An intelligent employment matching system and method based on a large language model

Abstract

The application relates to the technical field of language processing, and particularly discloses an intelligent employment matching system and method based on a large language model, which comprises an acquisition module, a determination module, a response module and a matching module. The acquisition module acquires multi-source dialogue data, cognitive level data, a post knowledge database and interview audio data of an interviewee, and determines interview question data and interview answer data. The determination module determines a plurality of deep question sets, a plurality of non-deep questions, knowledge-cognition data of each question number, post demand matching data of each answer number and each split number. The response module determines a time sequence response sequence of each post demand feature. The matching module determines an attenuation label, a semantic label of each deep question set or non-deep question and an interview evaluation report of the interviewee. The application improves the question and answer pairing accuracy of multi-role interview audio, quantifies the semantic attenuation trend in the dialogue process, and improves the semantic correlation recognition accuracy of the deep follow-up question chain.

Description

Technical Field

[0001] This invention relates to the field of language processing technology, and in particular to an intelligent job matching system and method based on a large language model. Background Technology

[0002] Early job matching relied on resume keyword retrieval, achieving only static text comparison. Subsequent AI interview assessments primarily depended on pre-set fixed question banks, mostly employing single-point static scoring, calculating matching scores based solely on individual responses, lacking temporal dimension analysis. While recent years have seen the emergence of recruitment recommendation systems based on large language models, and some solutions have incorporated knowledge graphs for knowledge processing of resume information, a job knowledge specification system adapted to dynamic interview processes based on ontology has yet to be built. The following technical shortcomings exist: the dialogue roles between interviewers and interviewees during the interview process rely on manual annotation or pre-set templates, failing to achieve real-time automatic separation; the assessment of interviewees' answers relative to job requirements is based solely on static matching of single text segments, lacking tracking and analysis of the semantic evolution trends within the answers, resulting in assessment results that fail to reflect the logical depth and temporal response stability of interviewees under pressure from in-depth questioning.

[0003] Therefore, this invention proposes an intelligent employment matching system and method based on a large language model. Summary of the Invention

[0004] This invention provides an intelligent job matching system and method based on a large language model. It acquires multi-source dialogue data, constructs a question-and-answer recognition model, obtains cognitive level data and a job knowledge database for required positions, acquires interview audio data of interviewees, determines interview question data and interview answer data, identifies multiple sets of deep questions and multiple sets of non-deep questions, determines the knowledge-cognitive data for each question number in the interview question data and the job requirement matching data for each split number of each answer number in the interview answer data, determines the temporal response sequence and temporal response trend graph for each job requirement feature, determines the decay label and semantic label for each set of deep questions or non-deep questions, and determines the interviewee's interview evaluation report. This improves the accuracy of question-and-answer matching for multi-role interview audio, achieves temporal continuous detection of semantic matching degree and automatic location of abrupt change points, quantifies the semantic decay trend during the dialogue process, and improves the semantic association recognition accuracy of deep questioning chains.

[0005] This invention provides an intelligent job matching system based on a large language model, comprising: Acquisition Module: Acquire multi-source dialogue data, build a question-and-answer recognition model, acquire cognitive level data and job knowledge database for required positions, acquire and determine interview questions and answers based on interviewee audio data; The module identifies multiple sets of in-depth questions and multiple sets of non-in-depth questions based on the interviewers' interview questions, interview answers, cognitive level data, and job knowledge database. It also identifies the knowledge-cognitive data for each question number in the interview question data and the job requirement matching data for each split number of each answer number in the interview answer data. Response module: Based on cognitive level data, knowledge-cognitive data of all question numbers, and job requirement matching data of all split numbers of all answer numbers, determine the time-series response sequence and time-series response trend chart of each job requirement feature; Matching module: Based on the time-series response sequence and time-series response trend chart of all job requirements, determine the decay label and semantic label of each set of deep questions or non-deep questions, and generate the interview evaluation report for the interviewee.

[0006] Preferably, an intelligent job matching system based on a large language model includes an acquisition module comprising: Multi-source dialogue data unit: Acquire multi-source dialogue data, which includes dialogue text data of multi-turn dialogues. The dialogue text data includes the dialogue feature vector and text labels of each dialogue text. The text labels include questions and statements. Question-answering recognition model unit: Constructing a question-answering recognition model based on multi-source dialogue data; Cognitive level data unit: Acquire cognitive level data, which includes the cognitive mapping value of each cognitive level and a set of level questions, where the set of level questions includes multiple question descriptions; Job Knowledge Database Unit: Obtain the job knowledge database for the required job position. The job knowledge database includes multiple knowledge points, as well as the knowledge point data and knowledge terminology set for each knowledge point.

[0007] Preferably, an intelligent job matching system based on a large language model includes an acquisition module that further comprises: Interview Recognition Data Unit: Acquires the interviewee's interview audio data, inputs the interview audio data into the voiceprint recognition model, and determines the interview recognition data. The interview recognition data includes multiple text fragment data and the speech number, text sequence number, and text time interval of each text fragment data. Speech text data unit: All text segments with the same speech number and adjacent text sequence numbers in the interview recognition data are merged to determine the speech text data. The speech text data includes multiple speech text sub-data and the speech number, speech sequence number, and text time interval of each speech text sub-data. Speech Feature Vector Unit: Extracts features from each speech text sub-data in the speech text data to determine the speech feature vector of each speech text sub-data in the speech text data; Text label unit: Input the speech feature vectors of all speech text sub-data in the speech text data into the question answering recognition model to determine the text label of each speech text sub-data in the speech text data; Speech Tag Unit: Based on the text tags of all speech text sub-data with the same speech number in the speech text data, determine the speech tag for each speech number, where the speech tag includes interviewee and interviewer; Question text data unit: Merge all speech text sub-data with the speech tag "interviewer" and adjacent speech sequence numbers in the speech text data to determine the question text data. The question text data includes multiple question text sub-data and the question time interval of each question text sub-data. Answer text sub-data unit: Each speech text sub-data unit in the speech text data with the speech tag of the interviewee is determined as the answer text sub-data unit, and the text time interval of the speech text sub-data unit is determined as the answer time interval of the answer text sub-data unit; Answer Numbering Unit: Based on the question time interval of all question text sub-data and the answer time interval of all answer text sub-data, determine the question number of each question text sub-data and the answer number of each answer text sub-data. Interview Response Data Unit: Based on the question text sub-data of all question numbers and the question time range, determine the interview question data of the interviewee; based on the answer text sub-data of all answer numbers and the answer time range, determine the interview response data of the interviewee.

[0008] Preferably, an intelligent job matching system based on a large language model includes a determination module, comprising: Word segmentation unit: Perform word segmentation on the sub-data of the question text for each question number in the interview question data, and determine the set of question words for each question number and the word position number of each word in the set of question words; The first matching unit: Match each word in the question word set for each question number in the interview question data with the knowledge term set of all knowledge points in the job knowledge database to determine the question knowledge set for each question number and the knowledge point position number set for each knowledge point in the question knowledge set; Judgment unit: Determines whether the same knowledge point exists in the question knowledge set of multiple adjacent question numbers. If it exists, determine that the multiple question numbers are a deep question set. If it does not exist, determine that each question number is a non-deep question. The second matching unit: matches each word in the question word set for each question number with the hierarchical question set for each cognitive level in the cognitive level data to determine the question cognitive set for each question number and the hierarchical position number of each cognitive level in the question cognitive set; Knowledge-cognitive data unit: Based on the question text sub-data of each question number, the question knowledge set, the knowledge point position number set of each knowledge point in the question knowledge set, the question cognition set, and the level position number of each cognition level in the question cognition set, the knowledge-cognitive data of each question number is determined. The knowledge-cognitive data includes multiple knowledge-cognitive tuples, and each knowledge-cognitive tuple includes at least one or more cognition levels. Set the number of words to be segmented: Based on the answer text sub-data of each answer number, determine the set number of words to be segmented for each answer number; Segmentation Unit: Based on the set number of words to be segmented and sentence boundaries for each answer number, the answer text sub-data for each answer number is segmented to determine multiple segmented text data for each answer number, as well as the segmentation number, segmentation time interval, and segmentation midpoint timestamp for each segmented text data. Job requirement matching data unit: The split text data of each split number of each answer number is sequentially input into the job requirement model of the job position to determine the job requirement matching data of each split number of each answer number. The job requirement matching data includes feature matching values ​​of multiple job requirement features.

[0009] Preferably, an intelligent job matching system based on a large language model includes a response module comprising: Calculation unit: Based on the cognitive mapping values ​​of all cognitive levels in the cognitive hierarchy data, the knowledge-cognitive data of each question number, and the feature matching value of each job requirement feature in the job requirement matching data of each split number of each answer number, calculates the corrected matching value of the job requirement feature; Time-series response sequence unit: Based on the midpoint timestamp of the splitting midpoint of all splitting numbers of all answer numbers and the corrected matching value of each job requirement feature, determine the time-series response sequence of each job requirement feature and draw the time-series response trend chart of each job requirement feature, where the horizontal axis of the time-series response trend chart is time and the vertical axis is the corrected matching value.

