initial question, converts the voice answer into standardized text; performs targeted information completeness parsing on the standardized text; constructs a result element identification model, and generates a set of competency element missing information identifiers;

[0081] The scoring matrix construction and intervention scheme module constructs a dynamic adaptive scoring matrix that links competence and follow-up questions, integrates the multi-dimensional monitoring status data to calibrate the matrix score and confidence level, and presets the fluctuation threshold for each scoring dimension; based on the matrix features and the competence element missing identifier set, it generates a follow-up questioning technology intervention scheme.

[0082] Role matching and follow-up question generation module: Pre-set differentiated virtual interviewer role library, each role is associated with specific technical intervention logic; Combine the candidate’s competency shortcomings in the previous round of responses, multi-dimensional monitoring status data and follow-up question technical intervention scheme, dynamically match virtual roles; Targeted matching between the technical intervention logic corresponding to the role and the competency element missing identifier set is used to form personalized follow-up question content;

[0083] Targeted metacognitive question output module: When the candidate's core meta-competency elements are not sufficiently represented or the score confidence level does not reach the preset threshold, the module extracts and integrates elements to finally output targeted metacognitive follow-up questions.

[0084] This example uses a smart interview for a software engineer position as a scenario. A candidate is applying for a mid-level backend developer position. The specific process is as follows:

[0085] Information Analysis and Missing Item Identification (corresponding to S1): The candidate's voice response to the initial question "Describe a project development experience you led" is converted into standardized text. Simultaneously, multi-dimensional monitoring data on response status (completeness, semantic fit, etc.), environmental status (noise interference), and operational status (transmission stability) are acquired. Analysis weights are adjusted based on this data. The text is analyzed according to core dimensions such as hard skills (programming, project management) and soft skills (problem solving). It is found that the candidate only describes the project context (S) and task (T), lacking core actions (A) and results (R), and contains vague expressions such as "we developed quickly." A result element identification model is constructed. After text preprocessing, feature extraction, and element comparison, "project execution actions" and "quantification of results" are identified as missing elements. A competency element missing item identifier set is generated, labeling the missing type, related dimensions, and priority. Simultaneously, potential contradictions between the candidate's technology choices and resume are identified. Information gaps, ambiguities, and contradictions are marked as follow-up questions. A comprehensive score is calculated using weighted indicators and ranked to form an ordered follow-up question sequence.

[0086] Score Matrix Construction and Intervention Plan Generation (corresponding to S2): A dynamic adaptive score matrix is ​​constructed using the core competency dimensions of software engineers (programming, problem solving, etc.) as rows and subdivided evaluation indicators as columns. The initial score is calculated based on the analysis results of S1, where the score and confidence level of "programming ability" are low due to missing information. The score and confidence level are calibrated by integrating multi-dimensional monitoring data (high candidate confidence, low environmental interference), and the fluctuation thresholds of each dimension are preset. Combined with the priority of the missing identifier set, a follow-up questioning intervention plan is generated. Targeted follow-up questions are executed for "programming ability," and stress test intervention is designed for "team collaboration ability," setting parameters such as questioning interval and maximum number of rounds.

[0087] Role matching and trait-based follow-up question generation (corresponding to S3): A pre-defined library of differentiated roles, such as "Conservative Technical Expert" and "Guiding Colleague," is established. Each role is associated with specific intervention logic (e.g., "Conservative Technical Expert" focuses on in-depth technical detail analysis). The candidate's competency gaps (lack of programming skills), emotional state (high confidence), and intervention goals are extracted and integrated to generate a comprehensive needs vector. Based on vector fit calculation, the "Conservative Technical Expert" role is matched. The intervention logic for this role is broken down and targeted with the missing identifier set to generate trait-based follow-up questions: "Are you responsible for coding the core modules of the project? What are the core advantages of the technical framework you use, and how do you solve high-concurrency performance issues?"

