A method for intelligently evaluating professional ability and post adaptation degree based on communication behavior

Abstract

The application discloses a kind of professional ability and post adaptation degree intelligent evaluation method based on communication behavior, including step one: obtaining target post description text and the original communication data of to-be-evaluated object;Step two: the data preprocessing of target post description text and original communication data;Step three: semantic embedding coding is carried out;Step four: by improved RetNet network, double time scale recursion modeling and professional ability embedding modeling are carried out;Step five: design post adaptation degree strategy and carry out matching calculation, obtain post adaptation degree score;Step six: based on post adaptation degree score, determine the adaptation degree grade of to-be-evaluated object;Step seven: the real performance evaluation grade of to-be-evaluated object is obtained, and the improved RetNet network is incrementally updated.The application improves the evaluation accuracy and evaluation stability of professional ability and post adaptation degree by improved RetNet network.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and human resource intelligent assessment technology, and in particular to an intelligent assessment method for professional competence and job suitability based on communication behavior. Background Technology

[0002] With the increasing demand for digital recruitment and intelligent human resource management, matching analysis technology that analyzes the relationship between candidate communication behavior characteristics and job competency requirements has received widespread attention. Existing job-person matching systems primarily rely on static resume information, keyword matching rules, or simple text similarity calculations for suitability assessment. However, these systems commonly suffer from the following problems in practical applications:

[0003] A candidate's true abilities during interviews, communication, or collaboration are often reflected in continuous dialogue behavior, expressive structure, and changes in the rhythm of interaction. However, traditional methods are mostly based on single text fragments or manual evaluations, failing to systematically model communication behavior sequences across time steps. This results in ability assessments being highly sensitive to occasional fluctuations in expression, exhibiting poor stability, and failing to reflect long-term ability characteristics. Job competency requirements are typically abstracted through job descriptions or historical performance standards. Existing methods often use simple vector averaging or keyword weighting to generate job profiles, lacking characterization of the structural relationships between competency dimensions and failing to create aligned models that are embedded in the same space as the candidate's competency representation. In the fit calculation stage, traditional similarity functions or linear weighted scoring methods cannot effectively distinguish the risk impact of key competency deficiencies from non-key competency differences. This results in a lack of sufficient penalty mechanisms for high-weight competency gaps in the matching results, thereby reducing the reliability and interpretability of fit determination.

[0004] Therefore, how to provide an intelligent assessment method for professional competence and job suitability based on communication behavior is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent assessment method for professional competence and job suitability based on communication behavior. This invention fully utilizes a pre-trained bidirectional Transformer text encoder and an improved RetNet network that includes dual-timescale Retention modeling to perform unified embedding modeling and competence decoupling mapping on the target job description text and the original communication data. It also achieves refined matching through a job suitability strategy, and has the advantages of high stability in competence modeling, accurate identification of key competence gaps, strong interpretability of matching results, and continuous incremental optimization of the model.

[0006] An intelligent assessment method for professional competence and job suitability based on communication behavior according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Obtain the target job description text and the original communication data between the candidate and the candidate to be evaluated;

[0008] Step 2: Perform data preprocessing on the target job description text and the original communication data to obtain structured job description text and standard communication text sequence;

[0009] Step 3: Semantically embed and encode the structured job description text and standard communication text sequence to obtain a set of job competency embedding vectors, a set of job dimension weights, and a sequence of communication behavior feature vectors;

[0010] Step 4: Input the communication behavior feature vector sequence into the improved RetNet network, perform dual-timescale recursive modeling and professional ability embedding modeling, and generate a set of professional ability embedding vectors; the improved RetNet network includes an embedding mapping layer, a backbone encoding network and an ability decoupling mapping module; wherein, the backbone encoding network introduces a low-rank projection structure, a shared basis structure and a dual-timescale Retention structure;

[0011] Step 5: Design a job fit strategy, match and calculate the set of professional ability embedding vectors, the set of job ability embedding vectors, and the set of job dimension weights to obtain a job fit score;

[0012] Step Six: Determine the suitability level of the candidate based on the job suitability score;

[0013] Step 7: Obtain the true performance rating of the object to be evaluated and incrementally update the improved RetNet network.

[0014] Optionally, the target job description text includes job responsibility items, job qualification items, job skill requirements, job competency requirements, and job performance evaluation standards; the original communication data includes voice communication data and text communication data.

[0015] Optionally, step two specifically includes:

[0016] The job description text and text communication data are segmented according to punctuation marks, and web page tags, special symbols and semantically meaningless characters are deleted to generate a structured job description text and structured communication text sequence.

[0017] The voice communication data is denoised, silence is detected, and speech is recognized and converted, and a speech-to-text sequence is generated in chronological order.

[0018] The speech-to-text sequence and the structured communication text sequence are aligned according to a set time step and merged at the text level to obtain a standard communication text sequence.

[0019] Optionally, step three specifically includes:

[0020] The structured job description text is segmented to generate a job term sequence. The job term sequence is then input into a pre-trained bidirectional Transformer text encoder to obtain a job term semantic vector sequence.

[0021] Initialize the capability dimension matrix, query linear mapping matrix, key linear mapping matrix, and value linear mapping matrix;

[0022] The sequence of semantic vectors of job terms is stacked according to time steps to form a job semantic matrix;

[0023] The capability dimension matrix is ​​transformed into a capability query matrix through a query linear mapping matrix, and the job semantic matrix is ​​transformed into a job key matrix and a job value matrix through a key linear mapping matrix and a value linear mapping matrix, respectively.

[0024] Multiply the capability query matrix by the transpose of the job key matrix, divide by the square root of the embedding dimension of the job term semantic vector, and then normalize using the Softmax function to generate the capability-job attention weight matrix.

[0025] Multiply the competency-job attention weight matrix with the job value matrix, and restructure the structure with competency dimension as rows and embedding dimension as columns to obtain a set of job competency embedding vectors.

[0026] Each job competency embedding vector is mapped to a scalar competency score through a fully connected layer. All scalar competency scores are then normalized using Softmax to generate a set of job dimension weights.

[0027] Based on the standard communication text sequence, at each time step, the standard communication text is segmented to generate a communication word sequence, and each communication word is input into a pre-trained bidirectional Transformer text encoder to generate a communication word semantic vector;

[0028] All semantic vectors of communication terms are averaged and pooled to obtain communication behavior feature vectors;

[0029] The communication behavior feature vectors are stacked in time step order to generate a communication behavior feature vector sequence.

