Multi-modal mixed interview question generation system and method based on artificial intelligence

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a multi-modal mixed interview question generation system and method based on artificial intelligence. According to the method, customized interview questions are provided for candidates with different backgrounds and different skill levels by dynamically selecting questions based on candidate resumes and knowledge maps, the question difficulty and knowledge emphasis are automatically adjusted through multiple rounds of answering effects and interviewer labeling data, the effectiveness and flexibility of interview evaluation are improved, and the interview evaluation efficiency is improved. Answering results and behavior anomaly detection are fused, knowledge mastering is evaluated, potential non-behavior factors can be found, manual question screening time is shortened through an automatic process, consistency of interview experiences of different candidates is guaranteed, subjective prejudice is reduced, new knowledge documents and post description can be accessed at any time, a multi-modal graph is continuously iterated, and the method is high in practicability and high in practicability. And changes of new technical fields and new post requirements are adapted.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to a multimodal hybrid interview question generation system and method based on artificial intelligence. Background Art

[0002] With the rapid development of artificial intelligence (AI), interview assessment, a core component of talent selection and competency certification, has become a hot topic of research. Traditional interview question generation methods often rely on manual design or static question banks. These methods suffer from a single format, lack of personalization, and weak cross-domain integration, making them difficult to adapt to the dynamic demands of diverse scenarios. While existing technologies attempt to incorporate some intelligent approaches, they still suffer from significant deficiencies in knowledge integration, real-time feedback, and anti-cheating measures.

[0003] Traditional interview question banks often rely solely on manual experience for maintenance and classification, resulting in a low degree of semantic relevance between the question bank content and the specific job skill requirements. Interviewers are required to manually sift through a massive amount of questions, which is time-consuming and difficult to ensure a close match with the actual job scenario, resulting in a disconnect between interview questions and the actual skills assessment. Online interview platforms are fixed once questions are generated, lacking the ability to dynamically adjust the difficulty and direction of questions based on the candidate's real-time performance. The interview process typically relies solely on accuracy rate and average time spent as evaluation metrics, failing to fully reflect a candidate's adaptability and depth of thinking in complex scenarios. It also makes it difficult to promptly compensate for a candidate's weaknesses in certain knowledge points, thus affecting the comprehensiveness and pertinence of the assessment. Summary of the Invention

[0004] The purpose of this invention is to provide a multimodal hybrid interview question generation system and method based on artificial intelligence, which can dynamically select topics based on resumes and knowledge graphs, automatically adjust the difficulty and focus, integrate behavior detection, automate processes to reduce manual labor, and continuously iterate to adapt to new needs.

[0005] The technical solutions adopted by the present invention are as follows: The artificial intelligence-based multimodal hybrid interview question generation method includes: Acquire multi-source data, including interview question banks, job skill descriptions, and domain knowledge documents, and build a correlation graph based on the multi-source data; Obtain target job requirements and interview data, where interview data includes job type, job difficulty, and candidate resume information. Obtain dynamic job questions based on the relationship graph, target job requirements, and interview data. Obtain candidate answer performance data based on the dynamic job question bank, obtain subsequent question adjustment strategies based on the answer performance data, and adjust subsequent questions of the dynamic job questions based on the subsequent question adjustment strategies; Obtain interviewer's marked question quality data, and obtain candidate's subsequent answer performance data based on the subsequent questions after adjusting the dynamic position question, obtain subsequent question fine-tuning strategy based on the marked question quality data and the subsequent answer performance data, and use the subsequent question fine-tuning strategy as the subsequent question adjustment strategy to adjust the subsequent questions of the dynamic position question; Obtain candidate answering behavior data, obtain candidate answering abnormality scores based on the answering behavior data, and generate an ability dimension analysis report based on all candidate answering effect data.

[0006] In a preferred solution, multi-source data is obtained, wherein the multi-source data includes an interview question bank, job skill descriptions, and domain knowledge documents. The steps of establishing a correlation map based on the multi-source data include: Acquire multi-source data, including interview question banks, job skill descriptions, and domain knowledge documents; Perform named entity recognition based on multi-source data to extract technical, theoretical, and business knowledge point entities; Obtain manual rules and, based on these rules, define the relationships between knowledge points. These relationships include hierarchical relationships, cross-domain dependencies, and application scenario relationships. The knowledge point entities are regarded as nodes and the association relationships as edges to construct an association graph.

[0007] In a preferred solution, target job requirements and interview data are obtained, wherein the interview data includes job type, job difficulty, and candidate resume information. The steps of obtaining dynamic job questions based on the relationship map, target job requirements, and interview data include: Obtain target job requirements and interview data, including job type, job difficulty, and candidate resume information; Target position question bank that meets the target position requirements based on the relationship map; Obtain the corresponding job type vector and job difficulty vector according to job type and job difficulty respectively; Obtain a candidate resume matrix and skill weights based on the candidate resume information, and obtain multiple candidate resume vectors based on the candidate resume matrix; Obtain preliminary question values ​​based on the job type vector, job difficulty vector, skill weight, and multiple candidate resume vectors; Obtaining a preliminary question table, wherein the preliminary question table includes a plurality of preliminary question value intervals and preliminary question strategies corresponding to each preliminary question value interval; Obtain the corresponding preliminary question strategy from the preliminary question table according to the preliminary question value interval corresponding to the preliminary question value; Generate dynamic job questions based on preliminary question strategy and target job question bank.

[0008] In a preferred embodiment, the steps of obtaining candidate answer performance data based on a dynamic job question bank, obtaining a subsequent question adjustment strategy based on the answer performance data, and adjusting subsequent questions of the dynamic job question based on the subsequent question adjustment strategy include: Get the number of stage answer evaluations based on the dynamic job question bank; Obtain answer performance data of candidates who meet the number of evaluation questions at each stage based on the dynamic job question bank; Based on the answering effect data, the corresponding answering error rate, time-consuming characteristic data of wrong questions, wrong question sets and wrong question correlation are obtained; Obtain related questions based on the relevance of wrong questions and the question bank of the target position; Based on the answer error rate, the time-consuming characteristic data of the wrong questions and the wrong question set, corresponding answer error weights, multiple wrong question time-consuming vectors and corresponding multiple wrong question vectors are obtained respectively; Obtaining examination point values ​​based on the weight of wrong answers, multiple wrong answer time vectors, and multiple wrong answer vectors; Obtaining an examination point table, wherein the examination point table includes a plurality of examination point value intervals and a subsequent question adjustment strategy corresponding to each examination point value interval; Obtain the corresponding subsequent question adjustment strategy from the examination point table according to the examination point value interval corresponding to the examination point value; Adjust the subsequent questions of dynamic job questions based on the subsequent question adjustment strategy, target job question bank and wrong question related questions.

