Competency model-based AI interview evaluation report generation system and method

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an AI interview evaluation report generation system and method based on a competency model, and the system comprises a multi-modal data synchronization module, a feature coding module, a behavior pattern screening module, a competency feature mapping module and an evaluation result generation module. According to the method, the dynamic synchronization of the voice image and the behavior sequence is realized by constructing a multi-source signal time alignment mechanism, the signal matching precision is improved by adopting a time compensation and window division strategy, so that the sequential relationship between expression change and intonation fluctuation is stably presented, and quantifiable behavior vectors are generated through multi-dimensional feature coding; according to the method, semantic and emotion changes are kept continuous and consistent through similarity screening, competency weight mapping is established in the feature fusion stage, behavior features and ability dimension self-adaptive matching is achieved, finally, a multi-dimensional evaluation result is output in a score and text description fusion mode, and the objectivity, interpretability and reliability of the evaluation result are enhanced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an AI interview assessment report generation system and method based on competency models. Background Technology

[0002] The field of artificial intelligence technology encompasses core aspects such as machine learning, natural language processing, and human-computer interaction. Through data acquisition, feature analysis, knowledge reasoning, and decision support, it simulates human intelligence in learning, cognition, and task execution. It is widely applied in systematic intelligent processing scenarios such as speech recognition, image recognition, autonomous driving, intelligent recommendation, and intelligent evaluation.

[0003] Traditional AI interview assessment report generation systems refer to systems designed for recruitment interviews that use candidates' interview answers as input. Human interviewers record candidates' verbal content, non-verbal behavior, and response performance according to a predetermined competency scoring sheet. The information is then manually compiled and summarized to form a text-based assessment result. Alternatively, pre-set keyword matching rules and simple score calculation steps are used to score the text content, and the report output is arranged according to the order of the scoring items. This method mainly relies on manual annotation and fixed rules to generate interview assessment reports.

[0004] Existing interview assessment reports rely on manual observation and fixed scoring sheets. The data collection process lacks temporal consistency of multi-source signals. Manual recording of verbal and non-verbal performances is susceptible to attentional bias and emotional interference. Rule-based scoring methods are based solely on word or sentence matching, failing to capture candidates' behavioral characteristics in changes in tone of voice, subtle facial expressions, and reaction rhythm. Information presentation remains at the static text level, lacking integrated analysis of behavioral dynamics and semantic logic. This results in low accuracy of the correlation between assessment results and competency dimensions, and the reports are prone to descriptive imbalances and judgment biases, making it difficult to support the quantitative identification and reliable assessment of candidates' comprehensive behavioral characteristics. Summary of the Invention

[0005] To address the shortcomings of existing interview assessment reports, which rely on manual observation and fixed scoring sheets, lack temporal consistency of multi-source signals during data collection, are susceptible to attentional bias and emotional interference in manually recording verbal and nonverbal performance, and whose rule-based scoring methods are based solely on word or sentence matching and fail to capture candidates' behavioral characteristics in changes in tone of voice, subtle facial expressions, and reaction rhythm, the presenting of information remains at the static text level, lacking the integration and analysis of behavioral dynamics and semantic logic. This results in low accuracy of the correlation between assessment results and competency dimensions, and the reports are prone to descriptive imbalances and judgment biases, making it difficult to support the quantitative identification and reliable assessment of candidates' comprehensive behavioral characteristics. This invention provides an AI-based interview assessment report generation system and method based on a competency model.

[0006] On the one hand, it provides an AI-powered interview assessment report generation system based on a competency model, which includes: The multimodal data synchronization module collects image, voice, and operational behavior signals during the candidate's interview process, calculates the time offset of the channel timestamp data, compensates for errors based on the dynamic time warping algorithm, generates a multimodal synchronization dataset, and transmits it to the feature encoding module. The feature encoding module receives the multimodal synchronous dataset, uses a multi-channel identifier embedding function to divide the data into fixed windows and mark time correction information, extracts facial expression, tone change and reaction time parameters, generates a multi-dimensional feature encoding vector, and transmits it to the behavior pattern screening module. The behavior pattern filtering module calculates the time window similarity coefficient of the multidimensional feature encoding vector, filters out time window features below a preset similarity threshold, retains time periods where the consistency between facial expression changes and tone changes exceeds a threshold, generates a competency-related behavior feature set, and passes it to the competency feature mapping module. The competency feature mapping module calls the competency-related behavioral feature set, uses a multimodal fusion function to match behavioral features with competency dimension weights, assigns contribution scores, generates competency assessment result data, and transmits it to the assessment result generation module.

[0007] As a further aspect of the present invention, the multimodal synchronization dataset includes a time synchronization index, a modality mapping matrix, and synchronization accuracy parameters; the multidimensional feature encoding vector includes facial expression features, speech prosody features, and reaction time features; the competency-related behavioral feature set includes key behavioral segments, emotion and tone consistency labels, and behavioral pattern feature identifiers; and the competency assessment result data includes dimensional result distribution, weight contribution coefficients, and level mapping indicators.

[0008] As a further aspect of the present invention, the multimodal data synchronization module includes: The channel signal acquisition submodule acquires image, voice and operation behavior signals during the candidate interview process, detects the stability of the sampling port input and records the channel identification code, performs time segmentation and frame division on the signal sequence according to the sampling frequency, calculates and sorts the frame start and end timestamps, and establishes a channel synchronization index set. The time offset calculation submodule, based on the channel synchronization index set, extracts the time interval between adjacent channels for the timestamp sequences of image, voice and operation behavior channels, calculates the interval difference and establishes a set of difference intervals, detects the extreme points of the offset boundary and performs weighted calculation to obtain the channel time offset coefficient. The dynamic time warping compensation submodule calls the channel time offset coefficient, performs synchronous adjustment on the channel signal time series based on the dynamic time warping algorithm, calculates the error interval based on the matching degree between signals and extracts boundary samples, performs interpolation calculation and sequence rearrangement on the error interval, and generates a multimodal synchronization dataset.

[0009] As a further aspect of the present invention, the feature encoding module includes: The synchronous data processing submodule acquires facial images, voice waves, and reaction time records from the multimodal synchronous dataset, compares the data according to the signal timestamp order, repairs missing frames using linear interpolation, reconstructs the index according to the channel identifier, unifies the sampling frequency, and generates a multimodal synchronous signal sequence set. The timing synchronization segmentation submodule, based on the multi-modal synchronization signal sequence set, uses a multi-channel identifier embedding function to encode the channel sequence, calculates the signal amplitude change rate and sets the sampling window length, and corrects the offset based on the reference difference to obtain the time-corrected segmented signal set; The feature vector extraction submodule calculates the amplitude of facial expression changes in the facial image channel, the rate of change of tone in the speech channel, and the delay interval in the reaction time channel based on the time-corrected segmented signal set. The three types of features are then concatenated over time to generate a multidimensional feature encoding vector.