[0030] Optionally, step four specifically includes:

[0031] In the embedding mapping layer, the sequence of communication behavior feature vectors is reshaped into a communication behavior feature matrix;

[0032] Define the core embedding dimension, where the core embedding dimension is equal to the embedding dimension of the semantic vector of the job term;

[0033] The trainable embedding mapping matrix is ​​initialized based on the backbone embedding dimension, and the communication behavior feature matrix is ​​transformed into a communication behavior embedding matrix through the embedding mapping matrix.

[0034] The backbone coding network consists of several stacked improved Retention Blocks. Each improved Retention Block includes a normalization layer, a low-rank projection query key generation module, a shared basis query key generation module, a short-timescale Retention path, a long-timescale Retention path, a scale-gated fusion module, and a feedforward residual module.

[0035] The input embedding matrix of each improved Retention Block is the output embedding matrix of the previous improved Retention Block; the input embedding matrix of the first improved Retention Block is the communication behavior embedding matrix, and the output embedding matrix of the last improved Retention Block is the professional ability feature matrix.

[0036] The backbone coding network generates a professional ability feature matrix through a dual-time-scale Retention structure.

[0037] In the capability decoupling mapping module, the professional capability feature matrix is ​​averaged in the time step dimension to generate a professional capability feature vector.

[0038] An independent capability mapping matrix is ​​initialized for each capability dimension. The professional capability feature vectors are linearly transformed through each capability mapping matrix to generate professional capability embedding vectors, and a set of professional capability embedding vectors is formed.

[0039] Optionally, the backbone coding network generates a professional competence feature matrix through a dual-time-scale Retention structure, specifically including:

[0040] In the normalization layer, the input embedding matrix is ​​subjected to layer normalization operation to generate a normalized embedding matrix;

[0041] In the low-rank projection query key-value generation module, the short-scale query mapping matrix, the short-scale key mapping matrix, and the short-scale value mapping matrix are initialized.

[0042] The normalized embedding matrix is ​​transformed into a short-scale query matrix, a short-scale key matrix, and a short-scale value matrix respectively through a short-scale query mapping matrix, a short-scale key mapping matrix, and a short-scale value mapping matrix;

[0043] Set the low-rank compression ratio, multiply the low-rank compression ratio by the number of time steps, and round up to obtain the low-rank length;

[0044] Based on the low-rank length, initialize the key low-rank projection matrix and the value low-rank projection matrix according to the structure of low-rank length × number of time steps;

[0045] Multiply the low-rank projection matrix of the bond with the short-scale bond matrix to obtain the short-scale low-rank bond matrix; multiply the low-rank projection matrix of the value with the short-scale value matrix to obtain the short-scale low-rank value matrix.

[0046] In the shared basis query key generation module, the shared basis transformation matrix, long-scale query mapping matrix, long-scale key mapping matrix, and long-scale value mapping matrix are initialized.

[0047] The shared basis matrix is ​​obtained by multiplying the transpose of the capability dimension matrix with the shared basis transformation matrix.

[0048] The normalized embedding matrix is ​​multiplied by the shared basis matrix to generate a shared intermediate feature matrix;

[0049] The shared intermediate feature matrix is ​​transformed into a long-scale query matrix, a long-scale key matrix, and a long-scale value matrix respectively through a long-scale query mapping matrix, a long-scale key mapping matrix, and a long-scale value mapping matrix.

[0050] In the short-timescale Retention path, the short-scale query matrix, short-scale key matrix, and short-scale value matrix are used to generate the short-scale Retention output matrix through recursive update rules and window truncation recursive update rules.

[0051] In the long-scale Retention path, the long-scale query matrix, long-scale key matrix, and long-scale value matrix are updated using a piecewise recursive update rule to generate the long-scale Retention output matrix.

[0052] In the scale-gated fusion module, the normalized embedding matrix is ​​linearly mapped and activated with Sigmoid to generate a gating matrix; and based on the gating matrix, the short-scale Retention output matrix and the long-scale Retention output matrix are gated and fused to generate a fused Retention output matrix.

[0053] In the feedforward residual module, the fused Retention output matrix and the input embedding matrix are residually concatenated to obtain the Retention residual matrix; the Retention residual matrix is ​​then subjected to a feedforward transformation to generate the feedforward output matrix; and the Retention residual matrix and the feedforward output matrix are residually concatenated to obtain the output embedding matrix.

[0054] The output embedding matrix of the last improved Retention Block is used as the professional competence feature matrix.

[0055] Optionally, in the short-timescale Retention path, the short-scale query matrix, short-scale key matrix, and short-scale value matrix are used to generate a short-scale Retention output matrix through a recursive update rule and a window truncation recursive update rule, specifically including:

[0056] Set the short-scale window W and the short-scale recursive decay coefficient, and initialize the short-scale recursive state matrix;

[0057] At time step t, based on the short-scale query matrix, the short-scale low-rank key matrix, and the short-scale low-rank value matrix, the short-scale query vector, the short-scale low-rank key vector, and the short-scale low-rank value vector are extracted.

[0058] When time step t is less than or equal to the short-scale window, the short-scale recursive state matrix is ​​updated according to the recursive update rule based on the short-scale low-rank key vector and the short-scale low-rank value vector.

[0059] When time step t is greater than the short-scale window, the short-scale low-rank key vector and short-scale low-rank value vector at time step tW are taken as the expired short-scale low-rank key vector and expired short-scale low-rank value vector, and the short-scale recursive state matrix is ​​updated according to the window truncation recursive update rule.

[0060] At time step t, the short-scale query vector is multiplied by the short-scale recursive state matrix to obtain the short-scale Retention output vector, and the short-scale Retention output vectors are stacked in the order of time steps to form the short-scale Retention output matrix.

[0061] Optionally, step five specifically includes:

[0062] For each competency dimension, the L2 norm of the vector difference between the job competency embedding vector and the professional competency embedding vector is used as the demand gap.

[0063] The demand gap is used to calculate the generation capacity satisfaction rate through an exponential decay function.

[0064] Set a target threshold, calculate the difference between the demand gap and the target threshold, and take the maximum value of the difference and 0 as the non-compliance gap;

[0065] Based on the set of job dimension weights, the job dimension weights of each competency dimension are multiplied by the gaps in competency, and all competency dimensions are summed to obtain a weighted gap penalty item.