[0010] Preferably, an intelligent job matching system based on a large language model includes a matching module comprising: Mutation timestamp unit: Perform Pettitt mutation detection on the time-series response sequence of each job requirement feature to determine the mutation timestamps of multiple negative mutation points for each job requirement feature; Mutation Feature Set Unit: Determine whether there are any questions whose midpoint timestamps corresponding to the mutation timestamps of negative mutation points of a preset number of job requirement features belong to the same depth question set or non-depth question set. If so, determine that all job requirement features are the mutation feature set of the depth question set or non-depth question set. Demand decay delay unit: Based on the mutation timestamp of the first mutation point of all job demand features in each set of deep questions or set of mutation features of non-deep questions, determine the earliest mutation timestamp and the latest mutation timestamp, and calculate the difference between the latest mutation timestamp and the earliest mutation timestamp as the demand decay delay of each set of deep questions or set of non-deep questions for candidates. Attenuation label unit: If the demand attenuation delay of the deep question set or non-deep question is less than the first preset delay threshold, the attenuation label of the deep question set or non-deep question is determined to be synchronous attenuation; if the demand attenuation delay of the deep question set or non-deep question is greater than the second preset delay threshold, the attenuation label of the deep question set or non-deep question is determined to be progressive attenuation; otherwise, the attenuation label of the deep question set or non-deep question is determined to be unattenuated.

[0011] Preferably, an intelligent job matching system based on a large language model includes a matching module comprising: First feature extraction unit: Performs feature extraction on the split text data of the first and last split numbers of the question number corresponding to each non-deep question, and determines the first semantic vector and the second semantic vector of each non-deep question; First semantic labeling unit: Calculate the cosine similarity value of the first semantic vector and the second semantic vector of each non-deep question. If the cosine similarity value is less than the first preset similarity threshold, determine that the semantic label of the non-deep question is semantically drifted; otherwise, determine that the semantic label of the non-deep question is semantically consistent. The second feature extraction unit extracts features from the split text data of the first split number of each question number in each deep question set to determine the first semantic vector of each question number in each deep question set. Semantic decay sequence unit: If the number of question numbers in the deep question set is greater than 2, calculate the cosine similarity value of the first semantic vector of the first question number in the deep question set and the first semantic vector of each question number other than the first question number, determine the semantic similarity value of each question number in the deep question set other than the first question number, calculate the difference of the semantic similarity of all adjacent question numbers in the deep question set, and determine the semantic decay sequence of the deep question set. Second semantic labeling unit: If the semantic decay sequence of the deep question set continues to decrease and the time-series response sequence of the job requirement feature as knowledge accuracy has a negative mutation point in the deep question set, the semantic label of the deep question set is determined to be semantic drift; otherwise, the semantic label of the deep question set is determined to be semantic consistency. Third semantic labeling unit: If the number of question numbers in the deep question set is equal to 2, calculate the cosine similarity value of the first semantic vectors of the two question numbers in the deep question set. If the cosine similarity value is less than the second preset similarity threshold and the time-series response sequence of the job requirement feature is knowledge accuracy has a negative mutation point in the deep question set, determine that the semantic label of the deep question set is semantic drift; otherwise, determine that the semantic label of the deep question set is semantic consistency. Interview Evaluation Report Unit: Based on the decay labels and semantic labels of all deep questions, the decay labels and semantic labels of all non-deep questions, and the time-series response trend chart of all job requirement characteristics, the interview evaluation report of the interviewee is determined.

[0012] This invention provides an intelligent job matching method based on a large language model, used to execute any one of the intelligent job matching systems based on a large language model in Examples 1 to 7, comprising: S1: Acquire multi-source dialogue data, construct a question-and-answer recognition model, acquire cognitive level data and job knowledge database for the required positions, acquire and determine the interview questions and answers based on the interviewee's interview audio data; S2: Based on the interview questions, interview answers, cognitive level data, and job knowledge database, determine multiple sets of in-depth questions and multiple sets of non-in-depth questions, and determine the knowledge-cognitive data for each question number in the interview question data and the job requirement matching data for each split number of each answer number in the interview answer data; S3: Based on cognitive level data, knowledge-cognitive data of all question numbers, and job requirement matching data of all split numbers of all answer numbers, determine the time-series response sequence and time-series response trend chart of each job requirement feature; S4: Based on the time-series response sequence and time-series response trend chart of all job requirements, determine the decay label and semantic label of each set of deep questions or non-deep questions, and determine the interview evaluation report of the interviewee.

[0013] The beneficial effects of this invention compared to existing technologies are as follows: It acquires multi-source dialogue data, constructs a question-and-answer recognition model, acquires cognitive level data and a job knowledge database for required positions, acquires interview audio data of interviewees, determines interview question data and interview answer data, identifies multiple sets of deep questions and multiple non-deep questions, determines the knowledge-cognitive data for each question number in the interview question data and the job requirement matching data for each split number of each answer number in the interview answer data, determines the temporal response sequence and temporal response trend graph for each job requirement feature, determines the decay label and semantic label for each set of deep questions or non-deep questions, and determines the interviewee's interview evaluation report. This improves the accuracy of question-and-answer matching in multi-role interview audio, achieves temporal continuous detection of semantic matching degree and automatic location of abrupt change points, quantifies the semantic decay trend during the dialogue process, and improves the semantic association recognition accuracy of deep questioning chains. This invention relies on ontology to standardize the knowledge system architecture for job positions, solving the problems of scattered knowledge points and chaotic hierarchy in traditional knowledge bases; it completes the structured transformation of interview texts and job-related professional information through standardized knowledge processing methods throughout the entire process; and it realizes knowledge point association retrieval and cross-cognitive level matching based on knowledge graphs, further enhancing the integrity of question-and-answer semantic parsing.

[0014] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] 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: Figure 1 This is a schematic diagram of an intelligent job matching system based on a large language model, as described in an embodiment of the present invention. Figure 2 This is a flowchart of an intelligent employment matching method based on a large language model, as described in an embodiment of the present invention. Detailed Implementation

[0017] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention. Example 1:

[0018] This invention provides an intelligent job matching system based on a large language model, with reference to... Figure 1 ,include: Acquisition Module: Acquire multi-source dialogue data, build a question-and-answer recognition model, acquire cognitive level data and job knowledge database for required positions, acquire and determine interview questions and answers based on interviewee audio data; The module identifies multiple sets of in-depth questions and multiple sets of non-in-depth questions based on the interviewers' interview questions, interview answers, cognitive level data, and job knowledge database. It also identifies the knowledge-cognitive data for each question number in the interview question data and the job requirement matching data for each split number of each answer number in the interview answer data. Response module: Based on cognitive level data, knowledge-cognitive data of all question numbers, and job requirement matching data of all split numbers of all answer numbers, determine the time-series response sequence and time-series response trend chart of each job requirement feature; Matching module: Based on the time-series response sequence and time-series response trend chart of all job requirements, determine the decay label and semantic label of each set of deep questions or non-deep questions, and generate the interview evaluation report for the interviewee.

[0019] In this embodiment, multi-source dialogue data refers to question-and-answer interaction records collected from different channels, including historical interview recordings, mock interviews, and publicly available question-and-answer corpora. The question-and-answer recognition model is a binary classification neural network trained using these dialogue data labeled with questions and statements, capable of automatically determining whether a text is a question or an answer. Cognitive hierarchy data is a predefined six-level cognitive ability hierarchy based on Bloom's Taxonomy of Cognitive Objectives, including memory, comprehension, application, analysis, evaluation, and creation. Each level includes a cognitive mapping value and a set of questions for that level. The job knowledge database is a structured knowledge repository built for the target recruitment position, containing knowledge elements such as professional terminology, technical concepts, and business processes required for the position. After acquiring the interviewee's interview audio data, the system automatically parses the original audio into structured interview question data and interview answer data using voiceprint recognition and speech segmentation technology.

[0020] In this embodiment, the determination module undertakes the task of converting raw dialogue into structured semantic information. Based on the interviewer's question and answer data, the system performs dual analysis on each question. The first analysis matches the words in the question with the knowledge terms in the job knowledge database to determine which knowledge points the question involves. The second analysis matches the words in the question with the typical question word set for each level in the cognitive level data to determine the cognitive level required by the question. Combining these two aspects of information, the system divides interview questions into two categories: deep question set and non-deep question set. The deep question set refers to a sequence of questions that involve multiple rounds of continuous follow-up questions around the same knowledge point, while the non-deep question set consists of independent questions without any follow-up relationship. For each question number, the system determines its knowledge-cognitive data, accurately recording which knowledge point the question requires and which cognitive level the answer to. For each answer number, the system splits the answer text into multiple semantically complete segments according to sentence boundaries and a preset length. Each segment is input into the job requirement model for evaluation, and the system outputs feature matching values ​​for each job requirement feature.

[0021] In this embodiment, the response module elevates the process from static scores to dynamic trends. Based on the cognitive mapping values ​​of each level in the cognitive hierarchy data, the knowledge cognitive data for each question number, and the job requirement matching data for each segment number of each answer number, the system performs two core processes. The first process corrects the original feature matching values ​​of each job requirement feature, making the scores of questions of different difficulties comparable on the same scale. The second process arranges all answer segments according to the chronological order of the midpoint timestamps, uses the corrected matching value of each segment as the ordinate, and the interview timeline as the abscissa to generate a time-series response sequence for each job requirement feature and plots a time-series response trend graph. This allows the dynamic changes in the interviewee's ability relative to the job requirements as the interview progresses to be fully presented.

[0022] In this embodiment, the matching module is the final decision output layer. Based on the time-series response sequences and trend charts of all job requirement features, the system performs attenuation labeling, semantic labeling, and interview evaluation report generation. Attenuation labeling uses the Pettitt mutation detection algorithm to locate negative mutation points in the time-series response sequences of each job requirement, i.e., the time points where the job requirement score significantly decreases. When multiple job requirement features simultaneously exhibit negative mutations within the same questioning time period, the system calculates the distribution span of these mutations on the time axis, i.e., the requirement attenuation delay. If the attenuation delay is less than a first preset delay threshold, it is marked as synchronous attenuation, meaning the interviewee experienced comprehensive cognitive overload during the question, causing all abilities to fail simultaneously. If the attenuation delay is greater than a second preset delay threshold, it is marked as progressive attenuation, meaning each job requirement feature disintegrates sequentially over time. If the mutation condition is not met or the delay is between the two thresholds, it is marked as no attenuation. Semantic labels determine whether the answer accurately addresses the core of the question and whether there are behavioral patterns such as thought loops or horizontal avoidance drift. By combining the decay labels and semantic labels of all questions with the temporal performance of the requirements of each position, the system finally determines the interview evaluation report of the interviewee, which is a comprehensive matching conclusion between the interviewee and the target position in multiple dimensions such as knowledge, skills, cognitive ability and thinking quality.

