Online interview system and method based on AI large model and digital human

Abstract

The invention discloses an online interview system and method based on an AI large model and a digital human, and mainly relates to the technical field of online interview. Comprising a user interaction layer, a service layer, a model layer and a data layer. Wherein the user interaction layer comprises a video interview interface module, a digital human display module and a candidate motion capture module; the service layer comprises an interview dialogue management module, a candidate scoring module, a test question management module and a digital person management module; the model layer comprises a resume analysis model, an interview dialogue generation model and a scoring model; the data layer comprises a user data storage module, an interview record storage module, a score data module and an interview question bank; according to the method, the AI large model is combined with the digital human, so that the accuracy of the information obtained by online interview is improved, the interview experience of candidates is improved, the interview efficiency in a large number of interview scenes is improved, and the manpower consumption of recruiters is effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of online interviews, in particular to an online interview system and method based on an AI large model and a digital person. BACKGROUND

[0002] In the recruitment process, after the initial input and screening of the text information of the resume, in order to further understand the information of the candidate, interview is a common screening method. Through the interview, the candidate's coping ability, the degree of clear pronunciation, the problem handling in the professional field and other contents that cannot be seen from the static resume can be investigated. Generally, due to the factors of the region and time of the company interviewer and the interviewee, online video interview is the most common interview method. The video interview is generally divided into multiple rounds. Except for the human resource related link, the technical interview in the professional field generally has 2 to 3 rounds, and the first round is generally more general and preliminary.

[0003] In addition, the company has a recruitment rate according to the different posts when recruiting. A large number of resumes are needed to support the demand of one recruitment. The resumes themselves are mixed with good and poor quality, and professional matching and professional mismatch. Screening these resumes and then finding the candidate meeting the requirements into the next round is a quite labor-consuming process. Moreover, when the number of recruitment demand is large, the base of the interview will also be large, which is very labor-consuming. In view of the problem, there are currently some automatic online video interview methods, and the form of digital person is used.

[0004] In the process of online interview, the common way is that after the candidate stops speaking, the recorded audio is recognized by ASR (Automatic Speech Recognition) as text input information input into the large model. The large model generates feedback information and outputs the information through TTS (Text-To-Speech) technology to convert the voice into speech and output.

[0005] However, in the existing online video interview through digital people, the main method is to play the interview materials corresponding to the fixed recruitment position, and then analyze the feedback of the candidate. In this process, on the one hand, the interview content itself is driven by background data content, and the existing materials are played in a rule matching manner or a fixed process manner. The content is relatively single, and the process is also relatively single, which cannot explore the differences between different candidates; on the other hand, the operation of the digital person is relatively simple, and it is still in the stage of playing existing animation, which cannot be combined with the actual played audio content, and the feedback of the interviewee cannot be effectively responded. Only video information can be recorded and the existing process can be executed, and the interviewee is a one-sided complaint. In addition, the whole process is through waiting for the candidate's speech pause, and then timing to judge the sentence, which also causes the unsmoothness of the interactive dialogue, and the interviewee will have a significant poor experience. SUMMARY

[0006] The purpose of the present application is to solve the problems existing in the prior art, provide an online interview system and method based on AI large model and digital person, improve the accuracy of information obtained by online interview through AI large model combined with digital person, improve the interview experience of candidates, improve the interview efficiency in a large number of interview scenes and effectively save the human consumption of recruitment personnel.