[0088] Targeted Metacognitive Questioning and Evaluation Results Output (corresponding to S4): After the candidate responds, it is found that their description of "performance optimization" is vague, the core meta-competency elements are not sufficiently represented, and the confidence score does not reach the threshold, triggering metacognitive questioning; based on the missing identifier set, the process reflection type and hypothesis subversion type of questioning are determined, the core logic and key nodes of the candidate's response are extracted, and the role's verbal parameters are integrated to generate the following question: "Looking back at the performance optimization stage of the project, what was the most challenging technical difficulty? If there is a problem with the adaptation of the core framework, how would you refactor the solution?" After the candidate responds, the core meta-competency elements and related arguments are extracted, verified and integrated with the previous data and missing identifier set, the score matrix is ​​updated, and finally, a multi-dimensional competency score and confidence report is output to support recruitment decisions.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An intelligent deep questioning method, characterized in that, include: S1: Acquire multi-dimensional monitoring status data and candidates' voice or text responses to the initial question, and convert the voice responses into standardized text; The multi-dimensional monitoring status data includes candidate response completeness, semantic fit, and emotional fluctuation parameters in the response status dimension; interference type and interference intensity quantification data in the follow-up questioning interaction scenario in the environmental status dimension; and system data transmission latency and program failure frequency in the running status dimension. Perform targeted information completeness analysis on the standardized text; construct a result element identification model to generate a competency element missing identifier set; The specific process of performing targeted information completeness analysis is as follows: preset core analysis dimensions and corresponding element templates, wherein the core analysis dimensions are used to target the core capability dimensions of the candidate's response text; extract information fragments under each dimension from the standardized text, and align the extraction results with the field requirements of the element template one by one; locate the uncovered core elements and ambiguous content fragments through field coverage statistics and semantic consistency verification. The specific process of constructing the result element recognition model is as follows: it is divided into a data preprocessing layer, a feature extraction layer, and an element matching layer; collect competency element data of multiple industry positions to construct a training set; the preprocessing layer performs noise reduction, word segmentation, and part-of-speech tagging on the parsed text; the feature extraction layer generates fixed-dimensional semantic feature vectors; the element matching layer compares the feature vectors with the competency element feature library through a semantic similarity calculation method, and establishes a missing element judgment rule by setting a matching threshold. The specific process for generating the competency element missing identifier set is as follows: based on the adjusted parsing weights, perform standardized text segmentation, entity recognition, and semantic association analysis to extract effective information fragments related to competency elements; judge the information against the preset information completeness judgment criteria; and mark the element items that do not match effective information or whose information does not meet the criteria as missing elements. Summarize the missing elements and rank them according to their impact on job competency to form a set of missing competency element identifiers; S2: Construct a dynamic adaptive scoring matrix that links competence and follow-up questions, integrate the multi-dimensional monitoring status data to calibrate the matrix score and confidence level, and preset the fluctuation threshold of each scoring dimension; based on the matrix features and the competence element missing identifier set, generate a follow-up questioning technology intervention scheme. The specific process of constructing the dynamic adaptive scoring matrix that links competence and follow-up questions is as follows: build a matrix framework with the core competence dimensions as rows and the subdivided evaluation indicators as columns; calculate the initial score based on element completeness and semantic fit. Establish a mapping relationship between follow-up questions and matrix cells. For each follow-up question with valid information, automatically update the corresponding cell score and dimension total score. Integrate multi-dimensional monitoring status data to calibrate score deviations and simultaneously calculate confidence values. S3: Pre-set a differentiated virtual interviewer role library, with each role associated with technical intervention logic; combine the candidate's competency shortcomings from the previous round of responses, multi-dimensional monitoring status data, and follow-up questioning technical intervention schemes to dynamically match virtual roles; through targeted matching of the technical intervention logic corresponding to the role with the competency element missing identifier set, form personalized follow-up question content; The specific process of targeted matching between the technical intervention logic corresponding to the role and the competency element missing identifier set is as follows: deconstruct the questioning direction, rhetoric style, and in-depth mining rules in the role's technical intervention logic, and extract the missing type, correlation dimension, and priority information from the competency element missing identifier set; Establish a core element mapping table, match the corresponding follow-up questions according to the missing type, determine the depth of exploration according to priority, and adapt the dialogue style according to the relevance dimension to form a logical chain of follow-up questions. S4: When the candidate's core meta-competency elements are not sufficiently represented or the score confidence level does not reach the preset threshold, the final output is a targeted metacognitive follow-up question through element extraction and association fusion. The specific process of element extraction and association fusion is as follows: core meta-competency elements and related arguments are extracted from candidate responses using keyword weight calculation methods and semantic role labeling technology; the core meta-competency elements focus on the metacognitive level, and the related arguments refer to the specific content in the candidate responses that supports the above meta-competency elements; the extraction results are cross-dimensionally correlated and verified with previous response data and competency element missing identifier sets; weights are assigned according to element importance and data credibility, and multi-source data are integrated into a unified feature vector through feature integration technology; The specific process of outputting targeted metacognitive follow-up questions is as follows: Based on a unified feature vector, the metacognitive follow-up question type is determined through a preset type mapping rule; a question generation framework containing sentence structure and semantic expansion rules is constructed for each type; after inputting candidate response fragments, the core expression logic and key information nodes are extracted through semantic analysis, and the missing element information, the rhetorical style parameters in the role intervention logic, and the key nodes are matched and integrated, and the sentence structure is constructed and the semantic expansion is completed according to the framework rules; finally, after grammatical rule verification and logical consistency detection, the metacognitive follow-up questions targeting the missing elements are output.