[0009] In a preferred embodiment, the steps of obtaining interviewer's marked question quality data, obtaining candidate's subsequent answer performance data based on the subsequent questions after adjusting the dynamic job question, obtaining a subsequent question fine-tuning strategy based on the marked question quality data and the subsequent answer performance data, and adjusting the subsequent questions of the dynamic job question using the subsequent question fine-tuning strategy as the subsequent question adjustment strategy include: Obtain interviewer's annotated question quality data; Obtain the candidate's historical answer performance data, and combine it with the marked question quality data to obtain the number of answer evaluations in the subsequent stage; Obtain the subsequent answer performance data of candidates for the subsequent questions after the dynamic job questions are adjusted based on the number of answer evaluations in the subsequent stage; Obtain subsequent examination point values ​​based on the marked question quality data and subsequent answering effect data; Obtaining a subsequent examination point table, wherein the subsequent examination point table includes multiple subsequent examination point intervals and a subsequent question fine-tuning strategy corresponding to each subsequent examination point interval; Obtain the corresponding subsequent question fine-tuning strategy from the subsequent examination point table according to the subsequent examination point value interval corresponding to the subsequent examination point value; The subsequent question fine-tuning strategy is returned as the subsequent question adjustment strategy to the subsequent questions for adjusting the dynamic position questions based on the subsequent question adjustment strategy until the number of dynamic position questions is met.

[0010] In a preferred embodiment, the steps of obtaining the candidate's historical answer performance data and combining it with the interviewer's annotated question quality data to obtain the number of answer evaluations in the subsequent stage include: Obtain the candidate's historical answer performance data, and based on the historical answer performance data, obtain the historical answer error rate and historical wrong answer time characteristic data; Obtain corresponding historical answer error values ​​and historical answer time characteristic values ​​based on historical answer error rates and historical answer time characteristic data; Obtain the number of standard stage answer evaluations, standard answer error values, and standard error time characteristic values; Obtain the quality weight of the marked questions based on the interviewer's marked question quality data; The number of answer evaluations in the subsequent stages is obtained based on historical answer error values, historical wrong question time-consuming characteristic values, marked question quality weights, the number of standard stage answer evaluations, standard answer error values, and standard wrong question time-consuming characteristic values.

[0011] In a preferred embodiment, the step of obtaining subsequent examination point values ​​based on the marked question quality data and the subsequent answer effect data includes: Obtain the quality weight of the marked questions based on the interviewer's marked question quality data; Based on the candidate's subsequent answering effect data, the corresponding subsequent answering error rate, subsequent wrong answering time characteristic data, and subsequent wrong answering set are obtained; Based on the subsequent answer error rate, the subsequent wrong question time-consuming feature data and the subsequent wrong question set, corresponding subsequent answer error weights, multiple subsequent wrong question time-consuming vectors and corresponding multiple subsequent wrong question vectors are obtained respectively; The subsequent examination point value is obtained according to the marked question quality weight, the subsequent answer error weight, the time-consuming vector of multiple subsequent wrong questions and the multiple subsequent wrong question vectors.

[0012] In a preferred embodiment, the steps of obtaining candidate answering behavior data, obtaining candidate answering abnormality scores based on the answering behavior data, and generating an ability dimension analysis report based on all candidate answering effect data include: Obtain candidate answering behavior data; Obtain screen switching frequency, clipboard operation data, and input interval characteristics based on answering behavior data; Based on the screen switching frequency, clipboard operation data and input interval characteristics, corresponding screen switching vectors, clipboard operation vectors and input feature vectors are respectively obtained; Obtaining an abnormal value according to the screen cutting vector, the clipboard operation vector and the input feature vector; Obtaining a score sheet, wherein the score sheet includes multiple outliers and an abnormal score of the answer corresponding to each outlier; Obtain the corresponding abnormal score of the answer from the score table according to the abnormal value; Get the scoring threshold and determine whether the anomaly score exceeds the scoring threshold; If the abnormal score does not exceed the scoring threshold, the answer is considered normal, and an ability dimension analysis report is generated based on all the candidate's answer performance data; If the abnormal score exceeds the scoring threshold, the answer is determined to be abnormal, and an ability dimension analysis report is generated based on all the candidate's answer effect data.

[0013] The present invention also provides an artificial intelligence-based multimodal mixed interview question generation system, which is used in the above-mentioned artificial intelligence-based multimodal mixed interview question generation method, including: The knowledge management module is used to obtain multi-source data, including interview question banks, job skill descriptions, and domain knowledge documents, and to build a correlation map based on the multi-source data; The question generation module is used to obtain target job requirements and interview data, where the interview data includes job type, job difficulty, and candidate resume information. Dynamic job questions are obtained based on the relationship map, target job requirements, and interview data. The question optimization module is used to obtain candidate answer performance data based on the dynamic position question bank, obtain subsequent question adjustment strategies based on the answer performance data, and adjust subsequent questions of the dynamic position questions based on the subsequent question adjustment strategies; The question feedback module is used to obtain the interviewer's marked question quality data, and obtain the candidate's subsequent answer performance data based on the subsequent questions after adjusting the dynamic position question, obtain the subsequent question fine-tuning strategy based on the marked question quality data and the subsequent answer performance data, and use the subsequent question fine-tuning strategy as the subsequent question adjustment strategy to adjust the subsequent questions of the dynamic position question; The comprehensive module is used to obtain candidate answering behavior data, obtain candidate answering abnormality scores based on the answering behavior data, and generate an ability dimension analysis report based on all candidate answering effect data.

[0014] And, the AI-based multimodal hybrid interview question generation terminal includes: one or more processors; a storage device having one or more programs stored thereon; When one or more programs are executed by one or more processors, the one or more processors implement an artificial intelligence-based multimodal hybrid interview question generation method.

[0015] The technical effects achieved by the present invention are: The present invention dynamically selects topics based on candidate resumes and knowledge graphs, and provides tailored interview questions for candidates with different backgrounds and skill levels. Through multiple rounds of answering results and interviewer annotation data, the difficulty of the questions and the focus of knowledge are automatically adjusted to improve the effectiveness and flexibility of interview evaluation. The integration of answering results and behavioral anomaly detection not only evaluates knowledge mastery but also discovers potential non-behavioral factors. The automated process reduces the time for manual screening, ensures the consistency of interview experience for different candidates, and reduces subjective bias. New knowledge documents and job descriptions can be accessed at any time, and the multimodal graph is continuously iterated to adapt to changes in new technology fields and new job requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the method provided by the present invention; Figure 2 This is a system module diagram provided by the present invention. DETAILED DESCRIPTION

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0018] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0019] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in a preferred embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it constitute a separate or selective embodiment that is mutually exclusive of other embodiments.