[0010] As a further aspect of the present invention, the behavior pattern screening module includes: The feature similarity calculation submodule obtains the facial expression change and tone change parameters of the time window in the multidimensional feature encoding vector sequence, calculates the difference of vectors of the same dimension and performs Euclidean distance operation and normalization, and calculates the similarity based on the normalized distance to generate a time window similarity coefficient set. The time window filtering submodule, based on the time window similarity coefficient set, marks time windows below a preset similarity threshold and retains the remaining windows. It calculates the difference between the rate of change of facial expression and the rate of change of tone in the retained windows and divides the intervals according to the continuity of the difference to obtain a consistency filtering interval sequence. The behavior consistency extraction submodule calls the consistency filtering interval sequence, performs feature aggregation operation on the multi-dimensional feature encoding vector within the interval, calculates the cross-correlation degree of facial expression change feature quantity and tone change feature quantity, and filters the synchronization segment according to the correlation degree threshold to generate a competency-related behavior feature set.

[0011] As a further aspect of the present invention, the preset similarity threshold is determined based on the average Euclidean distance distribution characteristics between facial expression changes and tone changes parameters of different individuals under the same task conditions. The correlation threshold is determined based on the statistical distribution characteristics of the cross-correlation between facial expression change features and tone change features in the original sample.

[0012] As a further aspect of the present invention, the competency feature mapping module includes: The behavior feature extraction submodule acquires semantic text, speech signals and operation log data from the competency-related behavior feature set, calculates semantic embedding vectors, speech spectrum parameters and operation sequence identifier values, compares overlapping index intervals based on a unified time axis and performs weighting, and generates behavior feature fusion vector values.

[0013] The multimodal fusion calculation submodule calls the behavioral feature fusion vector value, performs feature weighting combination in the multimodal fusion function according to the competency dimension weight parameter set, calculates the correlation coefficient between different modalities and identifies deviation components, performs proportional adjustment for deviation components, and generates a multimodal matching correlation coefficient set. The competency contribution allocation submodule calculates the cumulative weight of the competency dimension correlation coefficients based on the multimodal matching correlation coefficient set, performs linear conversion based on the difference between the score offset and the benchmark value, and generates competency assessment result data.

[0014] As a further aspect of the present invention, the evaluation result generation module, based on the competency assessment result data, performs a grade division calculation according to the competency dimension quantitative score, converts the score into four grade standards: excellent, good, average, and needing improvement, and combines confidence rating analysis with text description and chart display to construct the candidate AI interview evaluation result. The candidate AI interview assessment results include competency level conclusions, confidence level evaluation results, and result visualization content.

[0015] As a further aspect of the present invention, the evaluation result generation module includes: The competency data parsing submodule obtains the multi-dimensional original score sequence from the competency assessment result data, performs weighted and scaled calculations based on weight parameters and normalized benchmark values, compares and corrects the scores with the benchmark interval boundaries and records the differences, and generates a dimensional quantified score set. The grade division calculation submodule, based on the dimensional quantized score set, calls the preset grade interval threshold, performs interval comparison and labeling on the dimensional scores, calculates the grade ratio and overall score deviation, and generates a grade division result set.

[0016] The preset level range threshold is determined based on the dimensional quantification score distribution characteristics and cluster boundaries of candidates at different levels in the original competency assessment sample. The results presentation submodule, based on the graded result set, calls the dimension scores and confidence interval values, calculates the confidence rating index and matches the text template, integrates the proportional chart and text content, and generates the candidate AI interview assessment results. The confidence interval value is determined based on the fluctuation range and standard deviation distribution of the dimensional scores during multiple evaluation processes.

[0017] On the other hand, the competency-based AI interview assessment report generation method, which is executed based on the aforementioned competency-based AI interview assessment report generation system, includes the following steps: S1: By collecting image, voice and operational behavior signals during the candidate's interview process, the time offset of the channel timestamp data is calculated, and the error is compensated based on the dynamic time warping algorithm to generate a multimodal synchronous dataset; S2: Receive the multimodal synchronous dataset, use a multi-channel identifier embedding function to divide the data into fixed windows and mark time correction information, extract facial expression, tone change and reaction time parameters, and generate a multi-dimensional feature encoding vector; S3: Calculate the time window similarity coefficient of the multidimensional feature encoding vector, filter out time window features below the preset similarity threshold, retain the time period when the consistency between facial expression changes and tone changes exceeds the threshold, and generate a competency-related behavior feature set. S4: Call the competency-related behavioral feature set, use a multimodal fusion function to match behavioral features with competency dimension weights, assign contribution scores, and generate competency assessment result data; S5: Based on the competency assessment results, perform a grade division calculation according to the quantitative scores of the competency dimensions, convert the scores into four grade standards: excellent, good, average, and need improvement, and combine the confidence rating analysis with textual descriptions and charts to construct the candidate AI interview assessment results.

[0018] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: By constructing a time alignment mechanism for multi-source signals, dynamic synchronization of speech images and behavioral sequences is achieved. Time compensation and windowing strategies are used to improve signal matching accuracy, ensuring a stable presentation of the temporal relationship between facial expression changes and intonation fluctuations. After multi-dimensional feature encoding, quantifiable behavioral vectors are generated. Similarity screening maintains the continuity and consistency of semantic and emotional changes. In the feature fusion stage, a competency weight mapping relationship is established to achieve adaptive matching of behavioral features and ability dimensions. Finally, the multi-dimensional evaluation results are presented in the report output by fusing scores and textual descriptions, significantly enhancing the objectivity, interpretability, and reliability of the evaluation results. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0020] Figure 1 This is a system schematic diagram of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention; Figure 3 This is a flowchart of the multimodal data synchronization module in this invention; Figure 4 This is a flowchart of the feature encoding module in this invention; Figure 5 This is a flowchart of the behavior pattern filtering module in this invention; Figure 6 This is a flowchart of the competency feature mapping module in this invention; Figure 7 This is a flowchart of the evaluation result generation module in this invention; Figure 8 This is a flowchart of the method of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0022] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0023] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0024] In this embodiment of the invention, sometimes the subscript such as W1 is written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0025] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0026] This invention provides an AI-based interview assessment report generation system based on a competency model, such as... Figure 1-2 The diagram shown illustrates an AI-powered interview assessment report generation system based on a competency model. This system includes: The multimodal data synchronization module collects image, voice, and operational behavior signals during the candidate's interview process, calculates the time offset of the channel timestamp data, compensates for errors based on the dynamic time warping algorithm, generates a multimodal synchronization dataset, and transmits it to the feature encoding module. The feature encoding module receives a multimodal synchronous dataset, uses a multi-channel label embedding function to divide the data into fixed windows and mark time correction information, extracts facial expression, tone change and reaction time parameters, generates a multi-dimensional feature encoding vector, and passes it to the behavior pattern screening module. The behavior pattern screening module calculates the time window similarity coefficient of the multidimensional feature encoding vector, filters out time window features below the preset similarity threshold, retains time periods where the consistency between facial expression changes and tone changes exceeds the threshold, generates a competency-related behavior feature set, and passes it to the competency feature mapping module. The competency feature mapping module calls the competency-related behavioral feature set, uses a multimodal fusion function to match behavioral features with competency dimension weights, assigns contribution scores, generates competency assessment result data, and transmits it to the assessment result generation module. The assessment result generation module, based on the competency assessment result data, performs a level division calculation according to the quantitative score of the competency dimension, converts the score into four level standards: excellent, good, average, and need improvement, and combines confidence rating analysis with text description and chart display to construct the candidate AI interview assessment result; The multimodal synchronization dataset includes a time synchronization index, a modality mapping matrix, and synchronization accuracy parameters. The multidimensional feature encoding vectors include facial expression features, speech prosody features, and reaction time features. The competency-related behavioral feature set includes key behavioral segments, emotion and tone consistency labels, and behavioral pattern feature identifiers. The competency assessment results data includes dimensional result distribution, weight contribution coefficients, and level mapping indicators. The candidate AI interview assessment results include competency level conclusions, confidence evaluation results, and result visualization content.

