Interviewer evaluation bias correction method and system based on real-time feedback

Abstract

The present invention discloses a method and system for correcting interviewer evaluation bias based on real-time feedback, which specifically relates to the field of human resource management technology. The method and system are used to solve the problems in existing recruitment systems where evaluation standards are out of touch with industry characteristics due to standardized semantic parsing, and interviewers' implicit evaluation bias is difficult to correct dynamically. The method dynamically parses fuzzy terms in job requirements through an industry knowledge base, generates industry weights and terminology thresholds, collects interview interaction texts in real time, and analyzes the correlation between candidates and terminology thresholds. The method combines interviewer question context recognition and deviation calculation to mark terminology deviation signals and repeated question intention deviations, and finally dynamically adjusts the scoring weights according to industry correction rules. The semantic parsing and real-time feedback linkage mechanism adapted to industry characteristics achieves the alignment of evaluation standards with the actual job requirements, effectively eliminates evaluation bias caused by misunderstanding of terminology, and improves the objectivity and accuracy of talent selection.

Description

Technical Field

[0001] The present invention relates to the technical field of human resource management, and more particularly, to a method and system for correcting interviewer evaluation bias based on real-time feedback. Background Art

[0002] In existing human resource management systems, recruiters typically screen and match candidates based on job requirements described in natural language, relying on preset standardized competency tags to parse abstract terms in job descriptions to automatically associate candidates with job requirements. However, due to substantial differences in the actual competency requirements for the same terminology across different industries or companies, existing technologies use unified semantic mapping rules for demand parsing and fail to establish a dynamic association mechanism between term semantics and industry characteristics.

[0003] However, standardized parsing rules are difficult to accurately reflect the contextual semantics of ambiguous terms in job requirements, resulting in systematic deviations between the matching results generated by the system and the employer's true intentions. When interviewers evaluate candidates based on such biased data, they may form implicit evaluation biases due to the misalignment between their understanding of the terms and the actual ability standards of the position, thereby reducing the objectivity and accuracy of talent selection. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a method and system for correcting interviewer evaluation bias based on real-time feedback to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The interviewer evaluation bias correction method based on real-time feedback includes the following steps:

[0007] S1. Obtain the job requirements and industry attribute information of the target position, and collect the interactive text data between the interviewer and the candidate during the interview process;

[0008] S2. Extract term associations from the industry knowledge base based on industry attribute information, perform dynamic semantic analysis on fuzzy terms in job requirements, and generate industry weights and term thresholds.

[0009] S3. Extract the candidate's answer content regarding the fuzzy term from the interactive text data, analyze the correlation between the answer content and the term threshold, and generate correlation analysis results;

[0010] S4: Real-time identification of the interviewer's actual usage context of fuzzy terms and calculation of the initial deviation based on industry weights;

[0011] S5. Correcting the initial deviation based on the correlation analysis result, and marking it as a term deviation signal when the corrected deviation exceeds the term threshold;

[0012] S6. Identify the interviewer's repeated questioning behavior associated with the terminology bias signal. When the correlation analysis result meets the terminology threshold and repeated questioning behavior exists, generate an intention bias result.

[0013] S7. Based on the term deviation signal type, the corrected deviation degree, and the intention deviation result, the correction rules of the industry knowledge base are matched to generate feedback instructions and adjust the interviewer's real-time scoring weight.

[0014] In a preferred embodiment, S1 includes:

[0015] S1a. Obtain the target position's job requirements from a recruitment platform or enterprise database. The job requirements include job responsibilities and competency requirements described in natural language.

[0016] S1b. Extract industry attribute information based on the target position's enterprise registration information or user-specified parameters. The industry attribute information includes the industry classification code and the industry standard capability definition library identifier.

[0017] S1c. During the interview process, the conversation between the interviewer and the candidate is collected through real-time speech transcription technology to generate interactive text data with timestamps.

[0018] S1d. Associate and store the job requirement text, industry attribute information, and interactive text data according to the interview session number.

[0019] In a preferred embodiment, S2 includes:

[0020] S2a, based on the industry classification code, accesses the corresponding industry characteristic dataset in the preset industry knowledge base. The industry characteristic dataset contains the mapping relationship between industry terms and job competency indicators;

[0021] S2b. Extract term association rules from the industry knowledge base based on the industry standard capability definition library identifiers. The term association rules include the core capability indicators and association strength coefficients of fuzzy terms in different industry scenarios.

[0022] S2c. Semantically deconstruct the fuzzy terms in the job requirements text and generate dynamic semantic parsing results related to industry characteristics by combining core competency indicators and correlation strength coefficients.

[0023] S2d. Set industry weights based on the correlation strength coefficients of the core capability indicators in the dynamic semantic analysis results, and dynamically calculate term thresholds based on the historical score distribution of the core capability indicators.

[0024] In a preferred embodiment, S3 includes:

[0025] S3a, locate the candidate's answer segment from the interactive text data. The candidate's answer segment is the continuous dialogue text of the candidate after the interviewer asks the fuzzy term, and the speaker is marked as the candidate;

[0026] S3b. Perform contextual semantic segmentation on candidate responses to identify the depth of technical terms and the frequency of industry case citations related to core competency indicators, generating technical dimension analysis data.

[0027] S3c. Extract the logical structure markers and technical term density from the candidate's answer fragments and generate a semantic integrity score based on the integrity assessment rules pre-set in the industry knowledge base. The integrity assessment rules include indicators of logical chain integrity and term application adaptability.

[0028] S3d. Map the technical terminology depth and case citation frequency in the technical dimension analysis data to the standard reference value of the industry knowledge base, calculate the term matching deviation, and generate a qualified association analysis result when the term matching deviation is less than the term threshold and the semantic integrity score meets the standard.

[0029] In a preferred embodiment, S4 includes:

[0030] S4a, extracting the context fragments of the interviewer's questions about fuzzy terms from the interactive text data, and identifying the actual usage context type based on the preset context classification library;

[0031] S4b, matching the weight distribution rules of the corresponding scenarios in the industry knowledge base according to the identified context type, dynamically adjusting the industry weight according to the scenario keyword coverage, and generating the context adaptation weight;

[0032] S4c, analyzing the deviation direction of the core capability indicators of fuzzy terms in the question context fragment, and calculating the classification deviation degree in combination with the context adaptation weight;

[0033] S4d. Sum the classification deviations of different context types according to the context adaptation weights to generate an initial deviation.