2. The method according to claim 1, characterized in that, The specific process of the proposed follow-up questioning intervention scheme is as follows: the core of the intervention is to supplement key missing elements and correct score deviations, and it is divided into three types of intervention methods: targeted follow-up questioning, logical guidance, and rhythm adaptation; for targeted follow-up questioning, the question generation rules and the upper limit of the number of follow-up questioning rounds are set, and the questioning interval and speech speed adjustment parameters are preset.

3. The method according to claim 1, characterized in that, The specific process of the preset differentiated virtual interviewer role library is as follows: design role types according to the needs of intervention scenarios, write language style rules, questioning logic chains and intervention trigger thresholds for each role; preset the competency dimensions of each role and associate them with corresponding follow-up questioning strategies; and structurally store role attributes, triggering conditions and strategy templates to establish a role index library.

4. The method according to claim 1, characterized in that, The specific process of dynamically matching virtual roles is as follows: extract the competency gap vector, emotional state feature value and current intervention target parameter of the candidate's historical response, and integrate the three to generate a comprehensive demand vector; calculate the fit degree between the comprehensive demand vector and the adaptation feature vector of each role through the vector distance calculation method, and select roles with a fit degree within a preset range as candidate roles.

5. An intelligent deep questioning system, characterized in that, include: Information parsing and missing data identification module: acquires multi-dimensional monitoring status data and candidates' voice or text answers to the initial question, and converts the voice answers into standardized text; The multi-dimensional monitoring status data includes candidate response completeness, semantic fit, and emotional fluctuation parameters in the response status dimension; interference type and interference intensity quantification data in the follow-up questioning interaction scenario in the environmental status dimension; and system data transmission latency and program failure frequency in the running status dimension. Perform targeted information completeness analysis on the standardized text; construct a result element identification model to generate a competency element missing identifier set; The specific process of performing targeted information completeness analysis is as follows: preset core analysis dimensions and corresponding element templates, wherein the core analysis dimensions are used to target the core capability dimensions of the candidate's response text; extract information fragments under each dimension from the standardized text, and align the extraction results with the field requirements of the element template one by one; locate the uncovered core elements and ambiguous content fragments through field coverage statistics and semantic consistency verification. The specific process of constructing the result element recognition model is as follows: it is divided into a data preprocessing layer, a feature extraction layer, and an element matching layer; collect competency element data of multiple industry positions to construct a training set; the preprocessing layer performs noise reduction, word segmentation, and part-of-speech tagging on the parsed text; the feature extraction layer generates fixed-dimensional semantic feature vectors; the element matching layer compares the feature vectors with the competency element feature library through a semantic similarity calculation method, and establishes a missing element judgment rule by setting a matching threshold. The specific process for generating the competency element missing identifier set is as follows: based on the adjusted parsing weights, perform standardized text segmentation, entity recognition, and semantic association analysis to extract effective information fragments related to competency elements; judge the information against the preset information completeness judgment criteria; and mark the element items that do not match effective information or whose information does not meet the criteria as missing elements. Summarize the missing elements and