[0027] Specifically, such as Figure 2 , 3 As shown, the multimodal data synchronization module includes: The channel signal acquisition submodule acquires image, voice and operation behavior signals during the candidate interview process, detects the stability of the sampling port input and records the channel identification code, performs time segmentation and frame division on the signal sequence according to the sampling frequency, calculates and sorts the frame start and end timestamps, and establishes a channel synchronization index set. After acquiring the image, voice, and operational behavior signals during the candidate's interview, the input stability of the video channel is first detected. Specifically, the first 100 frames of the video signal are extracted, and the average difference of the luminance (Luma) component between consecutive frames is calculated. If this difference is lower than a preset stability benchmark of 5, the input is considered stable. For example, if the average luminance difference between consecutive frames of the first 100 acquired video is 3.8, which is less than 5, it is considered stable. Then, the image channel identifier is recorded. Similarly, for the voice channel, its stability is detected by calculating the signal-to-noise ratio (SNR) of the input signal. The SNR is calculated by comparing the signal power with the noise power, and a stable range is defined as an SNR greater than [value missing]. The acceptable range is to The unstable range is below If the signal-to-noise ratio of the currently acquired candidate speech signal is If it is stable, its channel identifier is recorded as follows: For operational behavior signals, detection is performed within a unit of time (e.g.) The determination is made by checking if the number of keyboard keystrokes or mouse clicks is 0. If the number of consecutive keystrokes or mouse clicks is 0, the result is positive. If no events occur, the channel is considered stable, and its channel identifier is recorded. Subsequently, the signal sequence was time-segmented based on the sampling frequency of the multiple channels, and the sampling frequency of the image channels was set to... The sampling frequency of the voice channel is The signal flow is uniformly divided into Within a given time period, the image signal is divided into frames, and the sampling rate is [missing information]. Each frame is approximately [length missing] Therefore, one The time period includes 30 image frames, while the speech signal is divided into... A short frame, that is, a The time period contains 40 audio frames. Next, starting from the beginning of the entire interview session... Starting from this point, calculate the start and end timestamps of each data frame, for example, the first time period after the interview begins ( arrive Within ) the first frame of the image channel The start and end timestamps are The second frame The timestamp of , and so on. The timestamp of the first frame of the voice channel is , the second frame has a timestamp of . The timestamp information of the frames of the channel is aggregated and sorted in ascending order of the start timestamp. Finally, a multi-dimensional list containing channel identification, frame number, frame start timestamp, and frame end timestamp is established to form a channel synchronization index set.

[0028] Time offset calculation sub-module. Based on the channel synchronization index set, for the timestamp sequences of the image, voice, and operation behavior channels, extract the time intervals between adjacent channels, calculate the interval differences, and establish a difference interval set. Detect the offset boundary extreme points and perform weighted operations to obtain the channel time offset coefficient; Based on the established channel synchronization index set, for the timestamp sequences of the image, voice, and operation behavior channels, extract the time intervals between adjacent channels. In this process, the voice channel is selected as the reference, and the time offsets of the image channel [[ID=I9]]and the operation behavior channel relative to the reference are calculated respectively. Taking the image and voice channels as an example, first, identify the starting point of pronunciation in the voice signal. For example, when the candidate says "Hello", the starting voice frame of the pronunciation of the word "Hello", its starting timestamp is . At the same time, in the image signal, identify the starting image frame where the mouth area of the human face undergoes significant deformation through the lip movement detection algorithm. Its starting timestamp is . Then the time interval between these two signals at this time is . Repeat this operation to find 10 such audio-visual synchronization event points during the interview to obtain a set of time interval differences. For example . This set is the difference interval set. Then, analyze the values in this set to detect the offset boundary extreme points. The specific method is to calculate the average value and standard deviation of this set. The average value is , and the standard deviation is . Set a detection interval. The boundaries of the interval are defined as the average value plus or minus times the standard deviation, that is . Calculate the obtained interval as . Identify the values and in the difference set that are not within this interval as offset boundary extreme points and remove them, retaining the set Then, a weighted operation is performed on the remaining 8 differences. The weights are set with reference to the energy value of the speech signal frame at the corresponding synchronization event point. The higher the energy value, the higher the pronunciation clarity, and the higher the reliability of the corresponding time interval difference. Therefore, a higher weight is assigned. Assuming that the weight coefficients obtained after normalizing the energy of the speech frames corresponding to the 8 synchronization event points are as follows: If the total weights are 1, then the weighted average calculation process is as follows: The calculation result is This value is the time offset coefficient of the image channel relative to the audio channel.

[0029] The dynamic time warping compensation submodule calls the channel time offset coefficient, performs synchronous adjustment on the channel signal time series based on the dynamic time warping algorithm, calculates the error interval based on the matching degree between signals and extracts boundary samples, performs interpolation calculation and sequence rearrangement on the error interval, and generates a multimodal synchronization dataset. Call the calculated channel time offset coefficient, for example, the time offset coefficient of the image channel relative to the audio channel is... First, the timestamp sequence of the image channels is preliminarily adjusted. The start and end timestamps of each frame in the channel are added To make it compatible with the voice channel After initial time alignment, to achieve finer synchronization, a link needs to be built between the two time series. The cost matrix, where M is the length of the image feature sequence, and M is the length of the speech feature sequence. Each element in the cost matrix... The value is the first The feature vector of the first image frame (such as a 136-dimensional vector composed of the coordinates of 68 key points extracted from the face) and the feature vector of the second image frame. The Euclidean distance between feature vectors (such as 13-dimensional Mel-frequency cepstral coefficients (MFCC) vectors of a speech frame, for example, the Euclidean distance calculated between the feature vector of the 100th image frame and the feature vector of the 125th speech frame (dimension aligned by zero padding or dimensionality reduction) is: ,but The value is After filling the entire cost matrix, a cumulative cost matrix is ​​further calculated. , of which elements The calculation method is to use the cost of the current position. With the left Below and the lower left The minimum values ​​at these three positions are added together, and the result is obtained by backtracking through this cumulative cost matrix. arrive Finding the minimum-cost path yields a regularized path, which reveals the optimal matching relationship between two sequence elements. Based on the signal matching degree corresponding to this path, i.e., the sum of the cumulative costs along the path, the error interval is calculated. First, the average cost of each step on the regularized path is calculated, assuming it is... The boundary value for error determination is set as the average cost. times, that is Iterate through the cost of each step on the regular path. If the cost of a certain step, for example, from... arrive The cost is Exceeded Then, the 151st image frame and the 181st speech frame are marked as samples within the error interval. When it is found that the cost of multiple consecutive points is higher than the boundary value, the sequence of points constitutes an error interval. The start and end samples of this interval are extracted, for example, frames 151 to 155 of the image sequence. Linear interpolation is performed on the image feature vectors within this interval, that is, the feature vectors of frames 152, 153, and 154 are recalculated using the feature vectors of frames 151 and 155. At the same time, the sequence is rearranged according to the normalized path. If the path appears from arrive The translation of will then make the first The image frame is copied once, so that it is consistent with the first image frame. and the By aligning individual speech frames and performing a series of adjustments, interpolations, and rearrangements, a time-aligned multimodal synchronization dataset is finally generated.