[0034] In a preferred embodiment, the context classification library contains semantic feature keywords of technical scenarios, management scenarios and collaboration scenarios; the core capability indicator offset direction is determined by comparing the focus of the question with the standard capability definition of the industry knowledge base.

[0035] In a preferred embodiment, S5 includes:

[0036] S5a. Verify the validity of the technical terminology depth and case citation frequency in the association analysis results. When the technical terminology depth is lower than the industry knowledge base standard level and the case citation frequency does not meet the standard, generate a correction factor;

[0037] S5b. Proportional adjustment of the initial deviation is made based on a correction factor, which is dynamically set based on the ratio of the industry weight to the deviation from the mean of historical scenarios;

[0038] S5c, performing weighted calculation on the adjusted deviation and the semantic completeness score to generate a composite deviation;

[0039] S5d. When the composite deviation exceeds the term threshold, it is marked as a term deviation signal and associated with the corresponding core capability indicator and context type.

[0040] In a preferred embodiment, S6 includes:

[0041] S6a. Based on the core capability indicators and context types associated with term deviation signals, the repeated question tolerance threshold for the corresponding indicators is extracted from the industry knowledge base. The repeated question tolerance threshold is dynamically set based on the deviation risk coefficient of industry weights and historical question behavior statistics;

[0042] S6b. Count the number of questions the interviewer asks about the same core competency indicator and the interval between questions from the interactive text data. If the number of questions exceeds the tolerance threshold for repeated questions and the interval between questions is less than the reasonable interval preset in the industry knowledge base, mark it as repeated questioning behavior.

[0043] S6c. Verify whether the technical terminology depth and semantic completeness scores in the association analysis results meet the terminology threshold. If both the terminology threshold and repeated questioning are met, generate an intent deviation result.

[0044] S6d. Bind the intention deviation results with the industry weights and context types associated with the term deviation signals to generate structured deviation records.

[0045] In a preferred embodiment, S7 includes:

[0046] S7a. Matching correction rules from the industry knowledge base based on the term deviation signal type and the associated context type. The correction rules include a weight adjustment ratio and a feedback trigger condition. The feedback trigger condition is set based on the difference between the number of repeated questions in the intent deviation result and the deviation tolerance threshold.

[0047] S7b, generating a dynamic adjustment coefficient based on the difference ratio between the corrected deviation degree and the deviation tolerance threshold in the intention deviation result, and calculating a real-time score weight correction value based on the weight adjustment ratio;

[0048] S7c. Determine whether to generate a real-time feedback instruction based on the feedback triggering condition. When the difference in the number of repeated questions exceeds the intervention threshold preset by the industry knowledge base, trigger a feedback instruction including a weight correction value and a deviation type prompt.

[0049] S7d. Add the real-time score weight correction value to the industry weight to generate an updated score weight and store it in the interview assessment decision record.

[0050] In another aspect, the present invention provides an interviewer evaluation bias correction system based on real-time feedback, comprising:

[0051] Data collection module: obtains the job requirements and industry attribute information of the target position, and collects the interactive text data between the interviewer and the candidate during the interview process;

[0052] Semantic parsing module: extracts term associations from the industry knowledge base based on industry attribute information, performs dynamic semantic parsing on fuzzy terms in job requirements, and generates industry weights and term thresholds;

[0053] Relevance Analysis Module: This module extracts candidates' responses to fuzzy terms from interactive text data, analyzes the correlation between the responses and term thresholds, and generates relevance analysis results.

[0054] Deviation calculation module: This module identifies the actual context in which the interviewer uses fuzzy terms in real time and calculates the initial deviation based on industry weights.

[0055] Deviation marking module: It corrects the initial deviation based on the correlation analysis results and marks it as a term deviation signal when the corrected deviation exceeds the term threshold;

[0056] Behavior recognition module: Identifies the interviewer's repeated questioning behavior associated with terminology deviation signals. When the correlation analysis results meet the terminology threshold and repeated questioning behavior exists, an intention deviation result is generated;

[0057] Feedback generation module: Based on the term deviation signal type, the corrected deviation degree and the intention deviation result, the correction rules of the industry knowledge base are matched to generate feedback instructions and adjust the interviewer's real-time scoring weight.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] 1. By building an evaluation and correction mechanism that integrates dynamic semantic analysis with real-time feedback, the objectivity and industry adaptability of interview evaluations are effectively improved, deeply embedding industry characteristics into the entire process of terminology analysis and evaluation monitoring: The terminology association rules based on the industry knowledge base dynamically analyze ambiguous terms in job requirements, and combine industry weights and terminology thresholds to generate multi-dimensional evaluation indicators, which can accurately reflect the actual requirements of different industries for the same competency item. Through real-time interactive text context recognition and correlation verification of candidate answers, potential deviations in interviewers' question focus and terminology usage habits can be dynamically captured. Term deviation signals can be promptly flagged based on composite deviation calculations to prevent implicit evaluation bias caused by standard misalignment from continuously influencing decision-making.

[0060] 2. The synergy between intention bias identification and dynamic score adjustment significantly enhances the self-correction capability of the assessment process. While monitoring interviewers' repeated questions, context deviations, and other behaviors in real time, the system combines the candidate's answer compliance status with the correction rules of the industry knowledge base to generate targeted feedback instructions and dynamically adjust the score weights to ensure that the assessment criteria are always aligned with the actual needs of the position. This two-way verification mechanism not only solves the semantic distortion problem caused by traditional one-way parsing, but also enhances the adaptability of the assessment model to complex scenarios through the dynamic adaptation of industry weights and real-time data, thereby achieving a better balance between accuracy and fairness in talent selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A flowchart of the interviewer evaluation bias correction method based on real-time feedback according to the present invention;

[0062] Figure 2 Schematic diagram of the structure of the interviewer evaluation bias correction system based on real-time feedback of the present invention. DETAILED DESCRIPTION

[0063] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0064] Example 1: Figure 1 The present invention provides a method for correcting bias in interviewer evaluation based on real-time feedback, which includes the following steps:

[0065] S1. Obtain the job requirements and industry attribute information of the target position, and collect the interactive text data between the interviewer and the candidate during the interview process;

[0066] S2. Extract term associations from the industry knowledge base based on industry attribute information, perform dynamic semantic analysis on fuzzy terms in job requirements, and generate industry weights and term thresholds.