[0020] Secondly, the present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention in detail, for the sake of convenience, the schematic diagrams are only examples and should not limit the scope of protection of the present invention.

[0021] Please see the attached Figure 1 As shown, a multimodal hybrid interview question generation method based on artificial intelligence is provided, including: S1. Obtain multi-source data, including interview question banks, job skill descriptions, and domain knowledge documents, and establish a correlation graph based on the multi-source data; S2. Obtain target job requirements and interview data, where interview data includes job type, job difficulty, and candidate resume information. Obtain dynamic job questions based on the relationship graph, target job requirements, and interview data. S3. Obtain candidate answer performance data based on the dynamic job question bank, obtain subsequent question adjustment strategies based on the answer performance data, and adjust subsequent questions of the dynamic job questions based on the subsequent question adjustment strategies; S4. Obtain interviewer's marked question quality data, and obtain candidate's subsequent answer performance data based on the subsequent questions after adjusting the dynamic position question, obtain subsequent question fine-tuning strategy based on the marked question quality data and the subsequent answer performance data, and use the subsequent question fine-tuning strategy as the subsequent question adjustment strategy to adjust the subsequent questions of the dynamic position question; S5. Obtain the candidate's answering behavior data, obtain the candidate's abnormal answering score based on the answering behavior data, and generate an ability dimension analysis report based on all the candidate's answering effect data.

[0022] As in steps S1 to S5 above, interview question banks (structured / semi-structured questions), job skill descriptions (JD, competency models) and domain knowledge documents (white papers, technical specifications, etc.) are collected, named entity recognition is performed on multi-source texts, and technical (such as algorithms, frameworks), theoretical (such as design patterns, system principles) and business (such as industry terminology, typical scenarios) knowledge points are extracted. Combined with manual experience rules, the hierarchical relationship (superordinate-inferior), cross-domain dependency (such as "deep learning" depends on "linear algebra") and application scenario association (such as the application of "caching" in high-concurrency systems) between knowledge points are defined. Knowledge point entities are used as nodes and association relationships as edges to generate a multimodal association graph to support subsequent intelligent retrieval and reasoning. The target job requirements (such as job type: back-end development; job difficulty: intermediate and senior) and candidate resume information (education background, project experience, skill tags) are input, and the knowledge subgraph matching the job requirements is located in the association graph to construct a "target job question bank". Preliminary question values ​​are obtained based on job type, job difficulty and candidate resume information, and Based on preset question value ranges and strategies, a dynamic question list for the first round is automatically generated. For candidates' first-round answers, the error rate, time spent answering, set of incorrect questions, and correlation between incorrect questions are recorded. The "test point values" for knowledge assessment points are then mapped to a pre-defined test point table. Corresponding adjustment strategies for subsequent questions (such as "strengthening basic modules" and "focusing on application scenarios") are then obtained. Based on these adjustment strategies, the target position question bank, and correlation between incorrect questions, dynamic questions for the second and subsequent rounds are generated. Interviewers then rate the quality of the generated questions (difficulty, coverage, innovation, etc.), re-evaluate the "subsequent test point values" based on the candidate's subsequent answer performance and the annotated quality weights, and obtain "fine-tuning strategies" from the subsequent test point table. These fine-tuning strategies are used as the basis for new question adjustments. This process is repeated until the question quantity and quality requirements are met. Behavioral characteristics such as the candidate's screen switching frequency, clipboard operations, and typing intervals are recorded. These behavioral characteristics are quantized and matched against a pre-defined scoring table to determine whether there are any anomalies such as cheating or excessive reliance on external resources. If the anomaly score does not exceed the threshold, the candidate is marked "normal."Otherwise, the result is marked as "abnormal answer." Ultimately, all answer performance and behavioral data are combined to generate a comprehensive analysis report divided by ability dimensions (such as knowledge mastery, adaptability, and problem-solving efficiency). Dynamic topic selection is based on the candidate's resume and knowledge graph, providing tailored interview questions for candidates with different backgrounds and skill levels. Based on multiple rounds of answer performance and interviewer annotation data, question difficulty and knowledge focus are automatically adjusted, improving the effectiveness and flexibility of interview assessments. The integration of answer results and behavioral anomaly detection not only assesses knowledge mastery but also identifies potential non-behavioral factors (such as reliance on cheating). The automated process reduces manual question screening time, ensures a consistent interview experience for different candidates, and reduces subjective bias. New knowledge documents and job descriptions can be accessed at any time, and the multimodal graph is continuously iterated to adapt to changes in new technology fields and job requirements.

[0023] In a preferred embodiment, multi-source data is obtained, wherein the multi-source data includes an interview question bank, job skill descriptions, and domain knowledge documents, and the step of establishing a correlation map based on the multi-source data includes: S101. Acquire multi-source data, where the multi-source data includes an interview question bank, job skill descriptions, and domain knowledge documents; S102. Perform named entity recognition based on multi-source data to extract technical, theoretical, and business knowledge point entities; S103: Obtaining manual rules and defining associations between knowledge points based on the manual rules, wherein the associations include hierarchical relationships, cross-domain dependency relationships, and application scenario associations; S104. Take the knowledge point entities as nodes and the association relationships as edges to construct an association graph.

[0024] As in steps S101 to S104 above, original materials are collected from three different dimensions: interview question banks, i.e., structured or semi-structured questions and answers from previous years, question labels (difficulty, knowledge point categories, etc.), job skill descriptions, i.e., recruitment requirements and capability models (such as skill matrices, job competency frameworks), and domain knowledge documents, i.e., unstructured texts such as industry white papers, technical standards, and academic papers. Pre-trained or self-trained NER models are used to perform word segmentation, part-of-speech tagging, and entity recognition on multi-source texts. Technical entities (algorithms, frameworks, tool names), theoretical entities (design patterns, principle concepts), and business entities (industry terms, business processes) are extracted from interview question banks, job skill descriptions, and domain knowledge documents respectively. A list of "knowledge point entities" is output to lay the foundation for subsequent graph nodes. Manual rules provided by senior interviewers and domain experts, such as experiential knowledge such as "advanced algorithms rely on linear algebra" and "caching is generally used in high-concurrency scenarios", define the relationship between knowledge points, and hierarchical Relationships (superordinate-inferior, such as "data structure" → "binary tree"), cross-domain dependencies (such as "deep learning" depends on "matrix operations"), application scenario associations (such as the application of "distributed cache" in "microservice architecture"), the rule engine matches the entities identified in the text in pairs, and automatically labels the corresponding relationship types according to the rules, treating all knowledge point entities as graph nodes and attaching attributes (such as source, category, weight), adding directed or undirected edges based on the three defined association relationships, and carrying relationship type labels and credibility indicators on the edges. It is saved in a graph database (such as Neo4j) or a knowledge graph framework (such as RDF / OWL), supports fast query, path search and reasoning, integrates heterogeneous data into the same structured graph, breaks down information silos, and clearly presents the implicit logical relationship between the interview question bank, job skill descriptions and domain documents. Graph-based proximity search and path query can quickly locate interview questions that are highly relevant to a job skill or knowledge point, reducing manual screening time.