[0030] Specifically, such as Figure 2 , 4 As shown, the feature encoding module includes: The synchronous data processing submodule acquires facial images, voice waves, and reaction time records from the multimodal synchronous dataset, compares the data according to the signal timestamp order, repairs missing frames using linear interpolation, reconstructs the index based on the channel identifier, unifies the sampling frequency, and generates a multimodal synchronous signal sequence set. After acquiring facial images, speech waves, and reaction time records from the multimodal synchronous dataset, the data is first processed according to the timestamp order of the signals. Frame-by-frame comparison is performed at intervals, with timestamps as... Location, discover image channels Existing data, voice channel Data exists, but operational behavior channels are missing. If a data frame is missing, the two most recent valid frames before and after the missing frame are retrieved, i.e., those located at... and The operational behavior data is used to perform linear interpolation to repair missing frames. The mouse X coordinate is , The X coordinate is After interpolation calculation The mouse X coordinate at that location has been corrected. After performing this type of inspection and repair on the channel data, based on the data frame embedded... , , After obtaining the channel identifier, a complete index strictly ordered by timestamp is reconstructed. Then, the sampling frequency of the channels is unified, and a unified target frequency is set, for example... That is, each A data point, for an original sampling frequency higher than The channel, for example, the original sampling rate is The voice channel performs a downsampling operation, specifically every... One point is extracted from each sampling point as... The new sampling points below, for original sampling frequencies lower than The channel, for example, the original sampling rate is (Right now For each frame of an image channel, an upsampling operation is performed by copying or inserting a calculated intermediate frame between the original frames. The frequency, for example, in the original first frame ( ) and the second frame ( Between ), it is necessary to A new data frame is inserted at this point. The content of this frame can be calculated by the weighted average of the first and second frames. The weights are determined by the time distance between the new frame and the two preceding and following frames. The new frame is located at the distance from the first frame. Distance from the second frame Then its pixel value is calculated as the pixel value of the first frame multiplied by Add the second frame pixel value multiplied by After this series of processes, a channel data point is finally generated that is strictly aligned on the time axis and has a sampling frequency of [missing information]. A set of multimodal synchronization signal sequences.

[0031] The timing synchronization segmentation submodule is based on a multi-modal synchronization signal sequence set. It uses a multi-channel identifier embedding function to encode the channel sequence, calculates the signal amplitude change rate and sets the sampling window length, and corrects the offset based on the reference difference to obtain a time-corrected segmented signal set. Based on the generated multimodal synchronization signal sequence set, each data point is first encoded using a multi-channel identifier embedding function. Specifically, this operation is performed on each... A one-hot vector is added before the data vector at each time point to identify its channel source. For example, if the system has three channels (image, audio, and operation), the image channel is identified as [1, 0, 0], the audio channel as [0, 1, 0], and the operation channel as [0, 0, 1]. Thus, in... The image frame feature vector [...] at a given location will be encoded as [1, 0, 0, ..., features...]. Next, the rate of change of the signal amplitude for each channel is calculated. Taking the audio channel as an example, the amplitude is the amplitude value of the audio sample. The rate of change at any given time is calculated as the current amplitude value. Amplitude value compared to the previous moment The absolute value of the difference divided by the time interval ,Right now Then, the length of the sampling window is dynamically set based on the mean of this rate of change, and the high-activity interval of the rate of change is set as the interval with a rate of change greater than 1. The low-activity range is defined as a rate of change of less than When continuous The mean rate of change calculated within the period, for example, is When the activity level is in a high-activity range, the subsequent analysis window length should be set to a shorter value. If the mean is If the activity level is in a low-activity zone, then set the window length to a longer value. Within the defined window, the minute offset is further corrected based on a reference difference. This reference difference is determined by a specific synchronization event identified within the window, such as the peak value of a keyboard keystroke and the corresponding image frame where the key was pressed. Within the window, the sound peak appears at the first... The action of pressing the button in the image occurs in the first... Then the benchmark deviation is Then, the entire sequence of operation signals within this window is moved forward. The signal is corrected to obtain a time-corrected segmented signal set.

[0032] The feature vector extraction submodule calculates the amplitude of facial expression changes in the facial image channel, the rate of pitch change in the speech channel, and the delay interval in the reaction time channel based on the time-corrected segmented signal set. The three types of features are then concatenated over time to generate a multidimensional feature encoding vector. Based on the obtained time-corrected segmented signal set, in each facial image channel... Within each segment, the amplitude of facial expression changes is calculated. Specifically, this involves extracting two key facial points representing the corners of the mouth from each frame, such as the coordinates of the left corner of the mouth. and the coordinates of the right corner of the mouth Calculate the Euclidean distance between two points. ,one The segments contain The frame image is used to obtain a sequence containing 25 distance values. The standard deviation of this sequence is calculated, for example, the value obtained is... This value represents the amplitude of facial expression changes in that segment. Next, the pitch change rate is calculated in the corresponding segment of the speech channel. The fundamental frequency (F0) sequence is extracted from the speech signal within the segment, resulting in a sequence containing 25 fundamental frequency values. For example... Calculate the average value of the sequence, assuming it is . Then, the deviation of each fundamental frequency value from the average value is calculated, and the average of the deviations is used as the pitch change rate. For example, the calculated value is... Finally, the delay interval is calculated in the reaction time channel. Here, the delay refers to the time difference between the end of the AI's question and the start of the candidate's response. The starting point of the response is determined by detecting the energy value of the voice channel. If the voice energy of the AI's question is within the timestamp... The noise level drops to the mute threshold (e.g., energy value is below). Below that, while the candidate's voice energy is... If the threshold is exceeded for the first time, the delay interval is calculated as follows: The feature values ​​calculated for these three categories within the same time segment, i.e., the amplitude of facial expression changes, are then analyzed. intonation variation rate and delay interval The time sequences are concatenated according to a preset order to generate a multidimensional feature encoding vector. .