[0067] S3. Extract the candidate's answer content regarding the fuzzy term from the interactive text data, analyze the correlation between the answer content and the term threshold, and generate correlation analysis results;

[0068] S4: Real-time identification of the interviewer's actual usage context of fuzzy terms and calculation of the initial deviation based on industry weights;

[0069] S5. Correcting the initial deviation based on the correlation analysis result, and marking it as a term deviation signal when the corrected deviation exceeds the term threshold;

[0070] S6. Identify the interviewer's repeated questioning behavior associated with the terminology bias signal. When the correlation analysis result meets the terminology threshold and repeated questioning behavior exists, generate an intention bias result.

[0071] S7. Based on the term deviation signal type, the corrected deviation degree, and the intention deviation result, the correction rules of the industry knowledge base are matched to generate feedback instructions and adjust the interviewer's real-time scoring weight.

[0072] S1. Obtain the job requirements and industry attribute information for the target position, and collect the interactive text data between the interviewer and the candidate during the interview process, including:

[0073] S1a. Obtain the target position's job requirements from a recruitment platform or enterprise database. The job requirements include job responsibilities and competency requirements described in natural language.

[0074] S1b. Extract industry attribute information based on the target position's enterprise registration information or user-specified parameters. The industry attribute information includes the industry classification code and the industry standard capability definition library identifier.

[0075] S1c. During the interview process, the conversation between the interviewer and the candidate is collected through real-time speech transcription technology to generate interactive text data with timestamps.

[0076] S1d. Associate and store the job requirement text, industry attribute information, and interactive text data according to the interview session number.

[0077] In step S1a, the job requirements for the target position are obtained by accessing the target position's job requirements through a public interface on a recruitment platform or enterprise database. The job requirements contain natural language descriptions of job responsibilities and competency requirements. For example, for a "Java Development Engineer" position, the job requirements might include natural language descriptions of competency requirements such as "proficiency in the Spring Framework" and "experience in distributed system design." The job requirements obtained in step S1a serve as input for dynamic semantic parsing in subsequent steps.

[0078] In step S1b, industry attribute information is extracted based on the industry classification code registered with the industry administration department by the company to which the target position belongs, or based on user-specified industry classification parameters in the recruitment system. Industry classification codes use the four-digit code from the national standard "National Economic Industry Classification," for example, "Software Development" corresponds to the industry classification code "6513." Industry standard capability definition library identifiers are used to associate capability definition data with pre-built industry knowledge bases, such as the "Internet Industry Capability Definition Library" identifier "IT_SKILL_LIB_2023."

[0079] In step S1c, the interactive text data is generated as follows: During the interview, the conversation between the interviewer and the candidate is converted into text data using real-time speech transcription technology, which is implemented using a speech recognition engine based on acoustic and language models. The converted text data is timestamped according to the order of the conversation. For example, the start time of each conversation is marked as "2023-08-20 14:30:05." The timestamps are used in subsequent time series analysis to identify the actual context in which the interviewer used ambiguous terms.

[0080] In step S1d, the specific method for data association storage is to assign a unique interview session number to each interview, such as "INTERVIEW_20230820_001", and associate and store the position requirement text obtained in step S1a, the industry attribute information extracted in step S1b, and the interactive text data generated in step S1c according to the interview session number. During storage, the position requirement text and industry attribute information are associated in the form of key-value pairs. For example, the key "POSITION_REQ" corresponds to the position requirement text content, and the key "INDUSTRY_ATTR" corresponds to the industry classification code and industry standard capability definition library identifier. The interactive text data is stored in timestamp order as a structured data table, which contains the fields "timestamp", "speaker identity", and "conversation content".

[0081] S2. Extract term associations from the industry knowledge base based on industry attribute information, perform dynamic semantic analysis on fuzzy terms in job requirements, and generate industry weights and term thresholds, including:

[0082] S2a, based on the industry classification code, accesses the corresponding industry characteristic dataset in the preset industry knowledge base. The industry characteristic dataset contains the mapping relationship between industry terms and job competency indicators;

[0083] S2b. Extract term association rules from the industry knowledge base based on the industry standard capability definition library identifiers. The term association rules include the core capability indicators and association strength coefficients of fuzzy terms in different industry scenarios.

[0084] S2c. Semantically deconstruct the fuzzy terms in the job requirements text and generate dynamic semantic parsing results related to industry characteristics by combining core competency indicators and correlation strength coefficients.

[0085] S2d. Set industry weights based on the correlation strength coefficients of the core capability indicators in the dynamic semantic analysis results, and dynamically calculate term thresholds based on the historical score distribution of the core capability indicators.

[0086] In step S2a, the specific method of accessing the corresponding industry characteristic data set in the preset industry knowledge base based on the industry classification code is: according to the industry classification code extracted in step S1b, the corresponding industry characteristic data set is matched from the industry knowledge base. The industry classification code adopts the four-digit code in the national standard "National Economic Industry Classification", and its industry characteristic data set contains the mapping relationship between the core terms of the industry and job competency indicators. In the industry characteristic data set of the software development industry, "code ability" is mapped to job competency indicators such as "code review pass rate" and "unit test coverage". The industry characteristic data set is constructed through the job competency white paper issued by the industry standardization organization and stored in the industry knowledge base.

[0087] In step S2b, the specific method for extracting term association rules is: according to the industry standard capability definition library identifier in step S1b, the corresponding capability definition library is loaded from the industry knowledge base, and the term association rules therein are parsed. The term association rules are stored in the form of a structured data table, and each rule contains fuzzy terms, core capability indicators and association strength coefficients. For example, for the fuzzy term "communication ability", in the term association rules of the sales industry, the core capability indicators are "customer demand conversion rate" and "cross-departmental collaboration frequency", and the association strength coefficients are 0.8 and 0.6 respectively, where the association strength coefficient is pre-set by industry experts based on the importance assessment of job capabilities. In the term association rules of the technology industry, the core capability indicators are "clarity of technical solution explanation" and "standardization of document writing", and the association strength coefficients are 0.7 and 0.5 respectively.