[0066] Based on the set of job dimension weights, the dimensional satisfaction of each ability dimension is weighted and averaged to obtain the weighted average satisfaction.

[0067] Set a gap penalty coefficient, multiply the gap penalty coefficient by the weighted gap penalty term to obtain the penalty amount, calculate the difference between the weighted satisfaction mean and the penalty amount, and perform a truncation operation using the truncation operator in the [0,1] interval to obtain the job suitability score.

[0068] Optionally, step six specifically includes:

[0069] Set basic adaptation threshold, good adaptation threshold, and high adaptation threshold;

[0070] When the job fit score is greater than or equal to the high fit threshold, the job fit level of the person to be evaluated is determined to be the high fit level.

[0071] When the job fit score is greater than or equal to the good fit threshold and less than the high fit threshold, the job fit level of the person being evaluated is determined to be the good fit level.

[0072] When the job fit score is greater than or equal to the basic fit threshold and less than the good fit threshold, the job fit level of the person to be evaluated is determined to be the basic fit level.

[0073] When the job fit score is less than the basic fit threshold, the job fit level of the person to be evaluated is determined to be a low fit level.

[0074] Output the job suitability level of the candidate to be evaluated.

[0075] Optionally, step seven specifically includes:

[0076] Obtain the actual performance evaluation level of the candidate after completing the target position; map the actual performance evaluation level numerically to generate the actual performance target value;

[0077] Calculate the mean squared error between the job fit score and the actual performance target value to obtain the supervised learning loss function;

[0078] Backpropagation is performed based on the supervised learning loss function to update the model parameters of the improved RetNet network. The model parameters include the embedding mapping layer parameters, the backbone coding network parameters, and the capability decoupling mapping module parameters.

[0079] The beneficial effects of this invention are:

[0080] This invention introduces an improved RetNet network incorporating a low-rank projection structure, a shared basis structure, and a dual-timescale Retention structure. It performs recursive modeling of communication behavior feature vector sequences on both short and long timescales, and generates a set of occupational ability embedding vectors through an ability decoupling mapping module. This achieves a hierarchical characterization of the instantaneous fluctuations and long-term stable features in the communication behavior sequence, improving the stability and discriminativeness of occupational ability representation. In the matching stage, a job fit strategy based on L2 norm requirement gap, exponential decay function, qualification threshold gap, and weighted gap penalty term is adopted to explicitly penalize the risk of failing to meet key competency standards. Simultaneously, an interval truncation operator ensures the controllability and consistency of job fit scoring, improving the interpretability of fitting results and the accuracy of risk identification. In the decision-making stage, fit level classification is completed based on a preset threshold range, achieving a structured output of the evaluation results. During the feedback phase, a mean squared error supervised learning loss function is constructed by introducing real performance evaluation levels, and backpropagation is performed to incrementally update the parameters of the improved RetNet network, forming a closed-loop mechanism of capability modeling, matching evaluation, performance feedback and model optimization, thereby improving the generalization ability and long-term adaptive optimization ability of the evaluation model in different job scenarios. Attached Figure Description

[0081] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0082] Figure 1 This is a schematic diagram of an intelligent assessment method for professional competence and job suitability based on communication behavior proposed in this invention.

[0083] Figure 2 This is a flowchart of the improved RetNet network structure in the intelligent assessment method for professional competence and job suitability based on communication behavior proposed in this invention.

[0084] Figure 3 This is a flowchart of job suitability scoring and grade determination in an intelligent assessment method for professional competence and job suitability based on communication behavior proposed in this invention. Detailed Implementation

[0085] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0086] refer to Figures 1-3 A smart assessment method for professional competence and job suitability based on communication behavior includes the following steps:

[0087] Step 1: Obtain the target job description text and the original communication data between the candidate and the candidate to be evaluated;

[0088] Step 2: Perform data preprocessing on the target job description text and the original communication data to obtain structured job description text and standard communication text sequence;

[0089] Step 3: Semantically embed and encode the structured job description text and standard communication text sequence to obtain a set of job competency embedding vectors, a set of job dimension weights, and a sequence of communication behavior feature vectors;

[0090] Step 4: Input the communication behavior feature vector sequence into the improved RetNet network, perform dual-timescale recursive modeling and professional ability embedding modeling, and generate a set of professional ability embedding vectors; the improved RetNet network includes an embedding mapping layer, a backbone encoding network and an ability decoupling mapping module; the backbone encoding network introduces a low-rank projection structure, a shared basis structure and a dual-timescale Retention structure;

[0091] Step 5: Design a job fit strategy, match and calculate the set of professional ability embedding vectors, the set of job ability embedding vectors, and the set of job dimension weights to obtain a job fit score;

[0092] Step Six: Determine the suitability level of the candidate based on the job suitability score;

[0093] Step 7: Obtain the true performance rating of the object to be evaluated and incrementally update the improved RetNet network.

[0094] In this embodiment, the target job description text includes job responsibility items, job qualification items, job skill requirements items, job competency requirements items, and job performance evaluation standards items; the raw communication data includes voice communication data and text communication data.

[0095] In this embodiment, step two specifically includes:

[0096] The job description text and text communication data are segmented according to punctuation marks, and web page tags, special symbols and semantically meaningless characters are deleted to generate a structured job description text and structured communication text sequence.

[0097] The voice communication data is denoised, silence is detected, and speech is recognized and converted, and a speech-to-text sequence is generated in chronological order.

[0098] The speech-to-text sequence and the structured communication text sequence are aligned according to a set time step and merged at the text level to obtain a standard communication text sequence.

[0099] In this embodiment, step three specifically includes:

[0100] The structured job description text is segmented to generate a job term sequence. The job term sequence is then input into a pre-trained bidirectional Transformer text encoder to obtain a job term semantic vector sequence.

[0101] Initialize the capability dimension matrix, query linear mapping matrix, key linear mapping matrix, and value linear mapping matrix; wherein, the capability dimension matrix is ​​structured as the set number of capability dimensions × the embedding dimension of the job term semantic vector; the query linear mapping matrix, key linear mapping matrix, and value linear mapping matrix are structured as the embedding dimension of the job term semantic vector × the embedding dimension of the job term semantic vector; in the implementation process, capability dimensions include language expression ability, logical organization ability, collaborative interaction ability, emotional stability ability, and job professional ability;

[0102] The sequence of semantic vectors of job terms is stacked according to time steps to form a job semantic matrix;

[0103] The capability dimension matrix is ​​transformed into a capability query matrix through a query linear mapping matrix, and the job semantic matrix is ​​transformed into a job key matrix and a job value matrix through a key linear mapping matrix and a value linear mapping matrix, respectively.