[0007] To achieve the above purpose, the following technical solutions are adopted: An online interview system based on AI large model and digital person, comprising a user interaction layer, a service layer, a model layer and a data layer; wherein: The user interaction layer is based on the HTML standard Web browser program, responsible for the interface display directly interacting with the user, and communicates with the service layer; the user interaction layer comprises a video interview interface module, a digital person display module and a candidate motion capture module; the video interview interface module is used for candidate identity verification, camera / microphone permission management, video stream real-time display and video recording of the interview process; the digital person display module is used for digital person image, expression, background configuration rendering and digital person voice and animation driving; the candidate motion capture module is used for capturing the position of the candidate subject in the picture, assisting the corresponding attention feedback action of the digital person, including candidate eye tracking, attention analysis, expression recognition and body language capture; The service layer coordinates the interaction of each module, processes the core business logic, calls the model layer capability, and exposes a unified interface to the outside; the service layer includes an interview dialogue management module, a candidate scoring module, a written test question management module, and a digital human management module; the interview dialogue management module is used to generate interview questions according to job information, fixed questions and in-depth follow-up questions, adopt multi-round dialogue logic control, adjust the sequence, trigger the follow-up condition, generate feedback prompts in real time, and perform exception handling; the candidate scoring module is responsible for calling the scoring model of the model layer, combining multi-modal data to generate structured scoring, and the multi-modal data includes voice, expression, and behavior; the written test question management module is responsible for calling the written test question and scoring the written test question; the digital human management module is responsible for driving the digital human in the foreground to perform corresponding performances and guide the user; The model layer encapsulates the AI large model capability and provides algorithm support; the model layer includes a resume analysis model, an interview dialogue generation model, and a scoring model; the resume analysis model is used to extract key information from the resume; the interview dialogue generation model generates basic interview questions according to job information and generates in-depth follow-up detail questions based on the candidate's experience; the scoring model is responsible for giving a score of the matching degree of the job information according to the candidate's interview question feedback information and giving a comprehensive score combined with the behavior data captured by the motion capture; The data layer stores and manages all data generated by the online interview system; the data layer includes a user data storage module, an interview record storage module, a scoring data module, and an interview question bank; the user data storage module is used to store the candidate resume, the corresponding job information, and the interview record text information; the interview record storage module stores video data of the interview process, including video interview process and written test process video records, candidate behavior data captured during the interview process, and abnormal video segments that need to be reviewed by humans; the scoring data module records all candidate interview record summaries and scoring results, and records the scoring results of all candidates who have applied for the position information under the classified position information; the interview question bank is used to save the interview questions entered by the enterprise itself as the default interview questions for some positions.

[0008] Preferably, in the user interaction layer, the video interview interface module uses the Web page mode to access the pure front-end environment based on the front-end and back-end separation architecture, records the entire interview, compresses the recorded data locally, and then transmits it to the background, and supports the picture-in-picture mode; the digital human display module uses UE5's Meta Human for the basic engine, and performs performance through the obtained lip animation and expression animation according to the performance text content through the service layer, and the data display model supports high and low models, which can be dynamically adjusted automatically on devices with different performances; the candidate motion capture module: analyzes the picture data captured by the camera, calculates, and completes it locally on the front-end page, and supports independent GPU card performance acceleration, and the face recognition uses the open source scheme of tensorflow, when the candidate appears to leave the screen, or the face deviates from the camera direction for a long time, the corresponding reminder signal is triggered, and the digital human management module of the service side calls the digital human display module to perform the corresponding animation performance to remind the candidate after receiving the signal.

[0009] Preferably, in the service layer, the interview dialogue management module is responsible for outputting process guide information in the entire process, and generating content according to the user's feedback information, so as to complete the multi-round interaction between the system and the candidate; the candidate scoring module summarizes all the interview information at the end of the interview, and then calls the scoring model in each dimension to evaluate the score; the written test question management module provides a set of independent front-end pages, and the candidate can complete the written test part by accessing the page, the written test questions are manually input into the system, and there are multiple sets of questions according to different business directions; the main calling party of the digital human management module is the interview dialogue management module, when the text feedback information is calculated, the TTS service is called to obtain the corresponding audio data and generate performance animation, and then the digital human display module of the interaction layer is called to perform performance.