[0088] In step S2c, the dynamic semantic parsing process involves semantically deconstructing the fuzzy terms in the job description obtained in step S1a. This deconstruction is based on a pre-built semantic rule base within the industry knowledge base, developed by industry experts based on the usage patterns of terms in different scenarios. For example, if the fuzzy term "leadership" appears in the job description and the industry attribute is "internet industry," the semantic rule base deconstructs it into "participation in technical decision-making" (with a correlation strength coefficient of 0.75) and "number of team project leadership" (with a correlation strength coefficient of 0.7). If the industry attribute is "financial industry," the semantic rule base deconstructs it into "risk decision-making accuracy" (with a correlation strength coefficient of 0.85) and "team resource coordination efficiency" (with a correlation strength coefficient of 0.65). The dynamic semantic parsing results are stored as key-value pairs, with the key being the fuzzy term and the value being a combination of the core competency indicator and its correlation strength coefficient.

[0089] In step S2d, the industry weight and term threshold are generated by setting the industry weight based on the absolute value of the correlation strength coefficient of the core capability indicator in the dynamic semantic analysis results of step S2c. For example, if the correlation strength coefficient of the core capability indicator "Technical Decision-Making Participation" is 0.75, its industry weight is 0.75.

[0090] The terminology threshold is dynamically calculated by taking the upper quartile of historical scoring data for the same core competency indicator from the industry standard competency definition library as the threshold. For example, if the upper quartile of historical scoring data for "Technical Decision Participation" in the internet industry is 85, the terminology threshold is set at 85. This historical scoring distribution data is extracted from the company's historical interview database and stored by industry classification code and core competency indicator.

[0091] S3. Extract the candidate's answers to the fuzzy terms from the interactive text data, analyze the correlation between the answers and the term threshold, and generate correlation analysis results, including:

[0092] S3a, locate the candidate's answer segment from the interactive text data. The candidate's answer segment is the continuous dialogue text of the candidate after the interviewer asks the fuzzy term, and the speaker is marked as the candidate;

[0093] S3b. Perform contextual semantic segmentation on candidate responses to identify the depth of technical terms and the frequency of industry case citations related to core competency indicators, generating technical dimension analysis data.

[0094] S3c. Extract the logical structure markers and technical term density from the candidate's answer fragments and generate a semantic integrity score based on the integrity assessment rules pre-set in the industry knowledge base. The integrity assessment rules include indicators of logical chain integrity and term application adaptability.

[0095] S3d. Map the technical terminology depth and case citation frequency in the technical dimension analysis data to the standard reference value of the industry knowledge base, calculate the term matching deviation, and generate a qualified association analysis result when the term matching deviation is less than the term threshold and the semantic integrity score meets the standard.

[0096] In step S3a, the specific method for locating candidate response fragments is to filter out conversations in which the speaker's identity is marked as the candidate from the interactive text data generated in step S1c, and then intercept the continuous conversation text after the interviewer asks the ambiguous term as the candidate's response fragment. For example, when the interviewer asks, "Please describe how you demonstrated leadership in the project," the text fragment in which the candidate responds, "I led a cross-departmental collaborative project and coordinated five teams to complete the microservices architecture upgrade," is marked as the response fragment. Speaker identity tagging is achieved by extracting the "Speaker Identity" field from the structured storage table of the interactive text data. This field is automatically annotated during speech transcription using existing voiceprint recognition technology (such as Azure Speaker Recognition), eliminating the need for custom algorithms.

[0097] In step S3b, the contextual semantic segmentation and technical dimension analysis data are generated by dividing the candidate's answer into multiple semantic segments based on natural paragraphs or semantic transitions (such as "first" and "secondly"). For each semantic segment, the depth of the technical terminology contained within is identified. This depth is determined based on a pre-defined hierarchy of technical terms in the industry knowledge base. This hierarchy is based on skill classifications in industry standard documents (such as the "Software Engineer Competency Standards"). For example, "microservice architecture" falls under the "architecture design" category and has a depth of 3. The frequency of industry case citations is also counted. For example, a mention of the "Financial Industry Risk Control System Migration Case" in the answer segment counts as one case citation. The case citation standard is based on a pre-defined list of industry cases in the industry knowledge base.

[0098] In step S3c, the semantic completeness score is calculated by extracting logical structure markers (such as "therefore" and "in summary") from the candidate's answer fragment, counting their number, and calculating the logical chain completeness score. For example, if the number of logical connectives in an answer fragment containing a complete "problem description-solution-implementation effect" logical chain is ≥3, the logical chain completeness score is 90. Technical term density is calculated as the ratio of the number of technical terms to the total number of words in the answer fragment. For example, if the proportion of technical terms exceeds 15%, the term density score is A. Combined with the completeness assessment rules developed by industry experts in the industry knowledge base (such as a 60% weight for logical chain completeness and a 40% weight for term application adaptability), a comprehensive semantic completeness score is generated.

[0099] The adaptability of terminology applications is determined by comparing the matching ratio of the technical terms used by the candidate with the mainstream technical term list of the current position in the industry knowledge base. For example, the matching ratio of "Spring Cloud" in the candidate's answer to "Kubernetes" in the mainstream term list is 70%.

[0100] In step S3d, the terminology mismatch calculation and correlation analysis results are generated by mapping the technical terminology depth and case citation frequency obtained in step S3b to the standard reference values for the same core competency indicator in the industry knowledge base. The standard reference values are based on historical interview data. For example, the standard technical terminology depth for "architecture design capability" in the internet industry is level 3 and the standard case citation frequency is 2. If the candidate's answer has a technical terminology depth of 2 and a case citation frequency of 1, the terminology mismatch calculation is (1 / 3 + 0 / 2) × 100% = 33%.

[0101] If the term matching deviation is less than the term threshold generated in step S2d (for example, the internet industry threshold is 40%) and the semantic integrity score in step S3c meets the standard (for example, ≥80 points), the result is marked as a qualified association analysis result. The qualified association analysis result is stored as structured data, including the technical term depth deviation value, case citation frequency deviation value, and semantic integrity score.