[0104] Multiply the capability query matrix by the transpose of the job key matrix, divide by the square root of the embedding dimension of the job term semantic vector, and then normalize using the Softmax function to generate the capability-job attention weight matrix.

[0105] Multiply the competency-job attention weight matrix with the job value matrix, and restructure the structure with competency dimension as rows and embedding dimension as columns to obtain a set of job competency embedding vectors.

[0106] Each job competency embedding vector is mapped to a scalar competency score through a fully connected layer. All scalar competency scores are then normalized using Softmax to generate a set of job dimension weights.

[0107] Based on the standard communication text sequence, at each time step, the standard communication text is segmented to generate a communication word sequence, and each communication word is input into a pre-trained bidirectional Transformer text encoder to generate a communication word semantic vector;

[0108] All semantic vectors of communication terms are averaged and pooled to obtain communication behavior feature vectors;

[0109] The communication behavior feature vectors are stacked in time step order to generate a communication behavior feature vector sequence.

[0110] In this embodiment, step four specifically includes:

[0111] In the embedding mapping layer, the sequence of communication behavior feature vectors is reshaped into a communication behavior feature matrix;

[0112] Define the core embedding dimension, where the core embedding dimension is equal to the embedding dimension of the semantic vector of the job term;

[0113] The trainable embedding mapping matrix is ​​initialized based on the backbone embedding dimension, and the communication behavior feature matrix is ​​transformed into a communication behavior embedding matrix through the embedding mapping matrix.

[0114] The backbone coding network consists of several improved Retention Blocks stacked together. Each improved Retention Block includes a normalization layer, a low-rank projection query key generation module, a shared basis query key generation module, a short-timescale Retention path, a long-timescale Retention path, a scale-gated fusion module, and a feedforward residual module.

[0115] The input embedding matrix of each improved Retention Block is the output embedding matrix of the previous improved Retention Block; the input embedding matrix of the first improved Retention Block is the communication behavior embedding matrix, and the output embedding matrix of the last improved Retention Block is the professional ability feature matrix.

[0116] The backbone encoding network generates a professional competence feature matrix through a dual-time-scale Retention structure;

[0117] In the capability decoupling mapping module, the professional capability feature matrix is ​​averaged in the time step dimension to generate a professional capability feature vector.

[0118] An independent capability mapping matrix is ​​initialized for each capability dimension. The professional capability feature vectors are linearly transformed through each capability mapping matrix to generate professional capability embedding vectors, and a set of professional capability embedding vectors is formed.

[0119] In this embodiment, the backbone coding network generates a professional ability feature matrix through a dual-time-scale Retention structure, specifically including:

[0120] In the normalization layer, the input embedding matrix is ​​subjected to layer normalization operation to generate a normalized embedding matrix;

[0121] In the low-rank projection query key-value generation module, the short-scale query mapping matrix, the short-scale key mapping matrix, and the short-scale value mapping matrix are initialized. The structure of the short-scale query mapping matrix, the short-scale key mapping matrix, and the short-scale value mapping matrix is ​​the backbone embedding dimension × the backbone embedding dimension.

[0122] The normalized embedding matrix is ​​transformed into a short-scale query matrix, a short-scale key matrix, and a short-scale value matrix respectively through a short-scale query mapping matrix, a short-scale key mapping matrix, and a short-scale value mapping matrix;

[0123] Set the low-rank compression ratio, multiply the low-rank compression ratio by the number of time steps, and round up to obtain the low-rank length;

[0124] Based on the low-rank length, initialize the key low-rank projection matrix and the value low-rank projection matrix according to the structure of low-rank length × number of time steps;

[0125] Multiply the low-rank projection matrix of the bond with the short-scale bond matrix to obtain the short-scale low-rank bond matrix; multiply the low-rank projection matrix of the value with the short-scale value matrix to obtain the short-scale low-rank value matrix.

[0126] In the shared basis query key generation module, the shared basis transformation matrix, long-scale query mapping matrix, long-scale key mapping matrix, and long-scale value mapping matrix are initialized. The shared basis transformation matrix has the structure of capability dimension number × shared basis space dimension, and the shared basis space dimension is smaller than the backbone embedding dimension. The long-scale query mapping matrix, long-scale key mapping matrix, and long-scale value mapping matrix have the structure of shared basis space dimension × backbone embedding dimension.

[0127] The shared basis matrix is ​​obtained by multiplying the transpose of the capability dimension matrix with the shared basis transformation matrix.

[0128] The normalized embedding matrix is ​​multiplied by the shared basis matrix to generate a shared intermediate feature matrix;

[0129] The shared intermediate feature matrix is ​​transformed into a long-scale query matrix, a long-scale key matrix, and a long-scale value matrix respectively through a long-scale query mapping matrix, a long-scale key mapping matrix, and a long-scale value mapping matrix.

[0130] In the short-timescale Retention path, the short-scale query matrix, short-scale key matrix, and short-scale value matrix are used to generate the short-scale Retention output matrix through recursive update rules and window truncation recursive update rules.

[0131] In the long-scale Retention path, the long-scale query matrix, long-scale key matrix, and long-scale value matrix are updated using a piecewise recursive update rule to generate the long-scale Retention output matrix, specifically:

[0132] Set the segmented switching time step, the long-scale decay coefficient A1 and the long-scale decay coefficient A2, and initialize the long-scale recursive state matrix;

[0133] At time step t, extract the long-scale query vector, long-scale key vector, and long-scale value vector;

[0134] When time step t is less than or equal to the segmented switching time step, the long-scale recursive state matrix is ​​updated according to the recursive update rule based on the long-scale decay coefficient A1, the long-scale key vector and the long-scale value vector.

[0135] When the time step t is greater than the segmented switching time step, the long-scale recursive state matrix is ​​updated according to the recursive update rule based on the long-scale decay coefficient A2, the long-scale key vector and the long-scale value vector.

[0136] At time step t, the long-scale query vector is multiplied by the long-scale recursive state matrix to obtain the long-scale Retention output vector, and the long-scale Retention output vectors are stacked in the order of time steps to form the long-scale Retention output matrix.

[0137] In the scale-gated fusion module, the normalized embedding matrix is ​​linearly mapped and activated with Sigmoid to generate a gating matrix; and based on the gating matrix, the short-scale Retention output matrix and the long-scale Retention output matrix are gated and fused to generate a fused Retention output matrix.