[0033] Specifically, such as Figure 2 , 5 As shown, the behavior pattern filtering module includes: The feature similarity calculation submodule obtains the facial expression and tone change parameters in the time window of the multidimensional feature encoding vector sequence, calculates the difference of vectors of the same dimension and performs Euclidean distance operation and normalization, and calculates the similarity based on the normalized distance to generate a set of time window similarity coefficients. First, extract the encoding vectors for two consecutive time windows from the multidimensional feature encoding vector sequence. For example, the first window vector is... The second window vector is Extract the facial expression change parameters and tone change parameters for comparison, and construct a two-dimensional vector, i.e., the two-dimensional vector of the first window is... The second one is Next, calculate the difference between these two vectors with the same dimension, that is, subtract the corresponding dimensions to obtain the difference vector. The result is Then, Euclidean distance is performed on the difference vector, specifically by summing the squares of the elements in the difference vector and taking the square root. The calculation process is as follows: The result is Approximately The distance was then normalized using a preset maximum distance value, which was set based on statistical analysis of a large number of interview sample data. Here, it is set to [value missing]. Normalization calculation involves dividing the actual distance by the maximum distance, i.e. The normalized distance is obtained as Finally, the similarity is calculated based on the normalized distance. The calculation method is as follows: Subtract the normalized distance, i.e. The similarity coefficient was obtained as This process is applied to adjacent time window pairs in the encoded vector sequence; for example, the two-dimensional vector of the third window is... Then it is related to the second window The Euclidean distance is Approximately The normalized distance is The similarity is This process is repeated until a set of time window similarity coefficients consisting of a series of similarity values ​​is generated.

[0034] The time window filtering submodule, based on the time window similarity coefficient set, marks time windows below the preset similarity threshold and retains the remaining windows. It calculates the difference between the rate of change of facial expression and the rate of change of tone in the retained windows and divides the intervals according to the continuity of the difference to obtain a consistency filtering interval sequence. Based on the generated time window similarity coefficient set, for example First, the time windows need to be marked according to a preset similarity threshold. This threshold is determined by referring to a dataset containing 50 different individuals completing the same task of "self-introduction," as shown in Table 1. The dataset includes the facial expression and tone change parameters of each person throughout the entire process, and the average Euclidean distance between any two adjacent windows is calculated. These 50 average Euclidean distance values ​​are then statistically analyzed, and their overall average is calculated. The standard deviation is Then the filter boundary was set to the average plus... The standard deviation is 10 times. ,this This is the distance threshold. Any window pair with a distance greater than this threshold is considered dissimilar and converted to a similarity threshold, using the same normalized maximum value as in the previous step. The normalized distance is calculated as follows: The similarity threshold is This threshold is applied to the similarity coefficient set, where the coefficients are... Below Then, the third time window corresponding to that coefficient is marked as to be eliminated, and the remaining windows are retained. Next, for the retained windows, the difference between the rate of change of facial expression and the rate of change of tone within them is calculated. Here, the rate is the parameter value in the encoding vector. For example, the vector of the first window is... The difference is The second window is The difference is The fourth window vector is The difference is This yields a sequence of differences within a retained window. Finally, based on the continuity of this difference sequence, intervals are divided, and the continuity criterion is set as the absolute value of the difference between two adjacent differences not exceeding [a certain value]. In the sequence middle, , Therefore, these three windows form a continuous, consistent interval. If the subsequent difference is... ,but The interval is broken here, and finally a consistent filtering interval sequence consisting of one or more continuous intervals is obtained.

[0035] Table 1: Statistics on Average Euclidean Distance of Individuals under Basic Tasks Individual ID Mean value of facial expression change parameters (pixels) Mean value of intonation variation parameter (Hertz) Average Euclidean Distance 1 4.1 7.2 4.5 2 3.9 8.5 5.1 3 5.2 6.1 4.2 ... ... ... ... 50 4.5 7.9 4.9 As shown in Table 1, this table lists some of the basic data used to determine the similarity threshold.

[0036] The behavioral consistency extraction submodule calls the consistency filtering interval sequence, performs feature aggregation operation on the multi-dimensional feature encoding vector within the interval, calculates the cross-correlation degree of facial expression change feature quantity and tone change feature quantity, and filters the synchronous segment according to the correlation degree threshold to generate a competency-related behavioral feature set. The obtained consistency filtering interval sequence is first processed by performing feature aggregation on the multiple multidimensional feature encoding vectors contained within each interval. Taking an interval containing three time windows as an example, its three complete encoding vectors are as follows: , and The feature aggregation method is to calculate the average value of each feature dimension within the interval. The aggregated value of the facial expression change feature is... The aggregate value of intonation variation features is The aggregate value of the reaction time characteristic is , to obtain the aggregate vector Next, the cross-correlation degree between the facial expression change features and the tone change features within this interval is calculated. Specifically, the time series of these two features within the interval, i.e., the facial expression sequence, is extracted. and intonation sequence Calculate the Pearson correlation coefficient between the two sequences, and the mean value of sequence E is calculated. The mean of sequence P is The covariance of the two sequences is approximately The standard deviation of sequence E is approximately The standard deviation of sequence P is approximately The final cross-correlation degree is The calculation result is approximately Subsequently, this correlation score is compared with a preset correlation score threshold to filter synchronization segments. This correlation score threshold is set based on statistical analysis of a large original interview sample database, and is calculated accordingly. The cross-correlation between facial expressions and intonation features in the segment revealed that... The correlation value of quantiles Therefore, the correlation threshold is set to That is, when the calculated absolute value of the correlation is higher than When this segment is determined to be a synchronization segment, based on the correlation calculated for the current segment... Greater than Therefore, this consistency screening interval is identified as a synchronization segment, and the aggregated feature vector of this segment is... Extract the relevant information and repeat this process on the selected synchronization segments to ultimately generate a competency-related behavioral feature set consisting of multiple aggregated feature vectors.