[0102] S4. Real-time identification of the interviewer's actual usage context of ambiguous terms and calculation of initial deviation based on industry weights, including:

[0103] S4a. Extract context fragments of ambiguous terms asked by the interviewer from the interactive text data and identify the actual usage context type based on the pre-set context classification library. The context classification library contains semantic feature keywords of technical scenarios, management scenarios, and collaboration scenarios.

[0104] S4b, matching the weight distribution rules of the corresponding scenarios in the industry knowledge base according to the identified context type, dynamically adjusting the industry weight according to the scenario keyword coverage, and generating the context adaptation weight;

[0105] S4c, analyze the core competency indicator deviation direction of fuzzy terms in the question context fragment, calculate the classification deviation degree based on the context adaptation weight, and determine the core competency indicator deviation direction by comparing the question focus with the standard competency definition in the industry knowledge base;

[0106] S4d. Sum the classification deviations of different context types according to the context adaptation weights to generate an initial deviation.

[0107] In step S4a, the process of real-time identification of the interviewer's actual usage context of fuzzy terms is as follows: the conversation content before and after the question position of the fuzzy term is extracted from the interactive text data generated in step S1c as context fragments. For example, the two sentences after the interviewer asks "Please explain how you demonstrate leadership in the project" "This project involves multiple teams, and I am responsible for technical decisions..." are intercepted as context fragments.

[0108] Context types are identified for context fragments based on a pre-built context classification library. This library, constructed by industry experts based on job competency scenarios, contains semantically significant keywords for technical, management, and collaboration scenarios. For example, keywords for technical scenarios include "architecture design" and "code performance," keywords for management scenarios include "team coordination" and "project planning," and keywords for collaboration scenarios include "cross-departmental communication" and "resource integration." Context type identification is performed by counting the number of matches between the context fragment and the keywords for each scenario. The scenario with the highest number of matches is determined as the current context type.

[0109] In step S4b, the dynamic adjustment of industry weights is achieved by reading the corresponding scenario weight adjustment coefficient from the industry knowledge base based on the context type determined in step S4a. This coefficient is determined by industry experts based on an analysis of the importance of the scenarios in historical interview data. For example, the adjustment coefficient for the technical scenario is +20%, for the management scenario is -15%, and for the collaboration scenario is -5%. Based on the original industry weight (the "Technical Decision-Making Participation" weight of 0.75 from step S2d), if the current scenario is technical, the adjusted context adaptation weight is 0.75 × 1.2 = 0.9; for the management scenario, it is adjusted to 0.75 × 0.85 = 0.64. The adjustment coefficient is set based on statistical correlations between ability indicators and hiring success across different scenarios within the same industry.

[0110] In step S4c, the specific method for determining the direction of core competency indicator deviation and calculating classification deviation is to compare the core competency indicators mentioned in the interviewer's question context with the standard competency definitions in the industry knowledge base. The standard competency definitions are derived from the term association rules in step S2b. For example, the standard definition of "architecture design capabilities" in the internet industry is "technology selection" (weighted 40%), "performance optimization" (weighted 35%), and "coding standards" (weighted 25%). If the interviewer focuses 80% of the question on "technology selection," the deviation is excessively focused on "technology selection." The deviation is calculated as the difference in target proportion (80% - 40% = 40%) multiplied by the context adaptation weight (0.9), resulting in a classification deviation of 36%.

[0111] In step S4d, the logic for generating the initial deviation is to weight the deviations of the classifications in multiple contexts by accumulating them according to the context adaptation weights. For example, in a particular interview, the deviation of the technical scenario classification is 36% (context adaptation weight 0.9) and the deviation of the management scenario classification is 10% (weight 0.64). If the context distribution is 70% for technical scenarios and 30% for management scenarios, then the initial deviation = 36% × 0.7 + 10% × 0.3 = 27%. The weighting is derived from the context adaptation weights generated in step S4b, ensuring that the deviations of higher-weighted scenarios have a greater impact on the final results.

[0112] S5. Correct the initial deviation based on the correlation analysis result. When the corrected deviation exceeds the term threshold, it is marked as a term deviation signal, including:

[0113] S5a. Verify the validity of the technical terminology depth and case citation frequency in the association analysis results. When the technical terminology depth is lower than the industry knowledge base standard level and the case citation frequency does not meet the standard, generate a correction factor;

[0114] S5b. Proportional adjustment of the initial deviation is made based on a correction factor, which is dynamically set based on the ratio of the industry weight to the deviation from the mean of historical scenarios;

[0115] S5c, performing weighted calculation on the adjusted deviation and the semantic completeness score to generate a composite deviation;

[0116] S5d. When the composite deviation exceeds the term threshold, it is marked as a term deviation signal and associated with the corresponding core capability indicator and context type.

[0117] In step S5a, the specific process for verifying the validity of the association analysis results is to extract the technical terminology depth and case citation frequency from the technical dimension analysis data generated in step S3b and compare them with the standard levels pre-set in the industry knowledge base. The standard levels are defined by the job competency level specifications published by industry standardization organizations. For example, the "architecture design capability" in the internet industry requires technical terminology depth to reach level 3 (covering sub-items such as architectural patterns and performance optimization) and case citation frequency to be ≥ 2. If the candidate's answer is level 2 in technical terminology depth and 1 in case citation frequency, it is considered to be below the standard, and a correction factor is generated. The correction factor is calculated as the weighted sum of the technical terminology depth deviation value (level 1) and the case citation frequency deviation value (1). For example, the correction factor = (1 / 3 + 1 / 2) × 0.5 = 0.42, where the weight of 0.5 is determined by the statistical results of the deviation from the mean of historical scenarios in the industry knowledge base.

[0118] In step S5b, the dynamic setting of the correction factor and the adjustment of the initial deviation are implemented by calculating the correction factor based on the ratio of the industry weight generated in step S2d to the historical scenario deviation mean. The historical scenario deviation mean is extracted from the industry knowledge base. For example, the average deviation of technical terminology depth in historical interviews for "architecture design capabilities" in the internet industry is 0.3 levels, and the average deviation of case citation frequency is 0.5. Therefore, the correction factor = industry weight (0.75) × (deviation value / historical mean) = 0.75 × (0.42 / 0.4) = 0.79. Multiplying the correction factor by the initial deviation (12.5% from step S4d) yields the adjusted deviation = 12.5% × 0.79 ≈ 9.88%.