[0138] In the feedforward residual module, the fused Retention output matrix and the input embedding matrix are residually concatenated to obtain the Retention residual matrix; the Retention residual matrix is ​​then subjected to a feedforward transformation to generate the feedforward output matrix; and the Retention residual matrix and the feedforward output matrix are residually concatenated to obtain the output embedding matrix.

[0139] The output embedding matrix of the last improved Retention Block is used as the professional competence feature matrix.

[0140] In this embodiment, in the short-timescale Retention path, the short-scale query matrix, short-scale key matrix, and short-scale value matrix are used to generate a short-scale Retention output matrix through a recursive update rule and a window truncation recursive update rule. Specifically, this includes:

[0141] Set the short-scale window W and the short-scale recursive decay coefficient, and initialize the short-scale recursive state matrix;

[0142] At time step t, based on the short-scale query matrix, the short-scale low-rank key matrix, and the short-scale low-rank value matrix, the short-scale query vector, the short-scale low-rank key vector, and the short-scale low-rank value vector are extracted.

[0143] When time step t is less than or equal to the short-scale window, the short-scale recursive state matrix is ​​updated according to the recursive update rule based on the short-scale low-rank key vector and the short-scale low-rank value vector.

[0144] The recursive update rule is as follows: multiply the short-scale recursive decay coefficient element-wise with the short-scale recursive state matrix of the previous time step to obtain the decay recursive state matrix; perform an outer product operation on the transpose of the short-scale low-rank key vector and the short-scale low-rank value vector to obtain the key value update matrix; add the decay recursive state matrix and the key value update matrix to obtain the short-scale recursive state matrix.

[0145] When time step t is greater than the short-scale window, the short-scale low-rank key vector and short-scale low-rank value vector at time step tW are taken as the expired short-scale low-rank key vector and expired short-scale low-rank value vector, and the short-scale recursive state matrix is ​​updated according to the window truncation recursive update rule.

[0146] The specific window-truncation recursive update rule is as follows: Multiply the short-scale recursive decay coefficient element-wise with the short-scale recursive state matrix of the previous time step to obtain the decayed recursive state matrix; perform an outer product operation on the transpose of the short-scale low-rank key vector and the short-scale low-rank value vector to obtain the key value update matrix; perform an outer product operation on the transpose of the expired short-scale low-rank key vector and the expired short-scale low-rank value vector to obtain the expired key value contribution matrix; calculate the short-scale window power of the short-scale recursive decay coefficient to obtain the window decay coefficient; multiply the window decay coefficient element-wise with the expired key value contribution matrix to obtain the decayed expired contribution matrix; add the decayed recursive state matrix to the key value update matrix and subtract the decayed expired contribution matrix to obtain the short-scale recursive state matrix.

[0147] At time step t, the short-scale query vector is multiplied by the short-scale recursive state matrix to obtain the short-scale Retention output vector, and the short-scale Retention output vectors are stacked in the order of time steps to form the short-scale Retention output matrix.

[0148] In this invention, the improved RetNet network retains the basic structural form of the original RetNet network in its overall framework, both employing a backbone encoding network composed of embedded mapping layers and stacked multi-layer Retention Blocks. The original RetNet network focuses on sequence modeling, constructing the Retention recursive state through query matrices, key matrices, and value matrices, and accumulating and updating it over time to achieve retention and decay control of historical information. Each Retention Block includes a layer normalization structure, a module for generating the query key-value matrix using linear mapping, a Retention recursive calculation module, a feedforward network module, and a residual connection structure. By performing matrix multiplication between the current query vector and the recursive state matrix at each time step, the output representation for the current time step is obtained, and the residual structure between layers ensures stable training of deep networks.

[0149] The improved RetNet network introduces several structural improvements based on the original RetNet network. First, it adds a low-rank projection structure. By setting a low-rank compression ratio, it constructs key-low-rank projection matrices and value-low-rank projection matrices, compressing the original key and value matrices into low-rank representations to reduce computational complexity and enhance information aggregation efficiency. Second, it introduces a shared basis structure. The capability dimension matrix is ​​mapped to a shared basis space through a shared basis transformation matrix and fused with the normalized embedding matrix, thus introducing prior information about the capability structure into the long-term Retention path. Third, it introduces a dual-timescale Retention structure in the Retention recursion mechanism. The original single-timescale recursion path is extended to a short-timescale Retention path and a long-timescale Retention path, and the outputs of the two paths are adaptively fused through a scale-gated fusion module. Specifically, the short-timescale Retention path controls the information contribution within a local time window through a window-truncation recursion update rule, while the long-timescale Retention path achieves phased modeling of long-term dependencies through a segmented recursion update rule and different decay coefficients. Finally, a capability decoupling mapping module is added to the output stage. By performing time-dimensional average pooling on the professional capability feature matrix and setting an independent capability mapping matrix, decoupling expression of multiple capability dimensions is achieved.

[0150] Through the aforementioned structural improvements, the improved RetNet network significantly outperforms the original RetNet network in terms of modeling capability and stability. The low-rank projection structure effectively reduces the dimensionality of the key-value matrix, improving computational efficiency and mitigating noise accumulation issues caused by high-dimensional representations. The shared basis structure introduces the capability dimension matrix into the long-term modeling path, allowing the network to explicitly incorporate prior knowledge of job capability structures during the recursive process, thereby enhancing the relevance and consistency of professional capability representation. The dual-timescale Retention structure can separately capture short-term expression fluctuations and long-term behavioral stability features in communication behavior, avoiding excessive smoothing of short-term information or excessive forgetting of long-term information by a single decay mechanism. The window-truncated recursive update rule effectively controls the range of short-term state accumulation, reducing interference from outdated information on the current output. The segmented recursive update rule adjusts the strength of stage-specific information retention through different decay coefficients, enhancing the ability to characterize complex communication behavior patterns. The capability decoupling mapping module makes the final professional capability embedding vector set more discriminative and interpretable across different capability dimensions. In summary, the improved RetNet network improves representation accuracy, model stability, and multi-dimensional capability separation in professional capability sequence modeling scenarios.

[0151] In this embodiment, step five specifically includes:

[0152] For each competency dimension, the L2 norm of the vector difference between the job competency embedding vector and the professional competency embedding vector is used as the demand gap.