[0010] Preferably, in the model layer, the resume analysis model supports pdf and docx file formats, the user's original resume is obtained from the user data storage module before the interview, the key information in the resume is extracted, and the data is unified into structured data, the key information includes personal basic information, contact information, work experience, project experience, skill keywords, the structured data is saved to the user data, and the corresponding mark field is added; the interview dialogue generation model prepares interview questions for each candidate in advance before the interview starts, combining the user's resume information and the recruitment position information, and combining the interview question bank; during the interview, feedback information is generated for each input information according to the input of the service layer, such as ending the answer to this question or initiating a follow-up question, and multiple languages are supported during the dialogue; the reference dimensions of the scoring model include professional ability, collaborative skill, enterprise culture adaptability, and comprehensive quality, and each dimension is scored by referring to the complete recorded information of the interview, and the multiple dimensions are integrated after independent scoring.

[0011] Preferably, in the data layer, the user data storage module is used to store the candidate's resume, the corresponding position information, and the interview record text information, wherein the resume and the position information come from manual input data before the interview starts, and are not changed in the subsequent interview process; the interview question bank saves the interview questions entered by the enterprise itself as default interview questions regardless of the position; when the position is matched, some questions will be selected from the interview question bank as default interview questions. The interview question bank is manually operated and updated regularly.

[0012] An online interview method based on AI large model and digital person, comprising the steps of: S1, preparation before the interview; extracting the candidate's resume from the system that needs to be interviewed online, and extracting the personal basic information therefrom, and saving all the extracted information to the data module; extracting the photo attached to the candidate's resume, extracting the corresponding facial features and saving them; extracting the position information of the candidate's interview from the position information, and generating corresponding basic interview questions according to the position; all the prepared interview contents will be saved in the interview question bank with a unique ID as a mark, waiting for use after the online interview starts, if the resume is updated before the online interview starts, the interview dialogue large model will update the interview contents and update them to the interview question bank; S2, start online video interview select a basic combination of appearance from the basic digital person combination library, select a background, and then start the digital person, and start the whole video recording after the system starts; First of all, ensure that the equipment for online interview is normal, including the input and output of video and audio; After confirming that all devices are normal, enter the candidate identity authentication step; S3, start the main process part of the interview The digital human guides the candidate to introduce himself. When the candidate introduces himself, the audio content is converted into text information in real time and saved, and is used as auxiliary input information in subsequent interview questions; During the entire process, the interview dialogue model understands the candidate's narration content according to real-time information, understands whether termination has occurred, and cooperates with the traditional waiting time to judge whether the candidate's answer is complete; According to the pre-fixed process, the digital human's mouth movements are driven by the speech content to perform real-time animation calculation; Start the formal interview session. The background service reads the pre-generated interview questions from the database and gives them to the digital human for performance; If the candidate does not answer the question itself after listening to the interview question, but asks questions, the interview dialogue model will respond accordingly according to the specific content; During the entire interview process of the candidate, real-time speech recognition is used for speech text analysis, and the obtained text is sent to the backend interview dialogue model. After the candidate finishes speaking, the model is called to analyze the entire block of answers; After the follow-up is completed, enter the next question and repeat the above process until all interview questions are completed; Complete all interview questions. All answers will be recorded in text form; S4, written test section For some positions with written tests, this process will be entered. The written test questions are randomly selected from a pre-prepared question bank. The digital human will remind the user of the preparation work and answer the questions during the test; The written test is conducted through an online test platform. The digital human will send the web link to the candidate after the test starts. After confirming that the test question has been opened successfully, the rules of the test are explained; At the beginning of the test, the digital human prompts the user to share the computer screen and start dual-screen recording. The video is saved for later review; During the entire test process, the digital human and the candidate will not have any other communication, and the input information of the interviewer at this time will be ignored; The candidate completes the written test questions and clicks the submit button on the online test paper, which means the test is complete. You can also submit the abandon button to abandon this test; After successfully completing the answers, the digital person guides the candidate to share his thinking process and records the corresponding information. The scoring model will score the candidate's thinking process based on the content of the written test. S5, Free questions The candidate asks some routine job information, and the answer data comes from the company's knowledge base for the corresponding position. The knowledge base is edited and entered by artificial basis, and then added through continuous accumulation. After the AI large model inputs the question, it will first query the knowledge base to find matching content for output. If the content of the inquiry is not in the knowledge base, use the general reply, and record these questions in the candidate's record. These questions will be extracted and supplemented as new QAs to the knowledge base during the subsequent manual review. Set the question duration. If the set time is exceeded or the digital person and the candidate confirm that there are no problems, this step is completed. S6, End of interview The digital person prompts the end of this online interview, thanks the candidate, and closes the video connection. At the same time, all video recordings are completed, and the real-time voice text content is completed. After receiving the online interview end notification, the background service calls the scoring module to extract all the interview information for scoring in various dimensions. Finally, the total score is calculated and recorded in the database. The full interview video will be used to analyze the candidate's attention concentration level in the future. The scoring system uses a structured scoring method. According to the job requirements, the evaluation dimensions are preset. The AI large model automatically scores based on the candidate's comprehensive information and generates a comprehensive evaluation report, providing the scoring reason and corresponding scoring basis. S7, Review process After the online interview, some candidates marked during the online interview will be manually reviewed based on their marked status during the interview. Candidates who lack photo information will be contacted to complete their recent photos and then perform face feature comparison on the video again after receiving the photos. Candidates who lack information in the free question section will also have their information supplemented manually. S8, Abnormal processing During the interview, if the network is unexpectedly interrupted, the system will wait until the network is restored. If the network cannot be restored, the existing information of this interview will be recorded, and then the interview will be recorded as an abnormal state. Subsequent manual review can continue the interview or terminate the candidate's interview process. If the candidate does not join the online interview at all or the candidate terminates the interview in advance, the corresponding abnormal state will be recorded, and the reason will be confirmed in the subsequent manual review step, and it will be decided whether to arrange the online interview again.