[0025] In a preferred embodiment, target job requirements and interview data are obtained, wherein the interview data includes job type, job difficulty, and candidate resume information. The steps of obtaining dynamic job questions based on the relationship map, target job requirements, and interview data include: S201. Obtain target job requirements and interview data, where the interview data includes job type, job difficulty, and candidate resume information; S202. Target position question bank that meets target position requirements based on the relationship map; S203. Obtain corresponding job type vectors and job difficulty vectors according to job type and job difficulty respectively; S204. Obtain a candidate resume matrix and skill weights based on the candidate resume information, and obtain multiple candidate resume vectors based on the candidate resume matrix; S205. Obtain preliminary question values ​​based on the job type vector, job difficulty vector, skill weight, and multiple candidate resume vectors; S206: Obtain a preliminary question table, wherein the preliminary question table includes a plurality of preliminary question value intervals and preliminary question strategies corresponding to each preliminary question value interval; S207, obtaining a corresponding preliminary question strategy from the preliminary question table according to the preliminary question value interval corresponding to the preliminary question value; S208. Generate dynamic job questions based on the preliminary question strategy and the target job question bank.

[0026] As in steps S201 to S208 above, the job type (such as "back-end development" or "data analysis"), job difficulty (such as "junior," "intermediate," or "senior"), and candidate resume information (structured extracted educational background, project experience, skill tags, etc.) are obtained from the recruitment system or user input interface. In the pre-constructed association graph, "job type" and "job skills" are used as query conditions to retrieve the question node subgraph that best matches the target job, forming an initial question bank. The job type (category label) and job difficulty (difficulty level) are respectively mapped to the vector space to obtain a job type vector and a job difficulty vector. The skill tags and related attributes in the candidate's resume are organized into a resume matrix and assigned skill weights. Through matrix decomposition or other dimensionality reduction techniques, multiple resume vectors of the candidate are obtained to reflect their overall skill distribution. The job type vector, job difficulty vector, skill weight, and candidate resume vector are combined to calculate the preliminary question value based on the current candidate's match with the job. This value can be understood as the "difficulty-matching weighted score" of the question. The preliminary question value calculation formula is: , where C represents the preliminary title value, i represents the number of multiple candidate resume vectors, i=1,2,3…n, It is represented as the resume vector of the i-th candidate, L is represented as the job type vector, and N is represented as the job difficulty vector. Represented as skill weights, a predefined table maps several "preliminary question value intervals" to corresponding preliminary question strategies (such as "mainly basic questions", "mainly intermediate questions", and "mainly challenging questions"). Based on which interval the preliminary question value of each question falls into, the corresponding question strategy is read from the table to determine whether the question should belong to the "basic", "intermediate", or "advanced" question bank. According to the selected preliminary question strategy, the corresponding number and type of test questions are extracted from the target position question bank, combined into a first-round dynamic question list, and output for use by the system or interviewer. Based on two-way vectorized calculations of positions and candidates, it is ensured that the selected questions not only meet the position requirements but also objectively reflect the candidate's current level. The preliminary question table and strategy can be freely configured according to corporate culture and interview preferences to meet the needs of different organizations or different positions.

[0027] In a preferred embodiment, the steps of obtaining candidate answer performance data based on a dynamic job question bank, obtaining a subsequent question adjustment strategy based on the answer performance data, and adjusting subsequent questions of the dynamic job question based on the subsequent question adjustment strategy include: S301. Obtain the number of stage answer evaluations based on the dynamic job question bank; S302. Obtaining answer performance data of candidates who meet the number of stage answer evaluations based on the dynamic job question database; S303, based on the answering effect data, obtaining the corresponding answering error rate, time-consuming characteristic data of wrong questions, wrong question set and wrong question correlation; S304, obtaining related questions of the wrong questions based on the relevance of the wrong questions and the question bank of the target position; S305: Based on the error rate of the questions, the characteristic data of the time consumed by the wrong questions, and the wrong question set, obtain the corresponding error weight of the questions, multiple wrong question time consumed vectors, and multiple corresponding wrong question vectors; S306, obtaining a test point value based on the wrong answer weight, multiple wrong answer time vectors, and multiple wrong answer vectors; S307: Obtain an examination point table, wherein the examination point table includes multiple examination point value intervals and subsequent question adjustment strategies corresponding to each examination point value interval; S308, obtaining the corresponding subsequent question adjustment strategy from the examination point table according to the examination point value interval corresponding to the examination point value; S309. Adjust the subsequent questions of the dynamic position questions based on the subsequent question adjustment strategy, the target position question bank, and the wrong question-related questions.

[0028] As in the above steps S301 to S309, the "stage answer evaluation quantity" is determined according to the scale of the dynamic question bank and the preset evaluation strategy (such as 5 questions need to be completed in each stage), to ensure that the data volume of each adjustment round is consistent and comparable, and the answer results of the current candidate for this batch of questions in the dynamic question bank in this stage are collected, including whether each question is correct, the answer time, the number of submissions, etc., to form a complete "answer effect data" set, and the answer error rate (number of wrong questions / total number of questions), the time-consuming characteristics of wrong questions (average time consuming, maximum time consuming, etc. of wrong questions), the wrong question set (list of wrong question IDs), the correlation of wrong questions (based on the graph or knowledge point) are calculated from the answer effect data. , measure the knowledge correlation between wrong questions), use the "knowledge point-question" mapping pre-built in the association map, find the "wrong question related questions" related to the current wrong question knowledge point according to the correlation of wrong questions (for subsequent supplementary exercises), the answer error weight is to map the error rate to the weight (such as a high error rate gives a higher weight), the wrong question time vector is to construct the time consumption of each wrong question into a vector to reflect the time consumption distribution, the wrong question vector is to convert the wrong question set into a knowledge point vector (such as the entity vector corresponding to each question), comprehensively consider the answer error weight, time consumption vector and wrong question vector, calculate the test point value, and quantify the candidate's weakness in knowledge points. The calculation formula of the test point value is: , where K represents the inspection point value, g represents the number of multiple wrong question time-consuming vectors and the number of multiple wrong question vectors, g=1,2,3…m, It is represented as the time-consuming vector of the g-th wrong question, Represented as the g-th wrong question vector, Expressed as error weight, the mapping of "examination point value interval → subsequent question adjustment strategy" is predefined. According to the interval in which the examination point value of each knowledge examination point falls, the corresponding subsequent question adjustment strategy is read from the mapping table to determine the rhythm of focusing on reinforcement or reducing difficulty. Combined with the read adjustment strategy, the target position question bank and the obtained wrong question-related questions, the dynamic question list for the next round is intelligently screened and combined to accurately fill the candidate's knowledge gaps. With the wrong questions and their related knowledge points as the center, the candidate's knowledge weaknesses are accurately located to avoid the problem coverage being too wide and distracting the evaluation focus. Starting from multi-dimensional quantitative indicators such as error rate, time consumption, and correlation, the strategy formulation is more scientific and reduces subjective experience-based deviations. With the help of correlation maps, the related questions of wrong questions are expanded, which can not only fill the direct wrong questions, but also strengthen the related upper-level or application scenarios, helping candidates to achieve more comprehensive preparation.