[0119] In step S5c, the composite deviation is calculated by weighting the deviation adjusted in step S5b with the semantic integrity score from step S3c. The weighting factor is determined by the integrity weighting rules pre-set in the industry knowledge base. For example, if the integrity score is ≥ 80, the deviation weight is 70% and the integrity weight is 30%. Assuming the adjusted deviation is 9.88% and the semantic integrity score is 85, the composite deviation is 9.88% × 0.7 + (100 - 85)% × 0.3 = 6.92% + 4.5% = 11.42%. The integrity weighting rules are based on the assessment of the impact of answer quality by industry experts.

[0120] In step S5d, the logic for marking term deviation signals is to compare the composite deviation with the term threshold generated in step S2d. For example, the term threshold for "architecture design capability" in the internet industry is 10%. If the composite deviation is 11.42% and exceeds 10%, it is marked as a term deviation signal. This marking is associated with the corresponding core capability indicator (such as "architecture design capability") and the context type (such as technical scenario), and stored in a structured data format, such as the fields "Deviation Signal Type: Technical Scenario - Architecture Design Capability" and "Deviation Value: 11.42%."

[0121] S6. Identify the interviewer's repeated questioning behavior associated with terminology deviation signals. When the correlation analysis results meet the terminology threshold and repeated questioning behavior exists, generate an intention deviation result, including:

[0122] S6a. Based on the core capability indicators and context types associated with term deviation signals, the repeated question tolerance threshold for the corresponding indicators is extracted from the industry knowledge base. The repeated question tolerance threshold is dynamically set based on the deviation risk coefficient of industry weights and historical question behavior statistics;

[0123] S6b. Count the number of questions the interviewer asks about the same core competency indicator and the interval between questions from the interactive text data. If the number of questions exceeds the tolerance threshold for repeated questions and the interval between questions is less than the reasonable interval preset in the industry knowledge base, mark it as repeated questioning behavior.

[0124] S6c. Verify whether the technical terminology depth and semantic completeness scores in the association analysis results meet the terminology threshold. If both the terminology threshold and repeated questioning are met, generate an intent deviation result.

[0125] S6d. Bind the intention deviation results with the industry weights and context types associated with the term deviation signals to generate structured deviation records.

[0126] In step S6a, the repeated question tolerance threshold is dynamically set by extracting historical questioning behavior statistics for the core competency indicator and context type associated with the term deviation signal identified in step S5d from the industry knowledge base. This historical questioning behavior statistics include the average frequency of questions asked about the same competency indicator by interviewers in previous interviews for the same position, as well as a deviation risk factor. The deviation risk factor is calculated based on the percentage of candidate scores that decrease due to repeated questions. For example, the average historical question frequency for "architecture design capability" in the internet industry is 2, and the deviation risk factor is 0.6 (indicating that 60% of candidates' scores decrease due to repeated questions).

[0127] The tolerance threshold is calculated as follows: Tolerance Threshold = Historical Frequency Average × (1 - Deviation Risk Factor) × Industry Weight. Assuming an industry weight of 0.75, the tolerance threshold = 2 × (1 - 0.6) × 0.75 = 0.6 times, rounded up to 1. The deviation risk factor is generated by a pre-configured historical data analysis model in the industry knowledge base. This model uses linear regression to analyze the relationship between question frequency and score changes in historical interview data.

[0128] In step S6b, the identification logic of repeated questioning behavior is: extract the interviewer's question records on the same core competency indicator from the interactive text data generated in step S1c, and count the number of questions asked and the time interval between adjacent questions. The reasonable interval preset in the industry knowledge base is set according to the position type. For example, the reasonable interval for technical positions is 15 minutes, and for management positions it is 10 minutes. If the interviewer asks about "architecture design ability" 3 times within 10 minutes (exceeding the tolerance threshold 1 time) and the interval time is less than the reasonable interval, it is marked as repeated questioning behavior. The question record is realized by matching the sentences in the interactive text data where the speaker is the interviewer and contains the core competency indicator keywords.

[0129] In step S6c, the relevance analysis results are verified by checking whether the technical terminology depth and semantic completeness scores in the qualified relevance analysis results generated in step S3d meet the terminology threshold. For example, the technical terminology depth must meet the industry knowledge base standard level 3, and the semantic completeness score must be ≥80. If the candidate's answer meets both the terminology threshold and contains repeated questions as flagged in step S6b, an intent deviation result is generated. The intent deviation result includes the number of repeated questions, the deviation from the tolerance threshold, and the associated context type (e.g., technical scenario).

[0130] In step S6d, the structured deviation record is generated by binding the intent deviation result with the industry weight (from step S2d) and context type (from step S4a) associated with the term deviation signal, forming structured data containing the fields "core capability indicator," "industry weight," "context type," "number of repeated questions," and "deviation tolerance threshold." For example, the structured data record might be "core capability indicator: architecture design capability; industry weight: 0.75; context type: technical scenario; number of repeated questions: 3; deviation tolerance threshold: 2." This structured data is stored in the historical deviation database of the industry knowledge base and used for correction rule matching and feedback instruction generation in the subsequent step S7.

[0131] S7. Based on the term deviation signal type, the corrected deviation degree, and the intention deviation result, the correction rules of the industry knowledge base are matched to generate feedback instructions and adjust the interviewer's real-time scoring weight, including:

[0132] S7a. Matching correction rules from the industry knowledge base based on the term deviation signal type and the associated context type. The correction rules include a weight adjustment ratio and a feedback trigger condition. The feedback trigger condition is set based on the difference between the number of repeated questions in the intent deviation result and the deviation tolerance threshold.

[0133] S7b, generating a dynamic adjustment coefficient based on the difference ratio between the corrected deviation degree and the deviation tolerance threshold in the intention deviation result, and calculating a real-time score weight correction value based on the weight adjustment ratio;

[0134] S7c. Determine whether to generate a real-time feedback instruction based on the feedback triggering condition. When the difference in the number of repeated questions exceeds the intervention threshold preset by the industry knowledge base, trigger a feedback instruction including a weight correction value and a deviation type prompt.

[0135] S7d. Add the real-time score weight correction value to the industry weight to generate an updated score weight and store it in the interview assessment decision record.