[0037] Specifically, such as Figure 2 , 6 As shown, the competency feature mapping module includes: The behavior feature extraction submodule acquires semantic text, speech signals and operation log data from the competency-related behavior feature set, calculates semantic embedding vectors, speech spectrum parameters and operation sequence identifier values, compares overlapping index intervals based on a unified time axis and performs weighting, and generates behavior feature fusion vector values. First, for a selected set of timestamps ranging from... to The system processes a 13-second synchronous segment in which the candidate answers the question, "Please describe your most challenging teamwork experience in a past project." The system extracts the corresponding raw data, and for the semantic text, "In that project, the biggest challenge came from poor cross-departmental communication. I proactively established weekly cross-departmental communication meetings and created a shared project progress dashboard," it calls a large-scale language model with a 512-dimensional output to convert this text into a 512-dimensional semantic embedding vector. The values ​​of its first four dimensions are Meanwhile, the speech signal of this 13-second segment is analyzed to calculate its speech spectrum parameters. Specifically, 13-dimensional Mel-frequency cepstral coefficients (MFCCs) are extracted, and the mean of each dimension is calculated over the entire segment, thus obtaining a 13-dimensional speech feature vector. For example, the values ​​of its first four dimensions Next, the operation log data within this period was retrieved. Since the candidate answered the questions verbally and did not perform any keyboard or mouse operations, the operation sequence identifier value was set to a 512-dimensional zero vector. That is, all elements are 0. Subsequently, based on a comparison of the overlapping index intervals of these three sets of features along a unified time axis, it was confirmed that they all originated from to The interval is then calculated, and weighted fusion is performed. Before weighting, feature vectors of different dimensions need to be mapped to the same dimensional space, which is set to 128 dimensions here. The 512-dimensional feature vectors are then mapped using a mapping matrix. 13-dimensional And 512-dimensional Converted to 128 dimensions respectively , , Next, based on the specific indicator of "communication skills" in the current assessment, we will set the weight of the modality and refer to the job requirements description for the semantic content weight in the communication skills assessment. for Voice performance weight for Operational behavior weights for Finally, the behavioral feature fusion vector is calculated by weighted summation. The first element value, assuming The first element is , The first element is , The first element is ,but This calculation is repeated for all 128 dimensions to generate the final 128-dimensional behavioral feature fusion vector value.

[0038] The multimodal fusion calculation submodule calls the behavioral feature fusion vector value, performs feature weighting combination in the multimodal fusion function according to the competency dimension weight parameter set, calculates the correlation coefficient between different modalities and identifies deviation components, performs proportional adjustment for deviation components, and generates a multimodal matching correlation coefficient set; The system retrieves a 128-dimensional behavioral feature fusion vector representing the candidate's performance on a specific issue. First, based on a pre-defined set of competency dimension weight parameters (as shown in Table 2), it performs a weighted combination of features in the multimodal fusion function. Taking the competency dimension of "logical thinking" as an example, its weight parameters are set such that semantic features account for a significant portion. Speech features account for Operational features account for The calculation here does not directly apply to the fused vectors, but instead calls the 128-dimensional mapping vector of the pre-fusion modality to calculate the L2 norm (i.e., the square root of the sum of squares of the elements in the vector) as the intensity representation of the modality. It is assumed that the calculated semantic strength... Speech intensity Operational intensity Therefore, the preliminary correlation coefficient for the "logical thinking" dimension is: Next, the correlation coefficients between different modalities are calculated to identify deviation components. Specifically, this involves analyzing whether the emotional polarity of the semantic content matches the fluctuation of the speech intonation, based on the semantic vector. The extracted emotion polarity score is (The interval is -1 to 1, indicating very positive content), while from the speech vector The extracted intonation variation is (The range is 0 to 1, representing a flat tone), and the threshold for deviation judgment is set as follows: This threshold was derived by analyzing baseline data from 100 employees and taking the 85th percentile of the absolute value of the difference between emotional polarity and tone fluctuation. The current difference is... This value is greater than Therefore, the speech modality is determined to be a deviation component, and a proportional adjustment is performed on this deviation component. The adjustment rule is to adjust the weight of the deviation component ( Multiply by an attenuation factor The new speech weights are obtained as follows: The weight values ​​that were reduced The original weight ratio of semantics to operations will be used. The weights are redistributed, with the semantic weights increased by . The weight added by the operation is The adjusted weight set is semantic ,voice ,operate The correlation coefficient of "logical thinking" was recalculated using the new weights. Repeat this process for the competency dimension to generate a multimodal matching correlation coefficient set consisting of the final correlation values ​​of the multiple dimensions.

[0039] Table 2: Set of Job Competency Weighting Parameters Competency Dimension Semantic feature weights Speech feature weights Operational feature weights Communication and Collaboration 0.60 0.35 0.05 Logical thinking 0.70 0.10 0.20 Innovation capability 0.65 0.15 0.20 Pressure resistance 0.20 0.50 0.30 As shown in Table 2, this table defines the weight distribution of modal features referenced when evaluating different competency dimensions.

[0040] The competency contribution allocation submodule calculates the cumulative weight of the competency dimension correlation coefficients based on the multimodal matching correlation coefficient set, performs linear conversion based on the difference between the score offset and the benchmark value, and generates competency assessment result data. Based on the generated set of multimodal matching association coefficients, for example, we obtain {communication and collaboration:} Logical thinking: Pressure resistance: First, the cumulative weight of the correlation coefficient for each competency dimension is calculated. In this embodiment, this cumulative weight is the correlation coefficient itself. Next, a linear conversion is performed based on the difference between the score offset and the preset benchmark value for the competency of the position. This benchmark value is set based on statistical analysis of data from more than 100 original successful recruitment cases for this position, taking the 70th percentile of the multi-dimensional scores as the benchmark. Assuming the benchmark value for "logical thinking" is... The benchmark value for "stress resistance" is Taking the "logical thinking" dimension as an example, its score offset is the candidate's correlation coefficient minus the benchmark value, i.e. Then, it is converted into a final percentage score according to an asymmetric linear conversion rule. The rule is as follows: when the score equals the benchmark value, the corresponding final score is 80 points; for every unit the score exceeds the benchmark value, the final score increases by 5 points; for every unit the score falls below the benchmark value, the final score decreases by 8 points. The current candidate's "Logical Thinking" score offset is... Therefore, its increased rating is The final score is [score]. Taking the "stress resistance" dimension as an example, the score offset is... Therefore, its reduced score is The final score is [score]. The scores are calculated for all competency dimensions, ultimately generating a set of competency assessment results data composed of specific scores that can intuitively reflect the candidate's multifaceted ability levels.

[0041] Specifically, such as Figure 2 , 7As shown, the evaluation result generation module includes: The competency data parsing submodule obtains the multi-dimensional raw score sequence from the competency assessment results data, performs weighted and scaled calculations based on weight parameters and normalized benchmark values, compares and corrects the scores with the benchmark interval boundaries and records the differences, and generates a dimensional quantified score set. After obtaining the raw score sequence of multiple dimensions from the competency assessment results data, such as a set containing {communication and collaboration: 87.22, logical thinking: 80.506, stress resistance: 73.6}, the first step is to apply the weight parameters specific to the current job posting, i.e., the weight of communication and collaboration is... Logical thinking is The compressive strength is The scores from multiple items are weighted for the purpose of generating a comprehensive index later. However, the scaling of individual scores uses a uniform normalized benchmark value, which is set to the theoretical minimum score. points and the highest score Score for "Communication and Collaboration" The scaling calculation is performed as (original score - lowest score) / (highest score - lowest score), i.e. The corrected score is Similarly, scores on logical thinking... Processing is performed to obtain The result is Score for ability to withstand pressure Processing is performed to obtain The result is Next, the revised score will be compared with the preset job competency benchmark range, which is based on the top performance of the position within the company. The average ability profile of employees is set, for example, the baseline range for "communication and collaboration" is... "Logical thinking" is The revised score for "communication and collaboration" will be adjusted. and interval The comparison revealed that it lies within the interval; the difference between it and the lower boundary of the interval was recorded as . Correcting the score for "logical thinking" and interval The comparison revealed that it was below the lower boundary of the interval; the difference was recorded as follows. This comparison and difference record is performed on the dimensions, ultimately generating a dimensional quantization score set consisting of scaled values ​​from multiple dimensions.