[0029] In a preferred embodiment, the steps of obtaining interviewer's marked question quality data, obtaining candidate's subsequent answer performance data based on the subsequent questions after adjusting the dynamic job question, obtaining a subsequent question fine-tuning strategy based on the marked question quality data and the subsequent answer performance data, and using the subsequent question fine-tuning strategy as a subsequent question adjustment strategy to adjust the subsequent questions of the dynamic job question include: S401. Obtaining interviewer's marked question quality data; S402: Obtain the candidate's historical answer performance data, and combine it with the marked question quality data to obtain the number of answer evaluations in the subsequent stage; S403. Obtaining subsequent answer performance data of candidates for subsequent questions after adjusting the dynamic position questions based on the number of answer evaluations in the subsequent stage; S404, obtaining subsequent examination point values ​​based on the marked question quality data and subsequent answer effect data; S405: Obtain a subsequent examination point table, wherein the subsequent examination point table includes multiple subsequent examination point intervals and a subsequent question fine-tuning strategy corresponding to each subsequent examination point interval; S406, obtaining a corresponding subsequent question fine-tuning strategy from the subsequent examination point table according to the subsequent examination point value interval corresponding to the subsequent examination point value; S407: Return the subsequent question fine-tuning strategy as the subsequent question adjustment strategy to the subsequent questions for adjusting the dynamic job questions based on the subsequent question adjustment strategy until the number of dynamic job questions is met.

[0030] As in steps S401 to S407 above, the interviewer scores or labels the generated subsequent questions according to preset dimensions (such as difficulty adaptability, content coverage, innovation, etc.) to form "labeled question quality data", and calculates the quality weight of each interviewer, reads the candidate's historical answer performance data (including error rate, time consumption, answer completion, etc.), and calculates the "subsequent stage answer evaluation number" in combination with the interviewer's quality weight to ensure that the new answer number can reflect the actual level and meet the ideal evaluation accuracy. According to the obtained subsequent stage answer evaluation number, the corresponding questions are pushed to the candidate, and the answer results of the candidate in this round are collected to form the "candidate's subsequent answer effect data", including accuracy rate, time consumption, wrong questions, etc., and the "subsequent examination point value" is comprehensively calculated for the knowledge examination points using the labeled question quality weight and the new round of answer effect data. For example, errors in high-quality questions will be given higher weights to prompt the system to focus on the key knowledge links identified by the interviewer. , pre-define the mapping table of "subsequent examination point value interval → subsequent question fine-tuning strategy", map the subsequent examination point value of each examination point to the corresponding interval, read the "subsequent question fine-tuning strategy" from the table, generate more targeted fine-tuning instructions (such as "increase difficulty", "reduce difficulty", "change question type" or "adjust knowledge focus"), and input the fine-tuning strategy as the new "subsequent question adjustment strategy" into the dynamic question generation process, and continuously iterate the "question-answer-feedback-fine-tuning" closed loop until the predetermined question quantity and quality standards are met. It not only relies on the automated data-driven capabilities but also integrates the professional judgment of the interviewer to achieve "machine + human" collaborative optimization, amplifies the feedback effect of high-value questions through quality weighting, and the difficulty and coverage of the questions can converge to the ideal range more quickly. The quantity and quality of the subsequent stage evaluation are determined by both the candidate's historical performance and the quality of the interviewer, reducing the deviation of a single dimension. The fine-tuning strategy can be cycled multiple times to ensure that the final question bank accurately matches the job requirements in terms of quantity, difficulty and knowledge point coverage.

[0031] In a preferred embodiment, the steps of obtaining the candidate's historical answer performance data and combining it with the interviewer's annotated question quality data to obtain the number of answer evaluations in the subsequent stage include: S4021. Obtain the candidate's historical answer performance data, and based on the historical answer performance data, obtain the historical answer error rate and historical wrong answer time characteristic data; S4022. Obtain corresponding historical answer error values ​​and historical answer time characteristic values ​​based on historical answer error rates and historical answer time characteristic data; S4023, obtaining the number of standard stage answer evaluations, standard answer error values, and standard error time characteristic values; S4024. Obtaining a quality weight of the marked questions based on the interviewer's marked question quality data; S4025. Obtain the number of evaluation questions in the subsequent stages based on the historical answer error values, the time-consuming characteristic values ​​of historical wrong questions, the quality weights of the marked questions, the number of evaluation questions in the standard stage, the standard answer error values, and the time-consuming characteristic values ​​of the standard wrong questions.

[0032] As in steps S4021 to S4025 above, key indicators are extracted from all previous answer records of the candidate, including the total number of answers, the number of incorrect questions, a list of incorrect questions, and the time taken to answer each incorrect question. The historical error rate (the number of incorrect questions divided by the total number of answers) and the time taken to answer incorrect questions (such as the average time taken to answer incorrect questions) are calculated and archived. The historical error rate is mapped to a historical error value, and the time taken to answer incorrect questions is mapped to a historical time taken to answer incorrect questions feature value. The number of standard stage answer evaluation questions (for example, 5 questions are recommended for each stage) is pre-set or obtained based on large-scale sample statistics. , standard answer error value (for example, the industry average error value is 30 points), standard wrong question time characteristic value (for example, the average wrong question time is 60 seconds), according to the interviewer's quality annotation of the question (difficulty, coverage, innovation and other dimensions), generate a marked question quality weight, according to the historical answer error value, historical wrong question time characteristic value, marked question quality weight, standard stage answer evaluation number, standard answer error value and standard wrong question time characteristic value, round up to the nearest natural number, get the subsequent stage answer evaluation number, the calculation formula for the subsequent stage answer evaluation number is: , where L represents the number of evaluation questions in the subsequent stage, Indicates the historical answer error value, It is represented by the time-consuming characteristic value of historical wrong questions. It represents the weight of the quality of the marked questions, F represents the number of evaluation questions in the standard stage, Represents the standard answer error value, Expressed as a standard time-consuming characteristic value for incorrect questions, it compares the candidate's actual historical performance with industry standards. The number of evaluations is neither too small to lack representativeness nor too large to cause fatigue. The interviewer quality weighting further modifies the algorithm so that the question allocation takes into account both "machine judgment" and "expert intention", better suiting the specific needs of the organization or position. Combining the two dimensions of error rate and time consumption can more comprehensively reflect the candidate's knowledge mastery and problem-solving proficiency, ensuring the quality of subsequent data samples.