[0013] Preferably, in step S1, the resume analysis process is to convert the original pdf or docx format document to text format through a conversion tool, and then use the resume analysis large model to extract the corresponding information in the resume according to the fixed data type, and finally form structured personal data; When extracting the photo attached to the candidate's resume, if the resume does not attach a photo, the corresponding face feature data will not be collected, and this step will be skipped at the beginning of the video, and the face authentication link will be marked as missing, and the information will be completed by the subsequent other processes; The source of the interview questions has two parts, one of which is from the system's reserved interview questions for this type of position, which is obtained and saved by manual or historical interview records and updated regularly; The other part comes from the dynamic interview questions generated by calling the interview dialogue generation model according to the user's resume experience, and some of the content will be asked in depth.

[0014] Preferably, in step S6, when analyzing the candidate's attention concentration, it is judged whether the candidate is fully engaged in the interview, if the candidate's attention is not concentrated during the interview, and the gaze is frequently stagnant or wandering for a long time, the system will evaluate the authenticity of the interview and record the corresponding information; In addition, the communication ability and adaptability are evaluated through voice tone analysis.

[0015] Preferably, in step S7, for the interview records with completed scores, part of them are randomly selected, and the video records of the entire interview are reviewed in combination with the final scores to confirm whether the AI large model gives a reasonable evaluation in this process; When it is found that the score of the interview does not match the actual interview process, feedback is given to the AI large model scoring module team to analyze and adjust the case data.

[0016] Compared with the prior art, the present application has the following advantages: The present application uses AI large model to deeply mine the resume, asks questions according to the candidate's experience and the needs of the position, obtains more rich feedback and professional skill information of the candidate; In the online interview process, the candidate's online video interview answer content is analyzed, and the questions are asked in depth in real time, which can more comprehensively and deeply understand the candidate's professional skills, so as to better judge the professional level; The digital person driven by voice and mouth shape is configured with dynamic expression management, and can interact with the candidate in the interview under the driving of the AI large model, which can greatly improve the user experience of the candidate during the interview, and make the candidate better play the level. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is the overall architecture diagram of the online interview system based on AI large model and digital person of the present application.