[0042] The grade classification calculation submodule, based on the dimension-quantified score set, calls the preset grade interval threshold, performs interval comparison and labeling on the dimension scores, calculates the grade ratio and calculates the overall score deviation, and generates a grade classification result set. The preset level interval threshold is determined based on the dimensional quantitative score distribution characteristics and cluster boundaries of candidates at different levels in the original competency assessment sample; The generated dimensional quantification score set, namely {communication and collaboration:} Logical thinking: Pressure resistance: First, a preset threshold table for grade intervals is invoked, as shown in Table 3. This threshold is determined based on a dataset containing 5000 original candidate competency assessment samples. The dataset is used to statistically analyze the multi-dimensional quantitative scores of various categories of candidates who were ultimately hired and whose performance ratings are "Excellent," "Good," and "Average," calculating the distribution characteristics of their scores. For example, the score of an "Excellent" employee... Concentrated The above steps, by finding the boundary with the least overlap in the score distributions of different level groups, ultimately determine the threshold for level division. Next, interval comparisons and label assignments are performed on the current candidate's dimensional scores, assigning the score for "communication and collaboration" to... Compared to the thresholds in Table 3, this value falls within the range of 0.70 or greater and 0.85, and is therefore rated as "Good". The score for "Logical Thinking" will be adjusted accordingly. For comparison, its score falls within the range of 0.55 or higher and less than 0.70, and is therefore labeled as "average." This score for "stress resilience" is... Compared to others, it is lower The samples were labeled as "needs improvement". Then, the percentages of these labels were statistically analyzed to determine the proportions of "good", "average", and "needs improvement". ,Right now Then, the overall score deviation is calculated. First, the weighted average score of the candidates is calculated using the aforementioned weights {0.5, 0.3, 0.2}, i.e. Then compare this score with the entry threshold for the "good" level of the position. By comparing the scores, the overall score deviation was calculated. This ultimately generates a set of grade classification results that includes specific levels and proportions across multiple dimensions.

[0043] Table 3: Threshold Table for Quantitative Score Level Intervals of Competency Dimensions Level Name Score range excellent [0.85，1.00] good [0.70，0.85) generally [0.55，0.70) Needs improvement [0.00，0.55) As shown in Table 3, this table defines the numerical range used to map dimensional quantification scores to specific ability levels.

[0044] The results presentation submodule divides the result set according to the level, calls the dimension scores and confidence interval values, calculates the confidence rating index and matches the text template, integrates the proportional charts and text content, and generates the candidate AI interview assessment results. The confidence interval values ​​are calculated and determined based on the fluctuation range and standard deviation distribution of the dimension scores in multiple evaluation processes; Based on the generated rating set, namely {Communication & Collaboration: Good, Logical Thinking: Average, Stress Resistance: Needs Improvement}, the original percentage scores for multiple dimensions {87.22, 80.506, 73.6} and their corresponding confidence intervals are first retrieved. The confidence intervals are calculated based on the fluctuations in the candidate's scores when answering multiple questions related to the same dimension (e.g., three questions about communication skills). If a candidate's three communication skills scores are... , , Then its standard deviation is calculated as follows: Subsequently based on The confidence level is used to calculate the confidence interval width to determine the "communication and collaboration" score. The confidence interval is Next, the confidence rating index is calculated, and the calculation method for this index is as follows: Subtract the confidence interval width and the score range width ( The ratio of ), i.e. Setting the index higher than For high confidence, in to The range is considered the middle confidence level, and the range is below that. For low confidence, currently The system assigns a medium confidence level. Then, based on the "Good" rating and "Medium" confidence level, it matches a pre-defined text template. For example, it retrieves a description coded "Good-Medium" from the template library: "The candidate demonstrated solid abilities in this dimension, meeting the basic requirements of the position; the stability of the evaluation results is at a moderate level." Subsequently, it integrates the dimension's score, rating, confidence level, and the rating proportion calculated in the preceding steps. The data is used to generate visual charts, such as a radar chart plotting multiple scores and a pie chart displaying the proportions of each level using different colored sectors. Finally, these charts are combined with matching descriptive text to generate a complete AI interview assessment result for the candidate, incorporating both data and textual interpretation.

[0045] Please see Figure 8The competency-based AI interview assessment report generation method is executed based on the aforementioned competency-based AI interview assessment report generation system, and includes the following steps: S1: By collecting image, voice and operational behavior signals during the candidate's interview process, the time offset of the channel timestamp data is calculated, and the error is compensated based on the dynamic time warping algorithm to generate a multimodal synchronous dataset; S2: Receive multimodal synchronous dataset, use multi-channel label embedding function to divide the data into fixed windows and mark time correction information, extract facial expression, tone change and reaction time parameters, and generate multidimensional feature encoding vector; S3: Calculate the time window similarity coefficient of the multidimensional feature encoding vector, filter out time window features below the preset similarity threshold, retain the time period when the consistency between facial expression changes and tone changes exceeds the threshold, and generate a competency-related behavioral feature set. S4: Call the competency-related behavioral feature set, use a multimodal fusion function to match the behavioral features with the competency dimension weights, assign contribution scores, and generate competency assessment result data; S5: Based on the competency assessment results, calculate the grade division according to the quantitative scores of the competency dimensions, convert the scores into four grade standards: excellent, good, average, and need improvement, and combine the confidence rating analysis with textual descriptions and charts to construct the candidate AI interview assessment results.

[0046] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An AI-based interview assessment report generation system based on a competency model, characterized in that, The system includes: The multimodal data synchronization module collects image, voice, and operational behavior signals during the candidate's interview process, calculates the time offset of the channel timestamp data, compensates for errors based on the dynamic time warping algorithm, generates a multimodal synchronization dataset, and transmits it to the feature encoding module. The feature encoding module receives the multimodal synchronous dataset, uses a multi-channel identifier embedding function to divide the data into fixed windows and mark time correction information, extracts facial expression, tone change and reaction time parameters, generates a multi-dimensional feature encoding vector, and transmits it to the behavior pattern screening module. The behavior pattern filtering module calculates the time window similarity coefficient of the multidimensional feature encoding vector, filters out time window features below a preset similarity threshold, retains time periods where the consistency between facial expression changes and tone changes exceeds a threshold, generates a competency-related behavior feature set, and passes it to the competency feature mapping module. The competency feature mapping module calls the competency-related behavioral feature set, uses a multimodal fusion function to match behavioral features with competency dimension weights, assigns contribution scores, generates competency assessment result data, and transmits it to the assessment result generation module.

2. The AI ​​interview assessment report generation system based on a competency model according to claim 1, characterized in that, The multimodal synchronization dataset includes a time synchronization index, a modality mapping matrix, and synchronization accuracy parameters. The multidimensional feature encoding vector includes facial expression features, speech prosody features, and reaction time features. The competency-related behavioral feature set includes key behavioral segments, emotion and tone consistency labels, and behavioral pattern feature identifiers. The competency assessment result data includes dimensional result distribution, weight contribution coefficients, and level mapping indicators.