[0033] In a preferred embodiment, the step of obtaining subsequent examination point values ​​based on the marked question quality data and the subsequent answer effect data includes: S4041. Obtaining a quality weight of the marked questions based on the interviewer's marked question quality data; S4042, based on the candidate's subsequent answering effect data, obtain the corresponding subsequent answering error rate, subsequent wrong answering time characteristic data, and subsequent wrong answering set; S4043, based on the subsequent question answer error rate, the subsequent wrong question time consumption characteristic data and the subsequent wrong question set, respectively obtain the corresponding subsequent question answer error weight, multiple subsequent wrong question time consumption vectors and corresponding multiple subsequent wrong question vectors; S4044. Obtain subsequent examination point values ​​based on the marked question quality weight, the subsequent answer error weight, multiple subsequent wrong question time consumption vectors, and multiple subsequent wrong question vectors.

[0034] As in steps S4041 to S4044 above, the interviewer scores the marked questions according to multiple dimensions (such as question novelty, coverage, and difficulty matching), and maps the score to a marked question quality weight. From the candidate's answer records in subsequent questions in this round, statistics are collected on the subsequent answer error rate, subsequent wrong question time-consuming characteristics, and subsequent wrong question sets. The subsequent error rate is converted into a weight index according to a preset table, and the time-consuming characteristics of each wrong question are constructed into multiple vectors. Using the question→knowledge point mapping in the knowledge graph, the wrong question set is converted into a corresponding knowledge point vector, indicating the candidate's weak distribution in each knowledge point. The marked question quality weight, the subsequent answer error weight, multiple subsequent wrong question time-consuming vectors, and multiple subsequent wrong question vectors are used to calculate the subsequent examination point value. The calculation formula for the subsequent examination point value is: , where It represents the value of the subsequent examination point, h represents the number of the subsequent wrong question time-consuming vectors and the number of the subsequent wrong question vectors, h=1,2,3…j, It is represented as the time-consuming vector of the hth subsequent wrong question, Represented as the hth subsequent wrong question vector, Denoted as the subsequent error weight, Expressed as the weight of the labeled question quality, it also takes into account error rate, time consumption, and knowledge distribution. The examination point value more comprehensively reflects the candidate's actual mastery of the knowledge. Subsequent examination point values ​​directly drive the question fine-tuning strategy, making each round of question generation more targeted and effective.

[0035] In a preferred embodiment, the steps of obtaining candidate answering behavior data, obtaining candidate answering anomaly scores based on the answering behavior data, and generating an ability dimension analysis report based on all candidate answering performance data include: S501, obtaining candidate answering behavior data; S502: Obtain screen switching frequency, clipboard operation data, and input interval characteristics based on the answering behavior data; S503, obtaining corresponding screen switching vectors, clipboard operation vectors, and input feature vectors based on the screen switching frequency, clipboard operation data, and input interval characteristics; S504, obtaining an abnormal value according to the screen cutting vector, the clipboard operation vector and the input feature vector; S505: Obtain a scoring table, wherein the scoring table includes multiple outliers and an abnormal score of the answer corresponding to each outlier; S506. Obtain the corresponding abnormal answer score from the score table according to the abnormal value; S507: Obtain a scoring threshold, and determine whether the abnormality score exceeds the scoring threshold; If the abnormal score does not exceed the scoring threshold, the answer is considered normal, and an ability dimension analysis report is generated based on all the candidate's answer performance data; If the abnormal score exceeds the scoring threshold, the answer is determined to be abnormal, and an ability dimension analysis report is generated based on all the candidate's answer effect data.

[0036] As in steps S501 to S507 above, during the candidate's answering process, the mouse and keyboard behavior logs are automatically collected, including focus switching events, clipboard usage records, timestamps of each input, etc., to extract key behavioral features. The screen switching frequency is the number of times the candidate switches browsers / windows in one answering session. The clipboard operation is the number of "copy-paste" events and the corresponding content length. The input interval feature is the calculation of the average time interval and its fluctuation between two consecutive typed characters. The screen switching frequency is mapped into a screen switching vector, the number of clipboard operations and the paste length are mapped into a clipboard operation vector, and the average value of the input interval constitutes an input feature vector. The three vectors are combined to calculate an outlier value to reflect the candidate's possible external resource dependence or cheating risk. For example, the outlier value calculation formula is: Where a represents the outlier, q represents the screen-slicing vector, t represents the clipboard operation vector, and z represents the input feature vector. A predefined mapping table maps different outlier value intervals to specific answer anomaly scores (e.g., a 0–10 scale, with higher scores indicating greater risk). Based on the obtained outlier value, the corresponding interval is searched in the score table, the candidate's answer anomaly score is read, and the preset score threshold is read. If the anomaly score does not exceed the threshold, the candidate is deemed "normal"; if it exceeds the threshold, the candidate is deemed "abnormal." Regardless of whether the candidate is abnormal, the candidate's answer performance data at all stages (correctness rate, time spent, test point score, etc.) is combined with the behavior anomaly score to generate a comprehensive report based on predefined ability dimensions (such as "knowledge mastery," "problem-solving efficiency," and "integrity"). By monitoring behavioral anomalies in real time, potential cheating or over-reliance on external resources can be promptly identified, ensuring interview fairness. In addition to knowledge and ability assessments, the dimension of "answer integrity" is added to build a more comprehensive candidate profile. The comprehensive report, driven by both the behavior anomaly score and answer performance data, can provide interviewers with detailed ability distribution and risk warnings.

[0037] Please see the attached Figure 2As shown, the present invention also provides an artificial intelligence-based multimodal mixed interview question generation system, which is used for the above-mentioned artificial intelligence-based multimodal mixed interview question generation method, including: The knowledge management module is used to obtain multi-source data, including interview question banks, job skill descriptions, and domain knowledge documents, and to build a correlation map based on the multi-source data; The question generation module is used to obtain target job requirements and interview data, where the interview data includes job type, job difficulty, and candidate resume information. Dynamic job questions are obtained based on the relationship map, target job requirements, and interview data. The question optimization module is used to obtain candidate answer performance data based on the dynamic position question bank, obtain subsequent question adjustment strategies based on the answer performance data, and adjust subsequent questions of the dynamic position questions based on the subsequent question adjustment strategies; The question feedback module is used to obtain the interviewer's marked question quality data, and obtain the candidate's subsequent answer performance data based on the subsequent questions after adjusting the dynamic position question, obtain the subsequent question fine-tuning strategy based on the marked question quality data and the subsequent answer performance data, and use the subsequent question fine-tuning strategy as the subsequent question adjustment strategy to adjust the subsequent questions of the dynamic position question; The comprehensive module is used to obtain candidate answering behavior data, obtain candidate answering abnormality scores based on the answering behavior data, and generate an ability dimension analysis report based on all candidate answering effect data.