[0136] In step S7a, the correction rule matching method is to retrieve preset correction rules from the industry knowledge base based on the term deviation signal type (such as "technical scenario-architecture design capability deviation") and the associated context type (such as technical scenario). Correction rules are formulated by industry experts based on the degree to which different deviation types in historical interview data affect scoring accuracy. For example, the correction rule for the deviation of "architecture design capability" in the technical scenario is: weight adjustment ratio -10% (i.e., the weight of this indicator is reduced by 10%), and the feedback trigger condition is set to "difference in the number of repeated questions ≥ 2 times." The difference in the number of repeated questions is calculated by subtracting the tolerance threshold in step S6a from the actual number of questions counted in step S6b. For example, actual number of questions asked - tolerance threshold of 1 time = difference of 2 times.

[0137] In step S7b, the logic for generating the real-time score weight correction is to generate a dynamic adjustment coefficient based on the proportional relationship between the corrected deviation (the 11.42% composite deviation from step S5c) and the deviation tolerance threshold difference (e.g., 2) in the intention deviation result. The dynamic adjustment coefficient = corrected deviation / (deviation tolerance threshold difference × preset unit impact factor). The unit impact factor is set by industry experts based on the impact of each deviation on the score in historical data (e.g., 5% for the internet industry).

[0138] In step S7c, the feedback triggering condition is determined based on the intervention threshold preset in the industry knowledge base, which is set based on the position type. For example, the intervention threshold for technical positions is a difference in the number of repeated questions ≥ 1.5. If the difference in step S6b is 2 (≥ 1.5), a feedback instruction is triggered. The feedback instruction contains the weight correction value (0.6645) and the deviation type (e.g., "Technical Scenario - Architecture Design Ability Weight Too High"). The instruction format is "Adjustment Indicator: Architecture Design Ability; Corrected Weight: 0.6645; Deviation Reason: Repeated Questioning Exceeds Threshold."

[0139] In step S7d, the scoring weight is updated by replacing the original industry weight (0.75) generated in step S2d with the real-time scoring weight correction value (0.6645), generating an updated scoring weight of 0.6645. The updated weight is associated with the interview session number, core competency indicator, and deviation type and stored in the interview assessment decision record. The record storage format includes the fields "Session Number: INTERVIEW_20230820_001; Indicator Name: Architecture Design Ability; Original Weight: 0.75; Revised Weight: 0.6645; Deviation Type: Technical Scenario - Repeated Question."

[0140] Example 2: Figure 2 A schematic diagram of the structure of the interviewer evaluation bias correction system based on real-time feedback of the present invention is provided. The interviewer evaluation bias correction system based on real-time feedback includes:

[0141] Data collection module: obtains the job requirements and industry attribute information of the target position, and collects the interactive text data between the interviewer and the candidate during the interview process;

[0142] Semantic parsing module: extracts term associations from the industry knowledge base based on industry attribute information, performs dynamic semantic parsing on fuzzy terms in job requirements, and generates industry weights and term thresholds;

[0143] Relevance Analysis Module: This module extracts candidates' responses to fuzzy terms from interactive text data, analyzes the correlation between the responses and term thresholds, and generates relevance analysis results.

[0144] Deviation calculation module: This module identifies the actual context in which the interviewer uses fuzzy terms in real time and calculates the initial deviation based on industry weights.

[0145] Deviation marking module: It corrects the initial deviation based on the correlation analysis results and marks it as a term deviation signal when the corrected deviation exceeds the term threshold;

[0146] Behavior recognition module: Identifies the interviewer's repeated questioning behavior associated with terminology deviation signals. When the correlation analysis results meet the terminology threshold and repeated questioning behavior exists, an intention deviation result is generated;

[0147] Feedback generation module: Based on the term deviation signal type, the corrected deviation degree and the intention deviation result, the correction rules of the industry knowledge base are matched to generate feedback instructions and adjust the interviewer's real-time scoring weight.

[0148] The above embodiments may be implemented in whole or in part through software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0149] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application of the technical solution and the invention constraints. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0150] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0151] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0152] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. Interviewer evaluation bias correction method based on real-time feedback, characterized by: The steps include: S1. Obtain the job requirements and industry attribute information of the target position, and collect the interactive text data between the interviewer and the candidate during the interview process; S2. Extract term associations from the industry knowledge base based on industry attribute information, perform dynamic semantic analysis on fuzzy terms in job requirements, and generate industry weights and term thresholds, including: S2a, based on the industry classification code, accesses the corresponding industry characteristic dataset in the preset industry knowledge base. The industry characteristic dataset contains the mapping relationship between industry terms and job competency indicators; S2b. Extract term association rules from the industry knowledge base based on the industry standard capability definition library identifiers. The term association rules include the core capability indicators and association strength coefficients of fuzzy terms in different industry scenarios. S2c. Semantically deconstruct the fuzzy terms in the job requirements text and generate dynamic semantic parsing results related to industry characteristics by combining core competency indicators and correlation strength coefficients. S2d. Set industry weights based on the correlation strength coefficients of the core capability indicators in the dynamic semantic analysis results, and dynamically calculate term thresholds based on the historical score distribution of the core capability indicators; S3. Extract the candidate's answers to the fuzzy terms from the interactive text data, analyze the correlation between the answers and the term threshold, and generate correlation analysis results, including: S3a, locate the candidate's answer segment from the interactive text data. The candidate's answer segment is the continuous dialogue text of the candidate after the interviewer asks the fuzzy term, and the speaker is marked as the candidate; S3b. Perform contextual semantic segmentation on candidate responses to identify the depth of technical terms and the frequency of industry case citations related to core competency indicators, generating technical dimension analysis data. S3c. Extract the logical structure markers and technical term density from the candidate's answer fragments and generate a semantic integrity score based on the integrity assessment rules pre-set in the industry knowledge base. The integrity assessment rules include indicators of logical chain integrity and term application adaptability. S3d. Map the technical terminology depth and industry case citation frequency in the technical dimension analysis data to the standard reference value of the industry knowledge base, calculate the term matching deviation, and generate a qualified association analysis result when the term matching deviation is less than the term threshold and the semantic integrity score meets the standard; S4: Real-time identification of the interviewer's actual usage context of fuzzy terms and calculation of the initial deviation based on industry weights; S5. Correcting the initial deviation based on the correlation analysis result, and marking it as a term deviation signal when the corrected deviation exceeds the term threshold; S6. Identify the interviewer's repeated questioning behavior associated with the terminology bias signal. When the correlation analysis result meets the terminology threshold and repeated questioning behavior exists, generate an intention bias result. S7. Based on the term deviation signal type, the corrected deviation degree, and the intention deviation result, the correction rules of the industry knowledge base are matched to generate feedback instructions and adjust the interviewer's real-time scoring weight.