[0023] The beneficial effects of the above technologies are as follows: acquiring multi-source dialogue data, constructing a question-and-answer recognition model, acquiring cognitive level data and a job knowledge database for required positions, acquiring interview audio data of interviewees, determining interview question data and interview answer data, identifying multiple sets of deep questions and multiple non-deep questions, determining the knowledge-cognitive data for each question number in the interview question data and the job requirement matching data for each split number of each answer number in the interview answer data, determining the temporal response sequence and temporal response trend graph for each job requirement feature, determining the decay label and semantic label for each set of deep questions or non-deep questions, and determining the interviewee's interview evaluation report. This improves the accuracy of question-and-answer matching in multi-role interview audio, achieves temporal continuous detection of semantic matching degree and automatic location of abrupt change points, quantifies the semantic decay trend during the dialogue process, and improves the semantic association recognition accuracy of deep questioning chains. Example 2:

[0024] Based on Example 1, an intelligent job matching system based on a large language model includes an acquisition module, comprising: Multi-source dialogue data unit: Acquire multi-source dialogue data, which includes dialogue text data of multi-turn dialogues. The dialogue text data includes the dialogue feature vector and text labels of each dialogue text. The text labels include questions and statements. Question-answering recognition model unit: Constructing a question-answering recognition model based on multi-source dialogue data; Cognitive level data unit: Acquire cognitive level data, which includes the cognitive mapping value of each cognitive level and a set of level questions, where the set of level questions includes multiple question descriptions; Job Knowledge Database Unit: Obtain the job knowledge database for the required job position. The job knowledge database includes multiple knowledge points, as well as the knowledge point data and knowledge terminology set for each knowledge point.

[0025] In this embodiment, multi-source dialogue data refers to dialogue interaction records collected from different channels and scenarios. These scenarios may include past interview recordings, mock interview drills, and publicly available question-and-answer dialogue data. Multi-turn dialogue text data refers to the text content converted from the original speech through speech recognition. This text content is organized according to the chronological order of the dialogue, preserving the contextual relationships between questions and answers. Dialogue feature vectors are fixed-length numerical vector representations of each dialogue text segment converted through a pre-trained language model. Text labels are the manually annotated results for each dialogue text segment, indicating whether the text is a question or a statement. This annotation provides a supervised learning training objective for the subsequent question-and-answer recognition model.

[0026] In this embodiment, the question-answering recognition model unit is responsible for automatically separating questions and answers from the original dialogue. Building a question-answering recognition model based on multi-source dialogue data involves training a deep neural network classifier using the large-scale dialogue data with question and statement labels obtained in the previous step. The model's input is a dialogue text and its corresponding dialogue feature vector, and its output is the probability value of whether the text belongs to a question or a statement. During training, the model learns the differences between questions and statements in terms of lexical composition, sentence structure, use of modal particles, and contextual position through a large number of labeled samples. After construction, the model can automatically label each segment of dialogue text transcribed from interview audio as either a question or a statement.

[0027] In this embodiment, the cognitive hierarchy data unit provides a benchmark system for assessing the depth of thinking. Cognitive hierarchy data refers to predefined cognitive ability grading standards, typically based on Bloom's Taxonomy of Cognitive Objectives in educational psychology, dividing cognitive abilities into six levels from low to high: memory, comprehension, application, analysis, evaluation, and creation. The cognitive mapping value for each cognitive level is a numerical weighting coefficient used to quantify the corresponding depth of thinking score; for example, the mapping value for the memory level is lower, while the mapping value for the creation level is higher. The hierarchical question set refers to the set of typical questioning methods and guiding vocabulary corresponding to each cognitive level. Each level's question set contains multiple question descriptions, which specifically describe how questions are typically posed at that cognitive level. For example, the set of questions for remembering cognitive levels includes what, definition, concept, enumeration, types, and introduction; the set of questions for understanding cognitive levels includes meaning, how to understand, and explanation; the set of questions for applying cognitive levels includes how to do it, how to implement it, and how to deploy it; the set of questions for analyzing cognitive levels includes why, the reasons for its occurrence, differences, comparisons, and bottlenecks; the set of questions for evaluating cognitive levels includes advantages and disadvantages, risks, suitable scenarios, and how to evaluate it; and the set of questions for creating cognitive levels includes how to optimize, how to design, how to build a solution, and how to restructure the thinking.

[0028] In this embodiment, the job knowledge database for the required position is a pre-built structured knowledge repository for a specific recruitment position, covering all professional knowledge areas required for that position. The job knowledge database includes multiple knowledge points, each representing an independent knowledge unit that the position requires mastery. The knowledge point data for each knowledge point refers to its detailed description, definition, and related background information. The knowledge terminology set for each knowledge point refers to a list of keywords used to identify and retrieve that knowledge point; these terms can be professional terms, technical terms, or common conceptual expressions.

[0029] The beneficial effects of the above technologies are: acquiring multi-source dialogue data, constructing a question-and-answer recognition model, and obtaining cognitive level data and a job knowledge database for required positions. This can improve the accuracy of question-and-answer matching in multi-role free-flowing dialogue scenarios, and provide data support for determining the knowledge-cognitive data for each question number in the interview question data and the job requirement matching data for each split number of each answer number in the interview answer data. Example 3:

[0030] Based on Example 1, an intelligent job matching system based on a large language model, including an acquisition module, further includes: Interview Recognition Data Unit: Acquires the interviewee's interview audio data, inputs the interview audio data into the voiceprint recognition model, and determines the interview recognition data. The interview recognition data includes multiple text fragment data and the speech number, text sequence number, and text time interval of each text fragment data. Speech text data unit: All text segments with the same speech number and adjacent text sequence numbers in the interview recognition data are merged to determine the speech text data. The speech text data includes multiple speech text sub-data and the speech number, speech sequence number, and text time interval of each speech text sub-data. Speech Feature Vector Unit: Extracts features from each speech text sub-data in the speech text data to determine the speech feature vector of each speech text sub-data in the speech text data; Text label unit: Input the speech feature vectors of all speech text sub-data in the speech text data into the question answering recognition model to determine the text label of each speech text sub-data in the speech text data; Speech Tag Unit: Based on the text tags of all speech text sub-data with the same speech number in the speech text data, determine the speech tag for each speech number, where the speech tag includes interviewee and interviewer; Question text data unit: Merge all speech text sub-data with the speech tag "interviewer" and adjacent speech sequence numbers in the speech text data to determine the question text data. The question text data includes multiple question text sub-data and the question time interval of each question text sub-data. Answer text sub-data unit: Each speech text sub-data unit in the speech text data with the speech tag of the interviewee is determined as the answer text sub-data unit, and the text time interval of the speech text sub-data unit is determined as the answer time interval of the answer text sub-data unit; Answer Numbering Unit: Based on the question time interval of all question text sub-data and the answer time interval of all answer text sub-data, determine the question number of each question text sub-data and the answer number of each answer text sub-data. Interview Response Data Unit: Based on the question text sub-data of all question numbers and the question time range, determine the interview question data of the interviewee; based on the answer text sub-data of all answer numbers and the answer time range, determine the interview response data of the interviewee.

[0031] In this embodiment, acquiring the interviewee's audio data refers to collecting the raw speech signal generated throughout the interview process using a recording device. This signal contains all the words spoken by the interviewer and the interviewee along the interview timeline. Inputting the interview audio data into the voiceprint recognition model means feeding this raw audio into a pre-trained voiceprint recognition neural network. This model can distinguish different speakers based on the unique acoustic features of each person's voice, such as timbre, formants, and prosodic patterns. After processing, the voiceprint recognition model outputs interview recognition data, which includes multiple text segment data. These segments are short speech fragments obtained by segmenting continuous audio according to speech activity detection and then transcribing them into text using speech recognition. Each text segment data is accompanied by three key attributes: a speech number to identify which speaker said the segment, a text sequence number to identify the sequence of the segment in the entire interview, and a text time interval to record the start and end times of the segment in the raw audio.

[0032] In this embodiment, the speech text data unit completes the merging of adjacent segments from the same speaker. Since voiceprint recognition and speech segmentation often divide a complete statement from the same speaker into multiple short segments, these fragmented segments need to be reassembled into complete semantic units. Merging all text segments with the same speech number and adjacent text sequence numbers in the interview recognition data means that the system traverses all text segments. When two consecutive segments are found to belong to the same speech number and have adjacent sequence numbers, the text content of these two segments is concatenated in chronological order, and the end of the time interval of the latter segment is used as the end of the time interval of the merged text. After merging, the speech text data is determined, which includes multiple speech text sub-data, each representing a complete continuous speech by a single person. Each speech text sub-data includes a speech number indicating who is speaking, a speech sequence number indicating the sequential position of the speech in the entire interview, and a text time interval indicating the start and end times of the speech in the original audio.

[0033] In this embodiment, the speech feature vector unit realizes the conversion from text to numerical vector. Feature extraction for each sub-data piece of the speech text data involves feeding each complete speech text segment into a pre-trained language coding model. This model can convert text of arbitrary length into a fixed-length numerical vector representation. This speech feature vector is a highly abstract and condensed representation of the semantic information of the text segment. In mathematical space, the more similar the semantics of two texts, the closer their feature vectors are. This vector does not contain specific words of the text, but rather encodes the overall meaning, topic distribution, and deep semantic structure of the text in numerical form.