[0038] The above-mentioned knowledge management module automatically obtains interview question banks, job skill descriptions, and domain knowledge documents, and cleans and structures them. Based on named entity recognition and manual rules, the extracted technical / theoretical / business knowledge points are built into nodes, and the hierarchical, dependent, and scenario associations are built into edges, which are stored in the graph database to provide a knowledge basis for subsequent retrieval and reasoning. The question generation module receives the target job type, difficulty, and skill labels and experience in the candidate's resume, locates the question subgraph that is highly relevant to the job in the knowledge graph, forms a "target job question bank", maps the job type, difficulty, and resume information into vectors, calculates the "preliminary question value", and generates the first round of dynamic questions based on the preset strategy table. The question optimization module collects data such as the candidate's error rate, time consumption, and correlation of wrong questions for the first round of questions, fuses the error weight, time consumption vector, and wrong question vector to obtain the value of each knowledge test point, and maps it to the "subsequent question adjustment strategy" table. Based on the strategy and the target question bank, the questions are supplemented or replaced to form the second and multiple rounds of dynamic questions. The question feedback module collects the interviewer's quality score for each round of questions , obtain the question quality weight, integrate the quality weight with the candidate's subsequent answering effect, recalculate the "subsequent examination point value", and obtain a more refined question fine-tuning plan through the "fine-tuning strategy table", iteratively adjust until the number and quality of questions meet the standards, comprehensive module, real-time collection of screen switching, clipboard operation, input rhythm and other behavior logs, vectorize behavioral features and perform anomaly detection, map them to the anomaly score table to determine whether the answer is abnormal, combine the behavioral anomaly score with the answering effect data of all rounds, and generate a comprehensive candidate ability analysis report according to dimensions such as knowledge mastery, problem-solving efficiency, and integrity. Covering the complete process from data collection, intelligent question selection, dynamic optimization to behavior monitoring and report generation, it greatly improves the level of interview automation. Multimodal data fusion and vectorized matching ensure that the questions are both in line with job requirements and candidate capabilities, making the assessment more targeted. Through multiple rounds of answer feedback and interviewer fine-tuning, dynamic and adaptive question bank updates and strategy iterations are achieved, and the assessment effect is continuously optimized. In addition to knowledge mastery, problem-solving efficiency and abnormal answering behavior can also be quantified, achieving a comprehensive and three-dimensional candidate portrait.

[0039] And, the AI-based multimodal hybrid interview question generation terminal includes: one or more processors; a storage device having one or more programs stored thereon; When one or more programs are executed by one or more processors, the one or more processors implement an artificial intelligence-based multimodal hybrid interview question generation method.

[0040] The foregoing is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained herein shall, unless otherwise specified or limited, be implemented in accordance with conventional means in the art.

Claims

1. The method for generating multimodal mixed interview questions based on artificial intelligence is characterized by: include: Acquire multi-source data, including interview question banks, job skill descriptions, and domain knowledge documents, and build a correlation graph based on the multi-source data; Obtain target job requirements and interview data, where interview data includes job type, job difficulty, and candidate resume information. Obtain dynamic job questions based on the relationship graph, target job requirements, and interview data. Obtain candidate answer performance data based on the dynamic job question bank, obtain subsequent question adjustment strategies based on the answer performance data, and adjust subsequent questions of the dynamic job questions based on the subsequent question adjustment strategies; Obtain interviewer's marked question quality data, and obtain candidate's subsequent answer performance data based on the subsequent questions after adjusting the dynamic position question, obtain subsequent question fine-tuning strategy based on the marked question quality data and the subsequent answer performance data, and use the subsequent question fine-tuning strategy as the subsequent question adjustment strategy to adjust the subsequent questions of the dynamic position question; Obtain candidate answering behavior data, obtain candidate answering abnormality scores based on the answering behavior data, and generate an ability dimension analysis report based on all candidate answering effect data.

2. The method for generating multimodal mixed interview questions based on artificial intelligence according to claim 1 is characterized in that: Acquire multi-source data, including interview question banks, job skill descriptions, and domain knowledge documents. The steps for building a correlation graph based on the multi-source data include: Acquire multi-source data, including interview question banks, job skill descriptions, and domain knowledge documents; Perform named entity recognition based on multi-source data to extract technical, theoretical, and business knowledge point entities; Obtain manual rules and, based on these rules, define the relationships between knowledge points. These relationships include hierarchical relationships, cross-domain dependencies, and application scenario relationships. The knowledge point entities are regarded as nodes and the association relationships as edges to construct an association graph.

3. The method for generating multimodal mixed interview questions based on artificial intelligence according to claim 1 is characterized in that: Obtain target job requirements and interview data, where interview data includes job type, job difficulty, and candidate resume information. The steps for obtaining dynamic job questions based on the relationship graph, target job requirements, and interview data include: Obtain target job requirements and interview data, including job type, job difficulty, and candidate resume information; Target position question bank that meets the target position requirements based on the relationship map; Obtain the corresponding job type vector and job difficulty vector according to job type and job difficulty respectively; Obtain a candidate resume matrix and skill weights based on the candidate resume information, and obtain multiple candidate resume vectors based on the candidate resume matrix; Obtain preliminary question values ​​based on the job type vector, job difficulty vector, skill weight, and multiple candidate resume vectors; Obtaining a preliminary question table, wherein the preliminary question table includes a plurality of preliminary question value intervals and preliminary question strategies corresponding to each preliminary question value interval; Obtain the corresponding preliminary question strategy from the preliminary question table according to the preliminary question value interval corresponding to the preliminary question value; Generate dynamic job questions based on preliminary question strategy and target job question bank.

4. The method for generating multimodal mixed interview questions based on artificial intelligence according to claim 3 is characterized in that: The steps of obtaining candidate answer performance data based on a dynamic job question bank, obtaining a subsequent question adjustment strategy based on the answer performance data, and adjusting subsequent questions of the dynamic job question based on the subsequent question adjustment strategy include: Get the number of stage answer evaluations based on the dynamic job question bank; Obtain answer performance data of candidates who meet the number of stage answer evaluations based on the dynamic job question bank; Based on the answering effect data, the corresponding answering error rate, time-consuming characteristic data of wrong questions, wrong question sets and wrong question correlation are obtained; Obtain related questions based on the relevance of wrong questions and the question bank of the target position; Based on the answer error rate, the time-consuming characteristic data of the wrong questions and the wrong question set, the corresponding answer error weight, multiple wrong question time-consuming vectors and corresponding multiple wrong question vectors are obtained respectively; Obtaining examination point values ​​based on the weight of wrong answers, multiple wrong answer time vectors, and multiple wrong answer vectors; Obtaining an examination point table, wherein the examination point table includes a plurality of examination point value intervals and a subsequent question adjustment strategy corresponding to each examination point value interval; Obtain the corresponding subsequent question adjustment strategy from the examination point table according to the examination point value interval corresponding to the examination point value; Adjust the subsequent questions of dynamic job questions based on the subsequent question adjustment strategy, target job question bank and wrong question related questions.