[0153] The demand gap is used to calculate the capacity satisfaction level using an exponential decay function. The exponential decay function is as follows: a decay coefficient is set, where the decay coefficient is greater than 0. The decay coefficient is multiplied by the demand gap and mapped through a negative exponential function to generate the capacity satisfaction level.

[0154] Set a target threshold, calculate the difference between the demand gap and the target threshold, and take the maximum value between the difference and 0 as the non-compliance gap;

[0155] Based on the set of job dimension weights, the job dimension weights of each competency dimension are multiplied by the gaps in competency, and all competency dimensions are summed to obtain a weighted gap penalty item.

[0156] Based on the set of job dimension weights, the dimension satisfaction of each ability dimension is aggregated by weighted average to obtain the weighted average satisfaction. Specifically, the job dimension weight of each ability dimension is multiplied by the dimension satisfaction, and all ability dimensions are summed. Then, the result is divided by the sum of the job dimension weights of each ability dimension to obtain the weighted average satisfaction.

[0157] A gap penalty coefficient is set, and the gap penalty coefficient is multiplied by the weighted gap penalty term to obtain the penalty amount. The difference between the weighted satisfaction mean and the penalty amount is calculated, and a truncation operation is performed using the truncation operator in the [0,1] interval to obtain the job suitability score. The truncation operator in the [0,1] interval means that the job suitability score is restricted to the interval [0,1]. If the job suitability score is greater than 1, the job suitability score is 1; if the job suitability score is less than 0, the job suitability score is 0.

[0158] In this embodiment, step six specifically includes:

[0159] Set a basic adaptation threshold, a good adaptation threshold, and a high adaptation threshold, where 0 < basic adaptation threshold < good adaptation threshold < high adaptation threshold < 1;

[0160] When the job fit score is greater than or equal to the high fit threshold, the job fit level of the person to be evaluated is determined to be the high fit level.

[0161] When the job fit score is greater than or equal to the good fit threshold and less than the high fit threshold, the job fit level of the person being evaluated is determined to be the good fit level.

[0162] When the job fit score is greater than or equal to the basic fit threshold and less than the good fit threshold, the job fit level of the person to be evaluated is determined to be the basic fit level.

[0163] When the job fit score is less than the basic fit threshold, the job fit level of the person to be evaluated is determined to be a low fit level.

[0164] Output the job suitability level of the candidate to be evaluated.

[0165] In this embodiment, step seven specifically includes:

[0166] Obtain the actual performance evaluation level of the candidate after completing the target position; map the actual performance evaluation level numerically to generate the actual performance target value;

[0167] Calculate the mean squared error between the job fit score and the actual performance target value to obtain the supervised learning loss function;

[0168] Backpropagation is performed based on the supervised learning loss function to update the model parameters of the improved RetNet network. The model parameters include the embedding mapping layer parameters, the backbone encoding network parameters, and the capability decoupling mapping module parameters.

[0169] Example 1: To verify the feasibility of this invention in practice, the method was applied to the recruitment scenario for a technical R&D position at a large internet technology company. Previously, the company primarily relied on subjective interviewer ratings and keyword matching systems for initial screening and ranking. However, in actual recruitment, it was found that the correlation between high interview scores and actual performance after onboarding was unstable. Some candidates performed fluently during the interview but had low performance evaluations three months after joining the company; conversely, some candidates appeared slightly nervous but demonstrated excellent performance after joining. The company hopes to achieve a more objective and quantifiable assessment of job fit by using structured modeling of interview communication behavior combined with job competency requirements.

[0170] In the implementation scenario, the following steps are taken: First, the text of job descriptions, job qualifications, job skills requirements, job competency requirements, and job performance evaluation criteria for the target position are collected. Additionally, voice communication data and written Q&A records from two rounds of technical interviews and one round of comprehensive interviews are obtained. The voice communication data is converted into a speech-to-text sequence after denoising, silence detection, and speech recognition. This sequence is then aligned with the text communication data according to time steps to generate a standard communication text sequence. The structured job description text and the standard communication text sequence are input into a pre-trained bidirectional Transformer text encoder for semantic embedding encoding, yielding a set of job competency embedding vectors and a set of job dimension weights, along with a sequence of communication behavior feature vectors.

[0171] The communication behavior feature vector sequence is input into an improved RetNet network. A low-rank projection structure and a shared basis structure are introduced into the backbone encoding network. A dual-timescale Retention structure is used to recursively model the short-term temporal scale expression fluctuations and long-term temporal scale expression stability, generating a set of professional ability embedding vectors. In the matching phase, the L2 norm of the job ability embedding vector and the professional ability embedding vector is calculated as the demand gap for each ability dimension. An exponential decay function is used to calculate the ability satisfaction, and a weighted gap penalty term is calculated by combining the qualification threshold and the job dimension weight set to finally obtain the job fit score. Based on preset basic fit thresholds, good fit thresholds, and high fit thresholds, candidates are divided into low fit, basic fit, good fit, and high fit levels. Three months after a candidate joins the company, their actual performance evaluation level is obtained and mapped to a numerical target value. The mean squared error between the job fit score and the actual performance target value is calculated, and backpropagation is performed to update the parameters of the improved RetNet network.

[0172] To further verify the effectiveness of the method of this invention, a comparative experiment was conducted with three alternative schemes. Comparison scheme one was a traditional keyword matching scoring model; comparison scheme two was a Transformer matching model based on resume text similarity; and comparison scheme three was a model using only a single-time-scale RetNet model without a job suitability strategy. The test data consisted of a sample of 326 valid candidates collected over a six-month recruitment period, of whom 118 were ultimately hired. The comparison results are shown in Table 1.

[0173] Table 1. Comparison of different approaches in assessing job fit.

[0174] Comparison indicators Comparison Option 1 Comparison Option 2 Comparison Option 3 Method of the present invention Performance forecast accuracy (%) 62.4 71.8 78.6 89.3 High performance recognition rate (%) 58.1 69.2 75.4 91.7 Low performance misjudgment rate (%) 27.6 19.4 15.8 6.9 Correlation coefficient between fit score and actual performance 0.41 0.56 0.63 0.82 Standard deviation of rating stability 0.173 0.142 0.119 0.071 Accuracy rate of identifying critical capability gaps (%) 52.7 66.5 72.3 90.4 Satisfaction with hiring decisions (internal assessment, maximum score 5) 3.1 3.7 4.0 4.6

[0175] As shown in Table 1, the method of this invention achieves a performance prediction accuracy of 89.3%, an improvement of 26.9 percentage points compared to Comparative Scheme 1, 17.5 percentage points compared to Comparative Scheme 2, and 10.7 percentage points compared to Comparative Scheme 3. The high-performance identification rate reaches 91.7%, indicating that the method of this invention significantly improves the ability to identify excellent candidates. The low-performance false positive rate decreases to 6.9%, demonstrating that the method of this invention can effectively identify the risk of key competency deficiencies through the job fit strategy. The correlation coefficient between job fit score and actual performance reaches 0.82, significantly higher than other comparative schemes, indicating that the dual-timescale retention structure constructed by the method of this invention can more accurately characterize the long-term communication ability characteristics of candidates.