[0034] In this embodiment, the text labeling unit determines the statement type of each speech segment. Inputting the speech feature vectors of all speech text sub-data into the question-answering recognition model means that the system sequentially feeds the numerical vector of each speech text into a pre-trained question-answering classifier. The question-answering recognition model is a binary classification neural network that has been trained on a large amount of dialogue data labeled with questions and statements. The model determines whether a speech segment is a question or a statement based on the input feature vectors and outputs the corresponding probability value. After processing, a text label is determined for each speech text sub-data in the speech text data. This label marks the functional role of each speech segment in the dialogue: whether it is a question asking a question or a statement responding to a question.

[0035] In this embodiment, the speech tagging unit determines the actual role attribution for each speech segment. The text tags based on all speech text sub-data of the same speech number in the speech text data refer to the system's aggregation and analysis of tags for all speech segments from the same speaker. For each speech number, i.e., each speaker, the system counts the quantity and proportion of question tags and declarative tags in that speaker's speech. Based on the statistical results, the speech tag for each speech number is determined, including two roles: interviewee and interviewer. The core logic of the determination is: in a standard interview, the role of asking questions is the interviewer, and the role of answering is the interviewee. Therefore, speech numbers with a significantly higher proportion of question tags should be marked as interviewers, and speech numbers with a significantly higher proportion of declarative tags should be marked as interviewees.

[0036] In this embodiment, the question text data unit merges the interviewer's scattered statements into a complete question. Merging all sub-data of statements tagged as "interviewer" and with adjacent statement sequence numbers involves the system filtering out all statements marked as "interviewer" and then concatenating adjacent statements into a longer text unit according to their sequence numbers. After merging, the question text data is determined, which includes multiple sub-data of question texts, each representing a complete round of questions posed by the interviewer. Each sub-data of question texts includes a question time interval record, recording the start and end times of that question in the interview audio.

[0037] In this embodiment, the answer text sub-data unit identifies each segment of the interviewee's speech as an answer. The system filters out all speech text sub-data tagged with the interviewee from the speech text data, directly identifying each such speech as the answer text sub-data. Simultaneously, the time interval of the original speech is directly identified as the answer time interval for that answer text sub-data. In a standard interview dialogue, all of the interviewee's statements are essentially responses to the interviewer's questions; therefore, cross-segment merging operations are not required as in the case of questions.

[0038] In this embodiment, the answer numbering unit enables the temporal pairing of questions and answers. Based on the question time intervals of all question text sub-data and the answer time intervals of all answer text sub-data, the system performs a timeline alignment operation on each round of questions and each answer. Specifically, the questions are sorted according to the start time of their question time intervals, and then each question is assigned an incrementing question number in chronological order. The answer number of the answer text sub-data corresponding to the question time interval closest to the question number is the same as that question number; that is, the answer belongs to the question that immediately follows the question.

[0039] In this embodiment, the interview answer data unit outputs the final structured interview data. Based on the question text sub-data of all question numbers and the corresponding question time intervals, the system integrates and generates the interviewee's interview question data, which includes all numbered questions and their time positions throughout the entire interview. Based on the answer text sub-data of all answer numbers and the corresponding answer time intervals, the system integrates and generates the interviewee's interview answer data, which includes all numbered answers and their time positions throughout the entire interview.

[0040] The beneficial effects of the above technology are: acquiring and determining the interview questions and answers based on the interviewee's audio data. This enables any form of free-flowing interview to be automatically parsed into structured interview question and answer data by the system. Example 4:

[0041] Based on Example 1, an intelligent job matching system based on a large language model includes a determination module, comprising: Word segmentation unit: Perform word segmentation on the sub-data of the question text for each question number in the interview question data, and determine the set of question words for each question number and the word position number of each word in the set of question words; The first matching unit: Match each word in the question word set for each question number in the interview question data with the knowledge term set of all knowledge points in the job knowledge database to determine the question knowledge set for each question number and the knowledge point position number set for each knowledge point in the question knowledge set; Judgment unit: Determines whether the same knowledge point exists in the question knowledge set of multiple adjacent question numbers. If it exists, determine that the multiple question numbers are a deep question set. If it does not exist, determine that each question number is a non-deep question. The second matching unit: matches each word in the question word set for each question number with the hierarchical question set for each cognitive level in the cognitive level data to determine the question cognitive set for each question number and the hierarchical position number of each cognitive level in the question cognitive set; Knowledge-cognitive data unit: Based on the question text sub-data of each question number, the question knowledge set, the knowledge point position number set of each knowledge point in the question knowledge set, the question cognition set, and the level position number of each cognition level in the question cognition set, the knowledge-cognitive data of each question number is determined. The knowledge-cognitive data includes multiple knowledge-cognitive tuples, and each knowledge-cognitive tuple includes at least one or more cognition levels. Set the number of words to be segmented: Based on the answer text sub-data of each answer number, determine the set number of words to be segmented for each answer number; Segmentation Unit: Based on the set number of words to be segmented and sentence boundaries for each answer number, the answer text sub-data for each answer number is segmented to determine multiple segmented text data for each answer number, as well as the segmentation number, segmentation time interval, and segmentation midpoint timestamp for each segmented text data. Job requirement matching data unit: The split text data of each split number of each answer number is sequentially input into the job requirement model of the job position to determine the job requirement matching data of each split number of each answer number. The job requirement matching data includes feature matching values ​​of multiple job requirement features.

[0042] In this embodiment, word segmentation is a fundamental step in text processing. The word segmentation unit performs word segmentation on the sub-data of the question text for each question number in the interview question data. Word segmentation refers to dividing a continuous sequence of Chinese characters into independent word units according to certain rules; this is the most basic operation in Chinese natural language processing. After word segmentation, the system determines the set of question words for each question number, which contains all the valid words segmented from the question. Simultaneously, the system also determines the word position number for each word in the question word set, which records the order in which the word appears in the original question text.

[0043] In this embodiment, the first matching unit is responsible for establishing the association between questions and job knowledge. The first matching unit processes each word in the question word set for each question number in the interview question data, matching each word with the knowledge terminology set of all knowledge points in the job knowledge database. Each knowledge point in the job knowledge database is pre-associated with a set of key terms to identify that knowledge point; the matching operation checks whether the words in the question are the same as or highly similar to these key terms. After matching, the question knowledge set for each question number is determined, which contains all the knowledge points involved in the question. Simultaneously, a set of knowledge point position numbers for each knowledge point in the question knowledge set is determined. This set records the position number of the matching word corresponding to that knowledge point in the original question, used to preserve the precise source of the knowledge point in the question.

[0044] In this embodiment, the judgment unit is responsible for identifying whether the interviewer is engaging in follow-up questioning. The judgment unit checks whether the knowledge sets of multiple adjacent question numbers contain the same knowledge point. This judgment process iterates through the knowledge sets of all adjacent questions and performs cross-comparisons. If multiple adjacent questions are found to involve the same knowledge point, it indicates that the interviewer is continuously and deeply questioning around that knowledge point, and the system merges these question numbers into a single set of deep questions. If adjacent questions do not share any knowledge points, it indicates that the questions are independent and there is no follow-up relationship between them, and the system identifies each question number as a separate non-deep question.

[0045] In this embodiment, the second matching unit is responsible for establishing the association between questions and cognitive levels. The second matching unit processes each word in the question word set for each question number, matching each word with the question set for each cognitive level in the cognitive level data. Each cognitive level in the cognitive level data has pre-set typical questioning methods and guiding words corresponding to that level; the matching operation identifies whether the words in the question match these typical descriptions. After matching, the question cognitive set for each question number is determined, which contains all cognitive levels triggered by that question. Simultaneously, the level position number of each cognitive level in the question cognitive set is determined; this number records the position of the matching word that triggered that cognitive level in the original question.

[0046] In this embodiment, the knowledge-cognition data unit reads the sub-data of the question text corresponding to each question number, the question knowledge set, the set of word segmentation position numbers corresponding to each knowledge point, the question cognition set, and the level position numbers corresponding to each cognition level. This unit extracts the minimum and maximum positions from the set of position numbers for each knowledge point, delineates the continuous coverage interval of that knowledge point in the question segmentation sequence, and clarifies the spatial arrangement order of each knowledge point according to the text sequence. The unit sequentially traverses all level position numbers corresponding to all cognition levels, determines the knowledge point coverage interval to which the position number belongs, and establishes an association pair between the cognition level and the corresponding knowledge point. Multiple cognition levels can belong to the same knowledge point coverage interval, and each pair of paired knowledge points and cognition levels together constitutes a knowledge-cognition binary. If the level position number of a single cognition level spans multiple knowledge point coverage intervals, the position number subset is split according to the knowledge point interval to which it belongs, and each subset is associated with the corresponding knowledge point to generate multiple knowledge-cognition binary sets. If multiple knowledge point coverage areas overlap, and the cognitive level trigger position falls into the overlapping area, the overlap ratio between the cognitive level position set and each overlapping knowledge point position set is calculated. The cognitive level is assigned to the knowledge point with the highest overlap ratio, generating a corresponding knowledge-cognition tuple. Finally, all tuples are summarized to form the complete knowledge and cognition data for the current question number.

[0047] Each unit summarizes all knowledge-cognition pairs to form the complete knowledge-cognition data corresponding to the current question number. Each knowledge-cognition pair consists of zero or more knowledge points and one or more cognitive levels.