2. The interviewer evaluation bias correction method based on real-time feedback according to claim 1 is characterized in that: S1 includes: S1a. Obtain the target position's job requirements from a recruitment platform or enterprise database. The job requirements include job responsibilities and competency requirements described in natural language. S1b. Extract industry attribute information based on the target position's enterprise registration information or user-specified parameters. The industry attribute information includes the industry classification code and the industry standard capability definition library identifier. S1c. During the interview process, the conversation between the interviewer and the candidate is collected through real-time speech transcription technology to generate interactive text data with timestamps. S1d. Associate and store the job requirement text, industry attribute information, and interactive text data according to the interview session number.

3. The interviewer evaluation bias correction method based on real-time feedback according to claim 1, characterized in that S4 include: S4a, extracting the context fragments of the interviewer's questions about fuzzy terms from the interactive text data, and identifying the actual usage context type based on the preset context classification library; S4b, matching the weight allocation rules of the corresponding scenarios in the industry knowledge base according to the identified actual usage context type, dynamically adjusting the industry weight according to the scenario keyword coverage, and generating the context adaptation weight; S4c, analyzing the deviation direction of the core capability indicators of fuzzy terms in the question context fragment, and calculating the classification deviation degree in combination with the context adaptation weight; S4d. Sum the classification deviations of different context types according to the context adaptation weights to generate an initial deviation.

4. The interviewer evaluation bias correction method based on real-time feedback according to claim 3 is characterized in that: The context classification library contains semantic feature keywords for technical scenarios, management scenarios, and collaboration scenarios; the deviation direction of core capability indicators is determined by comparing the focus of questions with the standard capability definitions in the industry knowledge base.

5. The interviewer evaluation bias correction method based on real-time feedback according to claim 1, characterized in that S5 include: S5a. Verify the validity of the technical terminology depth and industry case citation frequency in the association analysis results. When the technical terminology depth is lower than the standard level of the industry knowledge base and the industry case citation frequency does not meet the standard, generate a correction factor; S5b. Proportional adjustment of the initial deviation is made based on a correction factor, which is dynamically set based on the ratio of the industry weight to the deviation from the mean of historical scenarios; S5c, performing weighted calculation on the adjusted deviation and the semantic integrity score to generate a composite deviation; S5d. When the composite deviation exceeds the term threshold, it is marked as a term deviation signal and associated with the corresponding core capability indicator and context type.

6. The interviewer evaluation bias correction method based on real-time feedback according to claim 1, characterized in that S6 include: S6a. Based on the core capability indicators and context types associated with term deviation signals, the repeated question tolerance threshold for the corresponding indicators is extracted from the industry knowledge base. The repeated question tolerance threshold is dynamically set based on the deviation risk coefficient of industry weights and historical question behavior statistics; S6b. Count the number of questions the interviewer asks about the same core competency indicator and the interval between questions from the interactive text data. If the number of questions exceeds the tolerance threshold for repeated questions and the interval between questions is less than the reasonable interval preset in the industry knowledge base, mark it as repeated questioning behavior. S6c. Verify whether the technical terminology depth and semantic completeness scores in the association analysis results meet the terminology threshold. If both the terminology threshold and repeated questioning are met, generate an intent deviation result. S6d. Bind the intention deviation results with the industry weights and context types associated with the term deviation signals to generate structured deviation records.

7. The interviewer evaluation bias correction method based on real-time feedback according to claim 1, characterized in that S7 include: S7a. Matching correction rules from the industry knowledge base based on the term deviation signal type and the associated context type. The correction rules include a weight adjustment ratio and a feedback trigger condition. The feedback trigger condition is set based on the difference between the number of repeated questions in the intent deviation result and the deviation tolerance threshold. S7b, generating a dynamic adjustment coefficient based on the difference ratio between the corrected deviation degree and the deviation tolerance threshold in the intention deviation result, and calculating a real-time score weight correction value based on the weight adjustment ratio; S7c. Determine whether to generate a real-time feedback instruction based on the feedback trigger condition. When the difference in the number of repeated questions exceeds the intervention threshold preset in the industry knowledge base, trigger a feedback instruction including a weight correction value and a deviation type prompt. S7d. Add the real-time score weight correction value to the industry weight to generate an updated score weight and store it in the interview assessment decision record.

8. A system for correcting bias in interviewer evaluation based on real-time feedback, for implementing the method for correcting bias in interviewer evaluation based on real-time feedback according to any one of claims 1 to 7, characterized in that: include: Data collection module: obtains the job requirements and industry attribute information of the target position, and collects the interactive text data between the interviewer and the candidate during the interview process; Semantic parsing module: extracts term associations from the industry knowledge base based on industry attribute information, performs dynamic semantic parsing on fuzzy terms in job requirements, and generates industry weights and term thresholds; Relevance Analysis Module: This module extracts candidates' responses to fuzzy terms from interactive text data, analyzes the correlation between the responses and term thresholds, and generates relevance analysis results. Deviation calculation module: This module identifies the actual context in which the interviewer uses fuzzy terms in real time and calculates the initial deviation based on industry weights. Deviation marking module: It corrects the initial deviation based on the correlation analysis results and marks it as a term deviation signal when the corrected deviation exceeds the term threshold; Behavior recognition module: Identifies the interviewer's repeated questioning behavior associated with terminology deviation signals. When the correlation analysis results meet the terminology threshold and repeated questioning behavior exists, an intention deviation result is generated; Feedback generation module: Based on the term deviation signal type, the corrected deviation degree and the intention deviation result, the correction rules of the industry knowledge base are matched to generate feedback instructions and adjust the interviewer's real-time scoring weight.

Images