rank them according to their impact on job competency to form a set of missing competency element identifiers; The scoring matrix construction and intervention scheme module constructs a dynamic adaptive scoring matrix that links competence and follow-up questions, integrates the multi-dimensional monitoring status data to calibrate the matrix score and confidence level, and presets the fluctuation threshold for each scoring dimension; based on the matrix features and the competence element missing identifier set, it generates a follow-up questioning technology intervention scheme. The specific process of constructing the dynamic adaptive scoring matrix that links competence and follow-up questions is as follows: build a matrix framework with the core competence dimensions as rows and the subdivided evaluation indicators as columns; calculate the initial score based on element completeness and semantic fit. Establish a mapping relationship between follow-up questions and matrix cells. For each follow-up question with valid information, automatically update the corresponding cell score and dimension total score. Integrate multi-dimensional monitoring status data to calibrate score deviations and simultaneously calculate confidence values. Role matching and follow-up question generation module: Pre-set differentiated virtual interviewer role library, each role is associated with technical intervention logic; Combine the candidate’s competency shortcomings in the previous round of responses, multi-dimensional monitoring status data and follow-up question technical intervention scheme, dynamically match virtual roles; Targeted matching between the technical intervention logic corresponding to the role and the competency element missing identifier set is used to form personalized follow-up question content; The specific process of targeted matching between the technical intervention logic corresponding to the role and the competency element missing identifier set is as follows: deconstruct the questioning direction, rhetoric style, and in-depth mining rules in the role's technical intervention logic, and extract the missing type, correlation dimension, and priority information from the competency element missing identifier set; Establish a core element mapping table, match the corresponding follow-up questions according to the missing type, determine the depth of exploration according to priority, and adapt the dialogue style according to the relevance dimension to form a logical chain of follow-up questions. Targeted metacognitive question output module: When the candidate's core meta-competency elements are not sufficiently represented or the score confidence level does not reach the preset threshold, the module extracts and integrates elements to finally output targeted metacognitive follow-up questions. The specific process of element extraction and association fusion is as follows: core meta-competency elements and related arguments are extracted from candidate responses using keyword weight calculation methods and semantic role labeling technology; the core meta-competency elements focus on the metacognitive level, and the related arguments refer to the specific content in the candidate responses that supports the above meta-competency elements; the extraction results are cross-dimensionally correlated and verified with previous response data and competency element missing identifier sets; weights are assigned according to element importance and data credibility, and multi-source data are integrated into a unified feature vector through feature integration technology; The specific process of outputting targeted metacognitive follow-up questions is as follows: Based on a unified feature vector, the metacognitive follow-up question type is determined through a preset type mapping rule; a question generation framework containing sentence structure and semantic expansion rules is constructed for each type; after inputting candidate response fragments, the core expression logic and key information nodes are extracted through semantic analysis, and the missing element information, the rhetorical style parameters in the role intervention logic, and the key nodes are matched and integrated, and the sentence structure is constructed and the semantic expansion is completed according to the framework rules; finally, after grammatical rule verification and logical consistency detection, the metacognitive follow-up questions targeting the missing elements are output.

Images