[0048] In this embodiment, the question text sub-data is used when introducing oneself: the knowledge-cognition data of the question number includes 1 knowledge-cognition tuple, which includes zero knowledge points and memory cognition levels.

[0049] In this embodiment, the sub-data of the question text is: Please recall the core role definition of the Raft protocol, explain the basic execution process of log replication, design a data recovery scheme for single-node failure using this protocol, analyze the data inconsistency risk in network partition scenarios, evaluate its selection differences with the ZAB protocol in industrial scenarios, and attempt to propose a consistency optimization approach adapted to edge clusters. After performing word segmentation on the question text, each word is assigned a unique position number in sequence. The corresponding positions of the core matching words are as follows: 3-Recall, 4-Raft, 5-Protocol, 8-Role, 9-Definition, 10-Explanation, 11-Log, 12-Replication, 16-Process, 17-Application, 20 - Design, 22- Node, 23- Fault, 26- Recovery, 27- Solution, 28- Analysis, 29- Network, 30- Partition, 36- Risk, 37- Evaluation, 40- ZAB, 41- Protocol, 47- Difference, 50- Proposal, 53- Edge, 54- Cluster, 57- Optimization, 58- Idea; Five knowledge points are obtained through word matching. Each knowledge point corresponds to a set of matching word positions. The extreme values ​​are used to define the coverage range: Knowledge point 1: Core roles and log replication mechanism of the Raft protocol, matching positions 4, 5, 8, 9, 11, 12, 16, covering positions 4 to 16; Knowledge point 2: Raft single-node fault data recovery solution, matching position 1... 9, 22, 23, 25, 26, 27, covering the interval from position 19 to position 27; Knowledge Point 3: Data consistency risks in network partition scenarios; Matching positions 29, 30, 34, 35, 36, covering the interval from position 29 to position 36; Knowledge Point 4: Selection of Raft and ZAB protocols for industrial scenarios; Matching positions 38, 40, 41, 43, 46, 47, covering the interval from position 38 to position 47; Knowledge Point 5: Edge cluster consistency optimization scheme; Matching positions 53, 54, 56, 57, 58, covering the interval from position 53 to position 58; Six cognitive levels are obtained through hierarchical question word matching, and each cognitive level corresponds to the triggered word segmentation position number. The cognitive levels are as follows: Memory Layer (trigger positions 3 and 9, corresponding to the memory level question set); Understanding Layer (trigger positions 10 and 16, corresponding to the understanding level question set); Application Layer (trigger positions 17, 20, and 27, corresponding to the application level question set); Analysis Layer (trigger positions 28 and 36, corresponding to the analysis level question set); Evaluation Layer (trigger positions 37 and 47, corresponding to the evaluation level question set); Creation Layer (trigger positions 50, 57, and 58, corresponding to the creation level question set). Following the word segmentation sequence, the coverage areas of the five knowledge points are arranged sequentially, forming five independent knowledge discussion areas with no overlap. The trigger position of each cognitive level is determined individually: the core trigger position of that cognitive level falls within the coverage area of ​​which knowledge point, thus binding that cognitive level to the corresponding knowledge point.The memory layer trigger location falls within the coverage area of ​​knowledge point 1 and is bound to knowledge point 1. The understanding layer trigger location falls within the coverage area of ​​knowledge point 1 and is bound to knowledge point 1. The application layer trigger location falls within the coverage area of ​​knowledge point 2 and is bound to knowledge point 2. The analysis layer trigger location falls within the coverage area of ​​knowledge point 3 and is bound to knowledge point 3. The evaluation layer trigger location falls within the coverage area of ​​knowledge point 4 and is bound to knowledge point 4. The creation layer trigger location falls within the coverage area of ​​knowledge point 5 and is bound to knowledge point 5. A single knowledge point can correspond to one or more cognitive levels, and each binding relationship corresponds to a knowledge-cognition binary.

[0050] In this embodiment, a word segmentation quantity unit is set to provide granular control parameters for answer splitting. The word segmentation quantity unit is calculated based on the answer text sub-data for each answer number. The system first counts the total number of characters or words in the answer text, and then calculates the number of blocks the answer needs to be split into based on a preset target block size. The purpose of setting the word segmentation quantity is to ensure that each subsequently split text block is relatively balanced in length and contains sufficient semantic information, avoiding information compression distortion due to excessively long blocks or semantic fragmentation due to excessively short blocks.

[0051] In this embodiment, the splitting unit performs the actual segmentation of the answer text. The splitting unit segments the text based on a set number of words for each answer number and sentence boundaries. Sentence boundaries refer to the positions of punctuation marks such as periods, question marks, and exclamation marks that indicate the end of a sentence. The system first divides the answer into several natural sentences according to the sentence boundaries, and then merges adjacent sentences into appropriately sized text blocks, referencing the set number of words, without cutting off complete sentences. After splitting, multiple segmented text data for each answer number are determined, along with the splitting number, splitting time interval, and splitting midpoint timestamp for each segmented text data. The splitting number identifies the order of the segment within the answer, the splitting time interval records the start and end positions of the segment in the original audio, and the splitting midpoint timestamp is used to subsequently draw a time series trend chart.

[0052] In this embodiment, the job requirement matching data unit completes the conversion from text to quantitative assessment. The unit sequentially inputs the split text data of each answer number into the job requirement model according to the order of the split numbers. The job requirement model is a pre-trained multi-dimensional assessment model that can output the score of the input text on each job requirement dimension. After model processing, the unit determines the job requirement matching data for each split number of each answer number. This data includes feature matching values ​​for multiple job requirement features. Each feature matching value is a numerical score representing the strength of the split text's performance on the corresponding job requirement dimension. Job requirement features may include logical analysis, knowledge depth, knowledge accuracy, initiative and willingness to execute, stress resistance and time-series response stability, and teamwork.

[0053] The beneficial effects of the above technology are as follows: Based on interview questions, interview answers, cognitive level data, and job knowledge databases, multiple sets of deep questions and multiple sets of non-deep questions are identified. Furthermore, the knowledge-cognitive data for each question number in the interview question data and the job requirement matching data for each split number of each answer number in the interview answer data are determined. Unsupervised clustering of continuous questions is achieved, enabling automatic identification of deep questioning links, and achieving finer-grained temporal analysis while ensuring semantic integrity. Example 5:

[0054] Based on Example 1, an intelligent job matching system based on a large language model includes a response module comprising: Calculation unit: Based on the cognitive mapping values ​​of all cognitive levels in the cognitive hierarchy data, the knowledge-cognitive data of each question number, and the feature matching value of each job requirement feature in the job requirement matching data of each split number of each answer number, calculates the corrected matching value of the job requirement feature; Time-series response sequence unit: Based on the midpoint timestamp of the splitting midpoint of all splitting numbers of all answer numbers and the corrected matching value of each job requirement feature, determine the time-series response sequence of each job requirement feature and draw the time-series response trend chart of each job requirement feature, where the horizontal axis of the time-series response trend chart is time and the vertical axis is the corrected matching value.

[0055] In this embodiment, based on the cognitive mapping values ​​of all cognitive levels in the cognitive hierarchy data, the knowledge-cognitive data of each question number, and the feature matching value of each job requirement feature in the job requirement matching data of each split number, the corrected matching value of the job requirement feature is calculated. The calculation formula is expressed as follows: ; ; Where imN2 represents the number of knowledge terms in the m-th knowledge-cognition pair for the i-th question number, and imN3 represents the number of cognitive levels in the m-th knowledge-cognition pair for the i-th question number. This represents the nth cognitive level in the m-th knowledge-cognition pair for the i-th question number. This represents the p-th cognitive level in the cognitive hierarchy data. Let Ni represent the cognitive mapping value of the p-th cognitive level in the cognitive hierarchy data, N1 represent the number of cognitive levels in the cognitive hierarchy data, and iN4 represent the number of knowledge-cognitive pairs for the i-th question number. This represents the cognitive mapping value of the m-th knowledge-cognition pair for the i-th question number. This represents the feature matching value of the k-th job requirement feature in the job requirement matching data of the j-th split number of the i-th answer number. This represents the corrected matching value of the k-th job requirement feature in the job requirement matching data of the j-th split number of the i-th question number.

[0056] This indicates the larger of the number of knowledge terms and 1 in the knowledge-cognition binary. For example, if the question text for a certain question number is "self-introduction," only one knowledge term (0) and the cognitive level is "memory" can be obtained as a knowledge-cognition binary. The value is 1.

[0057] In this embodiment, based on the midpoint timestamp of the splitting midpoint of all splitting numbers for all answer numbers and the corrected matching value of each job requirement feature, the time-series response sequence for each job requirement feature is determined. The time-series response sequence is represented as follows: ; in, This represents the time-matching sequence of the demand feature for the k-th job position. These represent the midpoint timestamps of the splits of the 1st split of the 1st question number, the jth split of the 1st question number, and the Nu5th split of the N5th question number, respectively, where N5 represents the number of question numbers. This indicates the number of split numbers for the N5th question number. This represents the corrected matching value of the first split number of the first question number, the j-th split number of the ith question number, and the k-th job requirement feature of the nu5-th split number of the n5-th question number.