[0176] Furthermore, the standard deviation of the scoring stability of the method of this invention is only 0.071, which significantly reduces scoring fluctuations and makes the evaluation results more stable compared to the comparative scheme. The accuracy rate of key competency gap identification reaches 90.4%, verifying the effectiveness of the exponential decay function and gap penalty mechanism in key competency risk identification. The satisfaction score of internal recruitment decisions reaches 4.6 points, indicating that the method of this invention has good interpretability and usability in actual recruitment scenarios.

[0177] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligently assessing professional competence and job suitability based on communication behavior, characterized in that, Includes the following steps: Step 1: Obtain the target job description text and the original communication data between the candidate and the candidate to be evaluated; Step 2: Perform data preprocessing on the target job description text and the original communication data to obtain structured job description text and standard communication text sequence; Step 3: Semantically embed and encode the structured job description text and standard communication text sequence to obtain a set of job competency embedding vectors, a set of job dimension weights, and a sequence of communication behavior feature vectors; Step 4: Input the communication behavior feature vector sequence into the improved RetNet network, perform dual-timescale recursive modeling and professional ability embedding modeling, and generate a set of professional ability embedding vectors; the improved RetNet network includes an embedding mapping layer, a backbone encoding network and an ability decoupling mapping module; wherein, the backbone encoding network introduces a low-rank projection structure, a shared basis structure and a dual-timescale Retention structure; Step 5: Design a job fit strategy, match and calculate the set of professional ability embedding vectors, the set of job ability embedding vectors, and the set of job dimension weights to obtain a job fit score; Step Six: Determine the suitability level of the candidate based on the job suitability score; Step 7: Obtain the true performance rating of the object to be evaluated and incrementally update the improved RetNet network.

2. The intelligent assessment method for professional competence and job suitability based on communication behavior according to claim 1, characterized in that, The target job description text includes job responsibility items, job qualification items, job skill requirements items, job competency requirements items, and job performance evaluation standards items; the raw communication data includes voice communication data and text communication data.

3. The intelligent assessment method for professional competence and job suitability based on communication behavior according to claim 1, characterized in that, Step two specifically includes: The job description text and text communication data are segmented according to punctuation marks, and web page tags, special symbols and semantically meaningless characters are deleted to generate a structured job description text and structured communication text sequence. The voice communication data is denoised, silence is detected, and speech is recognized and converted, and a speech-to-text sequence is generated in chronological order. The speech-to-text sequence and the structured communication text sequence are aligned according to a set time step and merged at the text level to obtain a standard communication text sequence.

4. The intelligent assessment method for professional competence and job suitability based on communication behavior according to claim 1, characterized in that, Step three specifically includes: The structured job description text is segmented to generate a job term sequence. The job term sequence is then input into a pre-trained bidirectional Transformer text encoder to obtain a job term semantic vector sequence. Initialize the capability dimension matrix, query linear mapping matrix, key linear mapping matrix, and value linear mapping matrix; The sequence of semantic vectors of job terms is stacked according to time steps to form a job semantic matrix; The capability dimension matrix is ​​transformed into a capability query matrix through a query linear mapping matrix, and the job semantic matrix is ​​transformed into a job key matrix and a job value matrix through a key linear mapping matrix and a value linear mapping matrix, respectively. Multiply the capability query matrix by the transpose of the job key matrix, divide by the square root of the embedding dimension of the job term semantic vector, and then normalize using the Softmax function to generate the capability-job attention weight matrix. Multiply the competency-job attention weight matrix with the job value matrix, and restructure the structure with competency dimension as rows and embedding dimension as columns to obtain a set of job competency embedding vectors. Each job competency embedding vector is mapped to a scalar competency score through a fully connected layer. All scalar competency scores are then normalized using Softmax to generate a set of job dimension weights. Based on the standard communication text sequence, at each time step, the standard communication text is segmented to generate a communication word sequence, and each communication word is input into a pre-trained bidirectional Transformer text encoder to generate a communication word semantic vector; All semantic vectors of communication terms are averaged and pooled to obtain communication behavior feature vectors; The communication behavior feature vectors are stacked in time step order to generate a communication behavior feature vector sequence.

5. The intelligent assessment method for professional competence and job suitability based on communication behavior according to claim 1, characterized in that, Step four specifically includes: In the embedding mapping layer, the sequence of communication behavior feature vectors is reshaped into a communication behavior feature matrix; Define the core embedding dimension, where the core embedding dimension is equal to the embedding dimension of the semantic vector of the job term; The trainable embedding mapping matrix is ​​initialized based on the backbone embedding dimension, and the communication behavior feature matrix is ​​transformed into a communication behavior embedding matrix through the embedding mapping matrix. The backbone coding network consists of several improved Retention Blocks stacked together. Each improved Retention Block includes a normalization layer, a low-rank projection query key generation module, a shared basis query key generation module, a short-timescale Retention path, a long-timescale Retention path, a scale-gated fusion module, and a feedforward residual module. The input embedding matrix of each improved Retention Block is the output embedding matrix of the previous improved Retention Block; the input embedding matrix of the first improved Retention Block is the communication behavior embedding matrix, and the output embedding matrix of the last improved Retention Block is the professional ability feature matrix. The backbone coding network generates a professional competence feature matrix through a dual-time-scale Retention structure. In the capability decoupling mapping module, the professional capability feature matrix is ​​averaged in the time step dimension to generate a professional capability feature vector. An independent capability mapping matrix is ​​initialized for each capability dimension. The professional capability feature vectors are linearly transformed through each capability mapping matrix to generate professional capability embedding vectors, and a set of professional capability embedding vectors is formed.