3. The AI ​​interview assessment report generation system based on a competency model according to claim 1, characterized in that, The multimodal data synchronization module includes: The channel signal acquisition submodule acquires image, voice and operation behavior signals during the candidate interview process, detects the stability of the sampling port input and records the channel identification code, performs time segmentation and frame division on the signal sequence according to the sampling frequency, calculates and sorts the frame start and end timestamps, and establishes a channel synchronization index set. The time offset calculation submodule, based on the channel synchronization index set, extracts the time interval between adjacent channels for the timestamp sequences of image, voice and operation behavior channels, calculates the interval difference and establishes a set of difference intervals, detects the extreme points of the offset boundary and performs weighted calculation to obtain the channel time offset coefficient. The dynamic time warping compensation submodule calls the channel time offset coefficient, performs synchronous adjustment on the channel signal time series based on the dynamic time warping algorithm, calculates the error interval based on the matching degree between signals and extracts boundary samples, performs interpolation calculation and sequence rearrangement on the error interval, and generates a multimodal synchronization dataset.

4. The AI ​​interview assessment report generation system based on a competency model according to claim 3, characterized in that, The feature encoding module includes: The synchronous data processing submodule acquires facial images, voice waves, and reaction time records from the multimodal synchronous dataset, compares the data according to the signal timestamp order, repairs missing frames using linear interpolation, reconstructs the index according to the channel identifier, unifies the sampling frequency, and generates a multimodal synchronous signal sequence set. The timing synchronization segmentation submodule, based on the multi-modal synchronization signal sequence set, uses a multi-channel identifier embedding function to encode the channel sequence, calculates the signal amplitude change rate and sets the sampling window length, and corrects the offset based on the reference difference to obtain the time-corrected segmented signal set; The feature vector extraction submodule calculates the amplitude of facial expression changes in the facial image channel, the rate of change of tone in the speech channel, and the delay interval in the reaction time channel based on the time-corrected segmented signal set. The three types of features are then concatenated over time to generate a multidimensional feature encoding vector.

5. The AI ​​interview assessment report generation system based on a competency model according to claim 4, characterized in that, The behavior pattern filtering module includes: The feature similarity calculation submodule obtains the facial expression change and tone change parameters of the time window in the multidimensional feature encoding vector sequence, calculates the difference of vectors of the same dimension and performs Euclidean distance operation and normalization, and calculates the similarity based on the normalized distance to generate a time window similarity coefficient set. The time window filtering submodule, based on the time window similarity coefficient set, marks time windows below a preset similarity threshold and retains the remaining windows. It calculates the difference between the rate of change of facial expression and the rate of change of tone in the retained windows and divides the intervals according to the continuity of the difference to obtain a consistency filtering interval sequence. The behavior consistency extraction submodule calls the consistency filtering interval sequence, performs feature aggregation operation on the multi-dimensional feature encoding vector within the interval, calculates the cross-correlation degree of facial expression change feature quantity and tone change feature quantity, and filters the synchronization segment according to the correlation degree threshold to generate a competency-related behavior feature set.

6. The AI ​​interview assessment report generation system based on a competency model according to claim 5, characterized in that, The preset similarity threshold is determined based on the average Euclidean distance distribution characteristics between facial expression changes and tone changes of different individuals under the same task conditions; The correlation threshold is determined based on the statistical distribution characteristics of the cross-correlation between facial expression change features and tone change features in the original sample.

7. The AI ​​interview assessment report generation system based on a competency model according to claim 5, characterized in that, The competency feature mapping module includes: The behavior feature extraction submodule acquires semantic text, speech signals and operation log data from the competency-related behavior feature set, calculates semantic embedding vectors, speech spectrum parameters and operation sequence identifier values, compares overlapping index intervals based on a unified time axis and performs weighting, and generates behavior feature fusion vector values. The multimodal fusion calculation submodule calls the behavioral feature fusion vector value, performs feature weighting combination in the multimodal fusion function according to the competency dimension weight parameter set, calculates the correlation coefficient between different modalities and identifies deviation components, performs proportional adjustment for deviation components, and generates a multimodal matching correlation coefficient set. The competency contribution allocation submodule calculates the cumulative weight of the competency dimension correlation coefficients based on the multimodal matching correlation coefficient set, performs linear conversion based on the difference between the score offset and the benchmark value, and generates competency assessment result data.

8. The AI ​​interview assessment report generation system based on a competency model according to claim 1, characterized in that, The assessment result generation module, based on the competency assessment result data, performs a level division calculation according to the quantitative score of the competency dimension, converts the score into four level standards: excellent, good, average, and need improvement, and combines confidence rating analysis with text description and chart display to construct the candidate AI interview assessment result. The candidate AI interview assessment results include competency level conclusions, confidence level evaluation results, and result visualization content.

9. The AI ​​interview assessment report generation system based on a competency model according to claim 8, characterized in that, The evaluation result generation module includes: The competency data parsing submodule obtains the multi-dimensional original score sequence from the competency assessment result data, performs weighted and scaled calculations based on weight parameters and normalized benchmark values, compares and corrects the scores with the benchmark interval boundaries and records the differences, and generates a dimensional quantified score set. The grade division calculation submodule, based on the dimensional quantized score set, calls the preset grade interval threshold, performs interval comparison and labeling on the dimensional scores, calculates the grade ratio and overall score deviation, and generates a grade division result set. The results presentation submodule, based on the graded result set, calls the dimension scores and confidence interval values, calculates the confidence rating index and matches the text template, integrates the proportional charts and text content, and generates the candidate AI interview assessment results.

10. A method for generating AI interview assessment reports based on a competency model, characterized in that, The AI ​​interview assessment report generation system based on a competency model, as described in any one of claims 1-9, comprises the following steps: S1: By collecting image, voice and operational behavior signals during the candidate's interview process, the time offset of the channel timestamp data is calculated, and the error is compensated based on the dynamic time warping algorithm to generate a multimodal synchronous dataset; S2: Receive the multimodal synchronous dataset, use a multi-channel identifier embedding function to divide the data into fixed windows and mark time correction information, extract facial expression, tone change and reaction time parameters, and generate a multi-dimensional feature encoding vector; S3: Calculate the time window similarity coefficient of the multidimensional feature encoding vector, filter out time window features below the preset similarity threshold, retain the time period when the consistency between facial expression changes and tone changes exceeds the threshold, and generate a competency-related behavior feature set. S4: Call the competency-related behavioral feature set, use a multimodal fusion function to match behavioral features with competency dimension weights, assign contribution scores, and generate competency assessment result data; S5: Based on the competency assessment results, perform a grade division calculation according to the quantitative scores of the competency dimensions, convert the scores into four grade standards: excellent, good, average, and need improvement, and combine the confidence rating analysis with textual descriptions and charts to construct the candidate AI interview assessment results.

Images