[0058] The beneficial effects of the above technologies are: determining the time-series response sequence and time-series response trend chart for the requirements of each job position; eliminating the influence of questions of varying cognitive difficulty on the dimensionality of ability scores; improving the comparability of ability scores between different questions; realizing the transformation of ability performance from static single values ​​to dynamic time-series curves; and providing a standardized data foundation for subsequent mutation detection and trend analysis. Example 6:

[0059] Based on Example 1, an intelligent job matching system based on a large language model includes a matching module comprising: Mutation timestamp unit: Perform Pettitt mutation detection on the time-series response sequence of each job requirement feature to determine the mutation timestamps of multiple negative mutation points for each job requirement feature; Mutation Feature Set Unit: Determine whether there are any questions whose midpoint timestamps corresponding to the mutation timestamps of negative mutation points of a preset number of job requirement features belong to the same depth question set or non-depth question set. If so, determine that all job requirement features are the mutation feature set of the depth question set or non-depth question set. Demand decay delay unit: Based on the mutation timestamp of the first mutation point of all job demand features in each set of deep questions or set of mutation features of non-deep questions, determine the earliest mutation timestamp and the latest mutation timestamp, and calculate the difference between the latest mutation timestamp and the earliest mutation timestamp as the demand decay delay of each set of deep questions or set of non-deep questions for candidates. Attenuation label unit: If the demand attenuation delay of the deep question set or non-deep question is less than the first preset delay threshold, the attenuation label of the deep question set or non-deep question is determined to be synchronous attenuation; if the demand attenuation delay of the deep question set or non-deep question is greater than the second preset delay threshold, the attenuation label of the deep question set or non-deep question is determined to be progressive attenuation; otherwise, the attenuation label of the deep question set or non-deep question is determined to be unattenuated.

[0060] In this embodiment, the mutation timestamp unit is the execution stage for time-series mutation detection. The mutation timestamp unit performs recursive segmentation of the time-series response sequence for each job requirement feature using Pettitt mutation detection. Pettitt mutation detection is a non-parametric statistical test specifically designed to identify the existence of a significant mutation point in time-series data and its timing. First, an initial Pettitt test is performed on the complete time-series response sequence to locate the most significant mutation point. Using this point as the dividing boundary, the sequence is divided into two independent subsequences. Then, the Pettitt test is recursively performed on each subsequence, and the segmentation is iteratively repeated until no statistically significant mutation point is detected in any of the subsequences, resulting in mutation timestamps for multiple negative mutation points. The mutation timestamp of each negative mutation point is represented by a specific moment on the interview timeline, accurately recording the actual time of the decline.

[0061] In this embodiment, the mutation feature set unit performs correlation analysis on the mutation times of multiple job requirement features. The mutation feature set unit first checks whether there are a preset specified number of negative mutation points for job requirement features. This preset specified number is a configurable system parameter; for example, the system can be set to require at least three job requirement features to mutate simultaneously before triggering subsequent analysis. For candidate cases that meet the quantity condition, the system further determines whether the split midpoint timestamps corresponding to the mutation timestamps of these negative mutation points belong to the same depth question set or the same non-depth question. The logic behind this determination is that if multiple job requirement features decline simultaneously within a questioning time period, it indicates that the question triggered a synchronous decay of multiple competency dimensions of the interviewee. If all the above conditions are met, the system determines that these simultaneously mutating job requirement features are the mutation feature set of the depth question set or the non-depth question set.

[0062] In this embodiment, the demand decay delay unit identifies the earliest and latest timestamps corresponding to the earliest negative mutation points and the latest negative mutation points in the timeline of all job demand features across all job demand features in each set of deep or non-deep questions. The difference between the latest and latest timestamps is the demand decay delay for that set of deep or non-deep questions. This delay reflects the time span from the beginning of the decay of the first job demand feature to the beginning of the decay of the last job demand feature within the question period. A shorter delay indicates a more concentrated and synchronous decay, while a longer delay indicates a more dispersed decay over time.

[0063] In this embodiment, the attenuation labeling unit is responsible for converting the calculated quantified latency into classification labels with business semantics. The attenuation labeling unit compares the attenuation latency of either a set of deep questions or non-deep questions with two preset thresholds. The first preset latency threshold is a small latency value. If the calculated attenuation latency is less than this threshold, it indicates that multiple job requirement characteristics decline almost simultaneously. The system determines the attenuation label for this question to be synchronous attenuation. Synchronous attenuation usually means that the interviewee experienced comprehensive cognitive overload during this question, leading to the simultaneous failure of various abilities. The second preset latency threshold is a large latency value. If the attenuation latency is greater than this threshold, it indicates that the decline of multiple job requirement characteristics occurs sequentially rather than simultaneously. The system determines the attenuation label for this question to be progressive attenuation. Progressive attenuation usually means that the interviewee gradually exposes their ability shortcomings as the question is pursued more deeply, and the required abilities for each job gradually disintegrate. If the attenuation latency is between the two thresholds or does not meet the abrupt change condition, the system determines the attenuation label for this question to be unattenuated, indicating that the question did not trigger a significant decline in cognitive ability.

[0064] In this embodiment, the first and second preset time delay thresholds are determined based on the normal round-robin rhythm of interview questions and the physiological and psychological laws of cognitive decay. The first preset time delay threshold can be set to 10 seconds. When multiple job requirement characteristics undergo negative abrupt changes within 10 seconds, it indicates that the interviewee has almost no room for buffering or struggling when facing the question, and all abilities fail simultaneously. This corresponds to the most severe cognitive collapse, typically occurring when the interviewee is completely unable to answer the question or their answer is severely off-topic. 10 seconds is roughly equivalent to the time it takes for a normal person to speak 2 to 3 complete sentences; a distribution of abrupt changes within this time span means the collapse occurs instantaneously, exhibiting typical characteristics of synchronous decay. The second preset time delay threshold can be set to 45 seconds. When the negative abrupt changes in job requirement characteristics are distributed within a time span of more than 45 seconds, it indicates that the interviewee has experienced a complete process from barely holding on to gradually disintegrating in response to the question. A 45-second delay roughly corresponds to the total duration of a medium-length answer. Within this timeframe, the required skills for each position progressively decrease, reflecting that the interviewee has some knowledge and ability to handle the question, but deficiencies are gradually exposed under sustained cognitive pressure. A decay delay exceeding 45 seconds indicates that the collapse process is gradual rather than sudden, consistent with a progressive decay pattern. Delays between 10 and 45 seconds are categorized as non-decaying. The abrupt changes within this range are neither concentrated nor dispersed enough, possibly due to random fluctuations or mild cognitive pressure, insufficient to determine a significant cognitive decay pattern.

[0065] The beneficial effects of the above technologies are as follows: based on the time-series response sequences and trend charts of all job requirements, the decay labels for each set of deep questions or non-deep questions can be determined. This transforms time-series capability fluctuations into quantifiable labels, enabling refined classification and quantitative characterization of the capability decay process, and providing objective fluctuation pattern judgment results without manual intervention. Example 7:

[0066] Based on Example 1, an intelligent job matching system based on a large language model includes a matching module comprising: First feature extraction unit: Performs feature extraction on the split text data of the first and last split numbers of the question number corresponding to each non-deep question, and determines the first semantic vector and the second semantic vector of each non-deep question; First semantic labeling unit: Calculate the cosine similarity value of the first semantic vector and the second semantic vector of each non-deep question. If the cosine similarity value is less than the first preset similarity threshold, determine that the semantic label of the non-deep question is semantically drifted; otherwise, determine that the semantic label of the non-deep question is semantically consistent. The second feature extraction unit extracts features from the split text data of the first split number of each question number in each deep question set to determine the first semantic vector of each question number in each deep question set. Semantic decay sequence unit: If the number of question numbers in the deep question set is greater than 2, calculate the cosine similarity value of the first semantic vector of the first question number in the deep question set and the first semantic vector of each question number other than the first question number, determine the semantic similarity value of each question number in the deep question set other than the first question number, calculate the difference of the semantic similarity of all adjacent question numbers in the deep question set, and determine the semantic decay sequence of the deep question set. Second semantic labeling unit: If the semantic decay sequence of the deep question set continues to decrease and the time-series response sequence of the job requirement feature as knowledge accuracy has a negative mutation point in the deep question set, the semantic label of the deep question set is determined to be semantic drift; otherwise, the semantic label of the deep question set is determined to be semantic consistency. Third semantic labeling unit: If the number of question numbers in the deep question set is equal to 2, calculate the cosine similarity value of the first semantic vectors of the two question numbers in the deep question set. If the cosine similarity value is less than the second preset similarity threshold and the time-series response sequence of the job requirement feature is knowledge accuracy has a negative mutation point in the deep question set, determine that the semantic label of the deep question set is semantic drift; otherwise, determine that the semantic label of the deep question set is semantic consistency. Interview Evaluation Report Unit: Based on the decay labels and semantic labels of all deep questions, the decay labels and semantic labels of all non-deep questions, and the time-series response trend chart of all job requirement characteristics, the interview evaluation report of the interviewee is determined.

[0067] In this embodiment, the first feature extraction unit is responsible for extracting semantic vectors from the answer texts of non-deep questions. Non-deep questions refer to independent interview questions that do not have a follow-up relationship with each other, and each non-deep question corresponds to a unique question number. For each non-deep question, the system locates its corresponding answer number and extracts the split text data of the first split number and the last split number from multiple split text data of the answer, representing the beginning and end of the answer, respectively. The system inputs these two text segments into the feature extraction model to determine the first semantic vector of the non-deep question, i.e., the semantic representation of the beginning of the answer, and the second semantic vector, i.e., the semantic representation of the end of the answer.

[0068] In this embodiment, the first semantic labeling unit determines semantic drift or semantic consistency for non-deep questions. The system calculates the cosine similarity value between the first and second semantic vectors of each non-deep question. This value reflects the closeness of the beginning and end of the answer in the semantic space. A higher cosine similarity value indicates that the answer maintains semantic consistency from beginning to end without significant shifts, while a lower cosine similarity value indicates that the end of the answer has deviated significantly from the beginning in meaning. The system compares this cosine similarity value with a first preset similarity threshold. If the cosine similarity value is less than the threshold, it indicates that the end of the answer has undergone significant semantic drift compared to the beginning, manifested as the interviewee gradually deviating from the core of the question or contradictions appearing in the statements. In this case, the system determines the semantic label of the non-deep question as semantic drift. If the cosine similarity value is greater than or equal to the threshold, it indicates that the answer maintains semantic consistency from beginning to end. In this case, the system determines the semantic label of the non-deep question as semantic consistency. This label reflects the coherence of the interviewee's thinking in answering a single isolated question.