6. The intelligent assessment method for professional competence and job suitability based on communication behavior according to claim 5, characterized in that, The backbone coding network generates a professional competence feature matrix through a dual-time-scale Retention structure, specifically including: In the normalization layer, the input embedding matrix is ​​subjected to layer normalization operation to generate a normalized embedding matrix; In the low-rank projection query key-value generation module, the short-scale query mapping matrix, the short-scale key mapping matrix, and the short-scale value mapping matrix are initialized. The normalized embedding matrix is ​​transformed into a short-scale query matrix, a short-scale key matrix, and a short-scale value matrix respectively through a short-scale query mapping matrix, a short-scale key mapping matrix, and a short-scale value mapping matrix; Set the low-rank compression ratio, multiply the low-rank compression ratio by the number of time steps, and round up to obtain the low-rank length; Based on the low-rank length, initialize the key low-rank projection matrix and the value low-rank projection matrix according to the structure of low-rank length × number of time steps; Multiply the low-rank projection matrix of the bond with the short-scale bond matrix to obtain the short-scale low-rank bond matrix; multiply the low-rank projection matrix of the value with the short-scale value matrix to obtain the short-scale low-rank value matrix. In the shared basis query key generation module, the shared basis transformation matrix, long-scale query mapping matrix, long-scale key mapping matrix, and long-scale value mapping matrix are initialized. The shared basis matrix is ​​obtained by multiplying the transpose of the capability dimension matrix with the shared basis transformation matrix. The normalized embedding matrix is ​​multiplied by the shared basis matrix to generate a shared intermediate feature matrix; The shared intermediate feature matrix is ​​transformed into a long-scale query matrix, a long-scale key matrix, and a long-scale value matrix respectively through a long-scale query mapping matrix, a long-scale key mapping matrix, and a long-scale value mapping matrix. In the short-timescale Retention path, the short-scale query matrix, short-scale key matrix, and short-scale value matrix are used to generate the short-scale Retention output matrix through recursive update rules and window truncation recursive update rules. In the long-scale Retention path, the long-scale query matrix, long-scale key matrix, and long-scale value matrix are updated using a piecewise recursive update rule to generate the long-scale Retention output matrix. In the scale-gated fusion module, the normalized embedding matrix is ​​linearly mapped and activated with Sigmoid to generate a gating matrix; and based on the gating matrix, the short-scale Retention output matrix and the long-scale Retention output matrix are gated and fused to generate a fused Retention output matrix. In the feedforward residual module, the fused Retention output matrix and the input embedding matrix are residually concatenated to obtain the Retention residual matrix; the Retention residual matrix is ​​then subjected to a feedforward transformation to generate the feedforward output matrix; and the Retention residual matrix and the feedforward output matrix are residually concatenated to obtain the output embedding matrix. The output embedding matrix of the last improved Retention Block is used as the professional competence feature matrix.

7. The intelligent assessment method for professional competence and job suitability based on communication behavior according to claim 6, characterized in that, In the short-timescale Retention path, the short-scale query matrix, short-scale key matrix, and short-scale value matrix are used to generate a short-scale Retention output matrix through a recursive update rule and a window truncation recursive update rule. Specifically, this includes: Set the short-scale window W and the short-scale recursive decay coefficient, and initialize the short-scale recursive state matrix; At time step t, based on the short-scale query matrix, the short-scale low-rank key matrix, and the short-scale low-rank value matrix, the short-scale query vector, the short-scale low-rank key vector, and the short-scale low-rank value vector are extracted. When time step t is less than or equal to the short-scale window, the short-scale recursive state matrix is ​​updated according to the recursive update rule based on the short-scale low-rank key vector and the short-scale low-rank value vector. When time step t is greater than the short-scale window, the short-scale low-rank key vector and short-scale low-rank value vector at time step tW are taken as the expired short-scale low-rank key vector and expired short-scale low-rank value vector, and the short-scale recursive state matrix is ​​updated according to the window truncation recursive update rule. At time step t, the short-scale query vector is multiplied by the short-scale recursive state matrix to obtain the short-scale Retention output vector, and the short-scale Retention output vectors are stacked in the order of time steps to form the short-scale Retention output matrix.

8. The intelligent assessment method for professional competence and job suitability based on communication behavior according to claim 1, characterized in that, Step five specifically includes: For each competency dimension, the L2 norm of the vector difference between the job competency embedding vector and the professional competency embedding vector is used as the demand gap. The demand gap is used to calculate the generation capacity satisfaction rate through an exponential decay function. Set a target threshold, calculate the difference between the demand gap and the target threshold, and take the maximum value of the difference and 0 as the non-compliance gap; Based on the set of job dimension weights, the job dimension weights of each competency dimension are multiplied by the gaps in competency, and all competency dimensions are summed to obtain a weighted gap penalty item. Based on the set of job dimension weights, the dimensional satisfaction of each ability dimension is weighted and averaged to obtain the weighted average satisfaction. Set a gap penalty coefficient, multiply the gap penalty coefficient by the weighted gap penalty term to obtain the penalty amount, calculate the difference between the weighted satisfaction mean and the penalty amount, and perform a truncation operation using the truncation operator in the [0,1] interval to obtain the job suitability score.

9. The intelligent assessment method for professional competence and job suitability based on communication behavior according to claim 1, characterized in that, Step six specifically includes: Set basic adaptation threshold, good adaptation threshold, and high adaptation threshold; When the job fit score is greater than or equal to the high fit threshold, the job fit level of the person to be evaluated is determined to be the high fit level. When the job fit score is greater than or equal to the good fit threshold and less than the high fit threshold, the job fit level of the person being evaluated is determined to be the good fit level. When the job fit score is greater than or equal to the basic fit threshold and less than the good fit threshold, the job fit level of the person to be evaluated is determined to be the basic fit level. When the job fit score is less than the basic fit threshold, the job fit level of the person to be evaluated is determined to be a low fit level. Output the job suitability level of the candidate to be evaluated.

10. The intelligent assessment method for professional competence and job suitability based on communication behavior according to claim 1, characterized in that, Step seven specifically includes: Obtain the actual performance evaluation level of the candidate after completing the target position; map the actual performance evaluation level numerically to generate the actual performance target value; Calculate the mean squared error between the job fit score and the actual performance target value to obtain the supervised learning loss function; Backpropagation is performed based on the supervised learning loss function to update the model parameters of the improved RetNet network. The model parameters include the embedding mapping layer parameters, the backbone coding network parameters, and the capability decoupling mapping module parameters.

Images