5. The method for generating multimodal mixed interview questions based on artificial intelligence according to claim 1 is characterized in that: The steps of obtaining interviewer's marked question quality data, obtaining candidate's subsequent answer performance data based on the subsequent questions after adjusting the dynamic position question, obtaining subsequent question fine-tuning strategy based on the marked question quality data and the subsequent answer performance data, and adjusting the subsequent questions of the dynamic position question using the subsequent question fine-tuning strategy as the subsequent question adjustment strategy include: Obtain interviewer's annotated question quality data; Obtain the candidate's historical answer performance data, and combine it with the marked question quality data to obtain the number of answer evaluations in the subsequent stage; Obtain the subsequent answer performance data of candidates for the subsequent questions after the dynamic job questions are adjusted based on the number of answer evaluations in the subsequent stage; Obtain subsequent examination point values ​​based on the marked question quality data and subsequent answering effect data; Obtaining a subsequent examination point table, wherein the subsequent examination point table includes multiple subsequent examination point intervals and a subsequent question fine-tuning strategy corresponding to each subsequent examination point interval; Obtain the corresponding subsequent question fine-tuning strategy from the subsequent examination point table according to the subsequent examination point value interval corresponding to the subsequent examination point value; The subsequent question fine-tuning strategy is returned as the subsequent question adjustment strategy to the subsequent questions for adjusting the dynamic position questions based on the subsequent question adjustment strategy until the number of dynamic position questions is met.

6. The method for generating multimodal mixed interview questions based on artificial intelligence according to claim 5 is characterized in that: The steps for obtaining the candidate's historical answer performance data and combining it with the interviewer's annotated question quality data to obtain the number of answer evaluations in the subsequent stage include: Obtain the candidate's historical answer performance data, and based on the historical answer performance data, obtain the historical answer error rate and historical wrong answer time characteristic data; Obtain corresponding historical answer error values ​​and historical answer time characteristic values ​​based on historical answer error rates and historical answer time characteristic data; Obtain the number of standard stage answer evaluations, standard answer error values, and standard error time characteristic values; Obtain the quality weight of the marked questions based on the interviewer's marked question quality data; The number of answer evaluations in the subsequent stages is obtained based on historical answer error values, historical wrong question time-consuming characteristic values, marked question quality weights, the number of standard stage answer evaluations, standard answer error values, and standard wrong question time-consuming characteristic values.

7. The method for generating multimodal mixed interview questions based on artificial intelligence according to claim 5 is characterized in that: The steps for obtaining subsequent examination point values ​​based on the marked question quality data and subsequent answer effect data include: Obtain the quality weight of the marked questions based on the interviewer's marked question quality data; Based on the candidate's subsequent answering effect data, the corresponding subsequent answering error rate, subsequent wrong answering time characteristic data, and subsequent wrong answering set are obtained; Based on the subsequent answer error rate, the subsequent wrong question time-consuming feature data and the subsequent wrong question set, corresponding subsequent answer error weights, multiple subsequent wrong question time-consuming vectors and corresponding multiple subsequent wrong question vectors are obtained respectively; The subsequent examination point value is obtained according to the marked question quality weight, the subsequent answer error weight, the time-consuming vector of multiple subsequent wrong questions and the multiple subsequent wrong question vectors.

8. The method for generating multimodal mixed interview questions based on artificial intelligence according to claim 1 is characterized in that: The steps of obtaining candidate answering behavior data, obtaining candidate answering anomaly scores based on the answering behavior data, and generating a capability dimension analysis report based on all candidate answering effect data include: Obtain candidate answering behavior data; Obtain screen switching frequency, clipboard operation data, and input interval characteristics based on answering behavior data; Based on the screen switching frequency, clipboard operation data and input interval characteristics, corresponding screen switching vectors, clipboard operation vectors and input feature vectors are respectively obtained; Obtaining an abnormal value according to the screen cutting vector, the clipboard operation vector and the input feature vector; Obtaining a score sheet, wherein the score sheet includes multiple outliers and an abnormal score of the answer corresponding to each outlier; Obtain the corresponding abnormal score of the answer from the score table according to the abnormal value; Get the scoring threshold and determine whether the anomaly score exceeds the scoring threshold; If the abnormal score does not exceed the scoring threshold, the answer is considered normal, and an ability dimension analysis report is generated based on all the candidate's answer performance data; If the abnormal score exceeds the scoring threshold, the answer is determined to be abnormal, and an ability dimension analysis report is generated based on all the candidate's answer effect data.

9. An artificial intelligence-based multimodal mixed interview question generation system, applied to the artificial intelligence-based multimodal mixed interview question generation method according to any one of claims 1 to 8, characterized in that: include: The knowledge management module is used to obtain multi-source data, including interview question banks, job skill descriptions, and domain knowledge documents, and to build a correlation map based on the multi-source data; The question generation module is used to obtain target job requirements and interview data, where the interview data includes job type, job difficulty, and candidate resume information. Dynamic job questions are obtained based on the relationship map, target job requirements, and interview data. The question optimization module is used to obtain candidate answer performance data based on the dynamic position question bank, obtain subsequent question adjustment strategies based on the answer performance data, and adjust subsequent questions of the dynamic position questions based on the subsequent question adjustment strategies; The question feedback module is used to obtain the interviewer's marked question quality data, and obtain the candidate's subsequent answer performance data based on the subsequent questions after adjusting the dynamic position question, obtain the subsequent question fine-tuning strategy based on the marked question quality data and the subsequent answer performance data, and use the subsequent question fine-tuning strategy as the subsequent question adjustment strategy to adjust the subsequent questions of the dynamic position question; The comprehensive module is used to obtain candidate answering behavior data, obtain candidate answering abnormality scores based on the answering behavior data, and generate an ability dimension analysis report based on all candidate answering effect data.

10. The multimodal hybrid interview question generation terminal based on artificial intelligence is characterized by: include: one or more processors; a storage device having one or more programs stored thereon; When one or more programs are executed by one or more processors, the one or more processors implement the artificial intelligence-based multimodal hybrid interview question generation method as described in any one of claims 1 to 8.

Images