[0069] In this embodiment, although the beginning and end of the same answer are sequential in time, they discuss the same topic of the same question. If the interviewee's answer is semantically consistent and coherent, the cosine similarity of the semantic vectors of the beginning and end is usually high, generally between 0.60 and 0.85. The first preset similarity threshold can be set to 0.55. When the similarity is below 0.55, it indicates that the end of the answer has undergone significant semantic drift relative to the beginning. The beginning talks about A, but the end has deviated to B or C, which corresponds to the behavioral pattern of semantic drift. The threshold of 0.55 sets a relatively lenient judgment standard. Only significant deviations between the beginning and end are marked as semantic drift, avoiding misjudging normal answer conclusions or summaries as logical problems.

[0070] In this embodiment, the second feature extraction unit is responsible for extracting semantic vectors from the beginning of the answer to each question in the deep question set. A deep question set refers to a sequence of multiple questions that continuously follow up on the same knowledge point, where the question numbers are consecutively arranged on the timeline and share the same knowledge topic. For each question number in each deep question set, the system obtains the split text data of the first split number of the answer number corresponding to that question, i.e., the beginning of each round of answers. The system inputs the text of the beginning of each answer into the feature extraction model to determine the first semantic vector for each question number in the deep question set.

[0071] In this embodiment, the semantic decay sequence unit calculates the semantic change trend of the follow-up question chain in the deep question set. The unit first determines whether the number of question numbers in the deep question set is greater than 2, i.e., whether there are at least three rounds of continuous follow-up questions. When this condition is met, the system uses the first semantic vector of the first question number in the deep question set as a reference point, and calculates the cosine similarity value between this reference vector and the first semantic vector of each question number other than the first question number. This yields the semantic similarity value of each question number in the deep question set, excluding the first question number. These values ​​reflect the semantic similarity between the beginning of each subsequent round of answers and the beginning of the first round of answers. Subsequently, the system calculates the difference between the semantic similarity values ​​of all adjacent question numbers in the deep question set, i.e., the difference between the second round and the first round, the difference between the third round and the second round, and the difference between subsequent rounds. These differences are then arranged sequentially to form the semantic decay sequence of the deep question set.

[0072] In this embodiment, the second semantic labeling unit performs a logical judgment on the deep question set based on a decay trend. When performing joint judgment, the second semantic labeling unit first checks whether the semantic decay sequence is continuously decreasing, that is, whether the semantic similarity value of each subsequent round is lower than the value of the previous round. If the semantic decay sequence is indeed continuously decreasing, it indicates that the interviewee's answer semantics are gradually moving away from the core semantics of the first round in the follow-up question chain. At this time, the unit further checks whether there is at least one negative mutation point in the temporal response sequence of knowledge accuracy within the time interval corresponding to the deep question set, that is, the system checks whether a significant decline in knowledge accuracy score is marked within the range from the start time stamp to the end time stamp of the question set. Only when both conditions are met simultaneously—a continuous decrease in the semantic decay sequence and a negative mutation in knowledge accuracy within the same time interval—does the system determine that the semantic label of the deep question set is semantic drift. If the semantic decay sequence is continuously decreasing but the knowledge accuracy has not undergone a negative mutation, the system does not recognize semantic drift but determines that the semantic label of the deep question set is semantic consistency, because semantic deviation is a legitimate topic expansion rather than a true cognitive collapse.

[0073] In this embodiment, the third semantic labeling unit handles the special case where the deep question set contains only two question numbers. When the number of question numbers in the deep question set is equal to 2, the number of adjacent differences required by the semantic decay sequence unit is insufficient to constitute an effective trend determination. Therefore, the system adopts a pairwise comparison strategy similar to that used for non-deep questions. When the third semantic labeling unit performs joint determination, it first checks whether the cosine similarity value is less than the second preset similarity threshold, that is, whether the initial semantics of the second round of answers has deviated significantly from the initial semantics of the first round of answers. If the cosine similarity value is indeed less than the threshold, it means that the interviewee has already deviated from the semantic core of the first round of answers at the beginning of the second round of follow-up questions. At this time, the unit further checks whether there is at least one negative mutation point in the temporal response sequence of knowledge accuracy within the time interval corresponding to the deep question set. That is, the system checks whether a significant decline in knowledge accuracy score is marked within the range from the start time of the first question number to the end time of the second question number in the set. Only when both conditions are met—that is, the cosine similarity value is less than the preset threshold and the knowledge accuracy undergoes a negative mutation within the same time interval—does the system determine that the semantic label of the deep question set is semantic drift. If the cosine similarity value is less than the threshold but the knowledge accuracy does not experience a negative abrupt change, the system does not recognize semantic drift but instead determines that the semantic labels of that depth question set are semantically consistent. If the cosine similarity value itself is not less than the threshold, the system also determines that the semantics are consistent.

[0074] In this embodiment, cross-round semantic comparison faces more uncertainties. The second round of questioning is itself a follow-up or deepening of the first round. Even if the interviewee's answer is completely reasonable and semantically consistent, there will be some natural differences in the semantics of the beginning of the second round answer compared to the beginning of the first round answer due to the change in the angle of the question. Therefore, the threshold for cross-round comparison should be slightly more lenient than the threshold for comparing the beginning and end of the same answer to accommodate the semantic shift brought about by such normal follow-up questioning. The second preset similarity threshold can be set to 0.50, meaning that only when the semantic similarity between the beginning of the second round answer and the beginning of the first round answer is less than 0.50 is a significant semantic break considered to have occurred. Combined with joint verification of negative mutations in knowledge accuracy, this threshold can effectively capture true semantic drift while maintaining a low false positive rate.

[0075] In this embodiment, the interview evaluation report unit is the final decision-making synthesis layer. The system summarizes the decay labels and semantic labels of all deep questions, the decay labels and semantic labels of all non-deep questions, and the time-series response trend charts of all job requirement characteristics to determine the interview evaluation report for the interviewee. This provides complete and traceable quantitative evidence for recruitment decisions.

[0076] The beneficial effects of the above technologies are: determining the semantic labels for each set of deep questions or non-deep questions, and generating interview evaluation reports for interviewees. This improves the adaptability of semantic analysis under different dialogue structures, effectively distinguishes between normal topic expansion and true semantic deviation, reduces the false positive rate of semantic consistency determination, and enhances the completeness and traceability of dialogue analysis conclusions. Example 8:

[0077] This invention provides an intelligent job matching method based on a large language model, used to execute any one of the intelligent job matching systems based on a large language model in Examples 1 to 7, with reference to... Figure 2 ,include: S1: Acquire multi-source dialogue data, construct a question-and-answer recognition model, acquire cognitive level data and job knowledge database for the required positions, acquire and determine the interview questions and answers based on the interviewee's interview audio data; S2: Based on the interview questions, interview answers, cognitive level data, and job knowledge database, determine multiple sets of in-depth questions and multiple sets of non-in-depth questions, and determine the knowledge-cognitive data for each question number in the interview question data and the job requirement matching data for each split number of each answer number in the interview answer data; S3: Based on cognitive level data, knowledge-cognitive data of all question numbers, and job requirement matching data of all split numbers of all answer numbers, determine the time-series response sequence and time-series response trend chart of each job requirement feature; S4: Based on the time-series response sequence and time-series response trend chart of all job requirements, determine the decay label and semantic label of each set of deep questions or non-deep questions, and determine the interview evaluation report of the interviewee.

[0078] The beneficial effects of the above technologies are as follows: acquiring multi-source dialogue data, constructing a question-and-answer recognition model, acquiring cognitive level data and a job knowledge database for required positions, acquiring interview audio data of interviewees, determining interview question data and interview answer data, identifying multiple sets of deep questions and multiple non-deep questions, determining the knowledge-cognitive data for each question number in the interview question data and the job requirement matching data for each split number of each answer number in the interview answer data, determining the temporal response sequence and temporal response trend graph for each job requirement feature, determining the decay label and semantic label for each set of deep questions or non-deep questions, and determining the interviewee's interview evaluation report. This improves the accuracy of question-and-answer matching in multi-role interview audio, achieves temporal continuous detection of semantic matching degree and automatic location of abrupt change points, quantifies the semantic decay trend during the dialogue process, and improves the semantic association recognition accuracy of deep questioning chains.

[0079] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An intelligent job matching system based on a large language model, characterized in that, include: Acquisition Module: Acquire multi-source dialogue data, build a question-and-answer recognition model, acquire cognitive level data and job knowledge database for required positions, acquire and determine interview questions and answers based on interviewee audio data; The module identifies multiple sets of in-depth questions and multiple sets of non-in-depth questions based on the interviewers' interview questions, interview answers, cognitive level data, and job knowledge database. It also identifies the knowledge-cognitive data for each question number in the interview question data and the job requirement matching data for each split number of each answer number in the interview answer data. Response module: Based on cognitive level data, knowledge-cognitive data of all question numbers, and job requirement matching data of all split numbers of all answer numbers, determine the time-series response sequence and time-series response trend chart of each job requirement feature; Matching module: Based on the time-series response sequence and time-series response trend chart of all job requirements, determine the decay label and semantic label of each set of deep questions or non-deep questions, and generate the interview evaluation report for the interviewee.

2. The intelligent employment matching system based on a large language model according to claim 1, characterized in that, The acquisition module includes: Multi-source dialogue data unit: Acquire multi-source dialogue data, which includes dialogue text data of multi-turn dialogues. The dialogue text data includes the dialogue feature vector and text labels of each dialogue text. The text labels include questions and statements. Question-answering recognition model unit: Constructing a question-answering recognition model based on multi-source dialogue data; Cognitive level data unit: Acquire cognitive level data, which includes the cognitive mapping value of each cognitive level and a set of level questions, where the set of level questions includes multiple question descriptions; Job Knowledge Database Unit: Obtain the job knowledge database for the required job position. The job knowledge database includes multiple knowledge points, as well as the knowledge point data and knowledge terminology set for each knowledge point.

3. The intelligent employment matching system based on a large language model according to claim 1, characterized in that, The acquisition module also includes: Interview Recognition Data Unit: Acquires the interviewee's interview audio data, inputs the interview audio data into the voiceprint recognition model, and determines the interview recognition data. The interview recognition data includes multiple text fragment data and the speech number, text sequence number, and text time interval of each text fragment data. Speech text data unit: All text segments with the same speech number and adjacent text sequence numbers in the interview recognition data are merged to determine the speech text data. The speech text data includes multiple speech text sub-data and the speech number, speech sequence number, and text time interval of each speech text sub-data. Speech Feature Vector Unit: Extracts features from each speech text sub-data in the speech text data to determine the speech feature vector of each speech text sub-data in the speech text data; Text label unit: Input the speech feature vectors of all speech text sub-data in the speech text data into the question answering recognition model to determine the text label of each speech text sub-data in the speech text data; Speech Tag Unit: Based on the text tags of all speech text sub-data with the same speech number in the speech text data, determine the speech tag for each speech number, where the speech tag includes interviewee and interviewer; Question text data unit: Merge all speech text sub-data with the speech tag "interviewer" and adjacent speech sequence numbers in the speech text data to determine the question text data. The question text data includes multiple question text sub-data and the question time interval of each question text sub-data. Answer text sub-data unit: Each speech text sub-data unit in the speech text data with the speech tag of the interviewee is determined as the answer text sub-data unit, and the text time interval of the speech text sub-data unit is determined as the answer time interval of the answer text sub-data unit; Answer Numbering Unit: Based on the question time interval of all question text sub-data and the answer time interval of all answer text sub-data, determine the question number of each question text sub-data and the answer number of each answer text sub-data. Interview Response Data Unit: Based on the question text sub-data of all question numbers and the question time range, determine the interview question data of the interviewee; based on the answer text sub-data of all answer numbers and the answer time range, determine the interview response data of the interviewee.

4. The intelligent employment matching system based on a large language model according to claim 1, characterized in that, The module to be determined includes: Word segmentation unit: Perform word segmentation on the sub-data of the question text for each question number in the interview question data, and determine the set of question words for each question number and the word position number of each word in the set of question words; The first matching unit: Match each word in the question word set for each question number in the interview question data with the knowledge term set of all knowledge points in the job knowledge database to determine the question knowledge set for each question number and the knowledge point position number set for each knowledge point in the question knowledge set; Judgment unit: Determines whether the same knowledge point exists in the question knowledge set of multiple adjacent question numbers. If it exists, determine that the multiple question numbers are a deep question set. If it does not exist, determine that each question number is a non-deep question. The second matching unit: matches each word in the question word set for each question number with the hierarchical question set for each cognitive level in the cognitive level data to determine the question cognitive set for each question number and the hierarchical position number of each cognitive level in the question cognitive set; Knowledge-cognitive data unit: Based on the question text sub-data of each question number, the question knowledge set, the knowledge point position number set of each knowledge point in the question knowledge set, the question cognition set, and the level position number of each cognition level in the question cognition set, the knowledge-cognitive data of each question number is determined. The knowledge-cognitive data includes multiple knowledge-cognitive tuples, and each knowledge-cognitive tuple includes at least one or more cognition levels. Set the number of words to be segmented: Based on the answer text sub-data of each answer number, determine the set number of words to be segmented for each answer number; Segmentation Unit: Based on the set number of words to be segmented and sentence boundaries for each answer number, the answer text sub-data for each answer number is segmented to determine multiple segmented text data for each answer number, as well as the segmentation number, segmentation time interval, and segmentation midpoint timestamp for each segmented text data. Job requirement matching data unit: The split text data of each split number of each answer number is sequentially input into the job requirement model of the job position to determine the job requirement matching data of each split number of each answer number. The job requirement matching data includes feature matching values ​​of multiple job requirement features.

5. The intelligent employment matching system based on a large language model according to claim 1, characterized in that, The response module includes: Calculation unit: Based on the cognitive mapping values ​​of all cognitive levels in the cognitive hierarchy data, the knowledge-cognitive data of each question number, and the feature matching value of each job requirement feature in the job requirement matching data of each split number of each answer number, calculates the corrected matching value of the job requirement feature; Time-series response sequence unit: Based on the midpoint timestamp of the splitting midpoint of all splitting numbers of all answer numbers and the corrected matching value of each job requirement feature, determine the time-series response sequence of each job requirement feature and draw the time-series response trend chart of each job requirement feature, where the horizontal axis of the time-series response trend chart is time and the vertical axis is the corrected matching value.

6. The intelligent employment matching system based on a large language model according to claim 1, characterized in that, The matching module includes: Mutation timestamp unit: Perform Pettitt mutation detection on the time-series response sequence of each job requirement feature to determine the mutation timestamps of multiple negative mutation points for each job requirement feature; Mutation Feature Set Unit: Determine whether there are any questions whose midpoint timestamps corresponding to the mutation timestamps of negative mutation points of a preset number of job requirement features belong to the same depth question set or non-depth question set. If so, determine that all job requirement features are the mutation feature set of the depth question set or non-depth question set. Demand decay delay unit: Based on the mutation timestamp of the first mutation point of all job demand features in each set of deep questions or set of mutation features of non-deep questions, determine the earliest mutation timestamp and the latest mutation timestamp, and calculate the difference between the latest mutation timestamp and the earliest mutation timestamp as the demand decay delay of each set of deep questions or set of non-deep questions for candidates. Attenuation label unit: If the demand attenuation delay of the deep question set or non-deep question is less than the first preset delay threshold, the attenuation label of the deep question set or non-deep question is determined to be synchronous attenuation; if the demand attenuation delay of the deep question set or non-deep question is greater than the second preset delay threshold, the attenuation label of the deep question set or non-deep question is determined to be progressive attenuation; otherwise, the attenuation label of the deep question set or non-deep question is determined to be unattenuated.

7. The intelligent employment matching system based on a large language model according to claim 6, characterized in that, The matching module includes: First feature extraction unit: Performs feature extraction on the split text data of the first and last split numbers of the question number corresponding to each non-deep question, and determines the first semantic vector and the second semantic vector of each non-deep question; First semantic labeling unit: Calculate the cosine similarity value of the first semantic vector and the second semantic vector of each non-deep question. If the cosine similarity value is less than the first preset similarity threshold, determine that the semantic label of the non-deep question is semantically drifted; otherwise, determine that the semantic label of the non-deep question is semantically consistent. The second feature extraction unit extracts features from the split text data of the first split number of each question number in each deep question set to determine the first semantic vector of each question number in each deep question set. Semantic decay sequence unit: If the number of question numbers in the deep question set is greater than 2, calculate the cosine similarity value of the first semantic vector of the first question number in the deep question set and the first semantic vector of each question number other than the first question number, determine the semantic similarity value of each question number in the deep question set other than the first question number, calculate the difference of the semantic similarity of all adjacent question numbers in the deep question set, and determine the semantic decay sequence of the deep question set. Second semantic labeling unit: If the semantic decay sequence of the deep question set continues to decrease and the time-series response sequence of the job requirement feature as knowledge accuracy has a negative mutation point in the deep question set, the semantic label of the deep question set is determined to be semantic drift; otherwise, the semantic label of the deep question set is determined to be semantic consistency. Third semantic labeling unit: If the number of question numbers in the deep question set is equal to 2, calculate the cosine similarity value of the first semantic vectors of the two question numbers in the deep question set. If the cosine similarity value is less than the second preset similarity threshold and the time-series response sequence of the job requirement feature is knowledge accuracy has a negative mutation point in the deep question set, determine that the semantic label of the deep question set is semantic drift; otherwise, determine that the semantic label of the deep question set is semantic consistency. Interview Evaluation Report Unit: Based on the decay labels and semantic labels of all deep questions, the decay labels and semantic labels of all non-deep questions, and the time-series response trend chart of all job requirement characteristics, the interview evaluation report of the interviewee is determined.

8. An intelligent job matching method based on a large language model, characterized in that, An intelligent job matching system based on a large language model, as described in any one of claims 1 to 7, comprises: S1: Acquire multi-source dialogue data, construct a question-and-answer recognition model, acquire cognitive level data and job knowledge database for the required positions, acquire and determine the interview questions and answers based on the interviewee's interview audio data; S2: Based on the interview questions, interview answers, cognitive level data, and job knowledge database, determine multiple sets of in-depth questions and multiple sets of non-in-depth questions, and determine the knowledge-cognitive data for each question number in the interview question data and the job requirement matching data for each split number of each answer number in the interview answer data; S3: Based on cognitive level data, knowledge-cognitive data of all question numbers, and job requirement matching data of all split numbers of all answer numbers, determine the time-series response sequence and time-series response trend chart of each job requirement feature; S4: Based on the time-series response sequence and time-series response trend chart of all job requirements, determine the decay label and semantic label of each set of deep questions or non-deep questions, and determine the interview evaluation report of the interviewee.

Images