[0018] Figure 2 is the interview flowchart of the online interview method based on AI large model and digital person of the present application.

[0019] Figure 3 is the flowchart of the interview dialogue generation model. DETAILED DESCRIPTION

[0020] The present application will be further described below in connection with specific embodiments. It should be understood that these embodiments are only used to illustrate the present application and not used to limit the scope of the present application. In addition, it should be understood that after reading the content taught by the present application, those skilled in the art can make various modifications or changes to the present application, and these equivalent forms also fall within the scope defined by the present application.

[0021] Example 1: As shown in the attached Figure 1 The present application is an online interview system based on AI large model and digital person, which comprises a user interaction layer, a service layer, a model layer and a data layer, wherein: The user interaction layer is responsible for the interface display directly interacting with the user, based on the HTML standard Web browser program, while communicating with the service layer. It mainly includes a video interview interface module, a digital person display module and a candidate motion capture module. The video interview interface module provides candidate identity verification, camera / microphone permission management, real-time video stream display (digital person + candidate picture) and video recording function during the interview process. The digital person display module is responsible for the configuration and rendering of the appearance, expression and background of the digital person, as well as the voice and animation driving of the digital person. The candidate motion capture module is responsible for capturing the position of the candidate in the picture, assisting the corresponding attention feedback action of the digital person, as well as the functions of eye tracking, attention analysis, expression recognition and body language capture of the candidate.

[0022] Service layer: coordinates the interaction between modules, handles core business logic, calls model layer capabilities, and exposes unified interfaces to the outside. Mainly includes interview dialogue management module, candidate scoring module, written test question management module, digital human management module. The interview dialogue management module is responsible for generating interview questions according to job information, fixed questions + in-depth follow-up questions, multi-round dialogue logic control, sequence adjustment, follow-up trigger conditions, and real-time generation of feedback prompts. Such as "Please further explain the specific contribution in the project", exception handling, such as network interruption recovery, candidate deviation question guidance. The candidate scoring module is responsible for calling the scoring model of the model layer, combining multi-modal data, voice, expression, behavior, etc. to generate structured scoring. The candidate scoring module is mainly responsible for summarizing information and calling the scoring module of the model layer for scoring. The written test question management module is responsible for calling the written test questions and scoring the written test questions. The digital human management module is responsible for driving the digital human in the foreground to perform corresponding performances and guide the user.

[0023] Model layer: encapsulates AI large model capabilities, provides resume analysis, dialogue generation, scoring and other core algorithm support. Including resume analysis model, interview dialogue generation model, scoring model. The resume analysis model extracts key information from the resume, such as personal basic information, contact information, work experience, project experience, skill keywords, etc. The interview dialogue generation model generates basic interview questions based on job information, and generates in-depth follow-up detail questions based on the candidate's experience. When generating questions, multiple languages are supported, such as English or Chinese. The scoring model is responsible for giving a score of the matching degree of the job information according to the candidate's interview question feedback information and giving a comprehensive score combined with the behavior data captured by the motion capture.

[0024] Data layer: stores and manages all data generated by the system, including user information, interview records, model training data, etc. Including user data storage module, interview record storage module, scoring data module and interview question library. The user data storage module stores the candidate's resume, corresponding job information, and interview record text information. The interview record storage module stores the video data of the interview process, including the video interview process and the video record of the written test process, including the candidate's behavior data captured during the interview process, such as eye tracking, expression labels, etc., and stores abnormal video segments that need to be reviewed by humans. The scoring data module records the interview record summary and scoring results of all candidates. At the same time, under the classification of job information, it records the scoring results of all candidates who have applied for this job information. The interview question library saves the interview questions entered by the enterprise itself as the default interview questions for some positions.

[0025] Embodiment 2: As shown in the accompanying Figure 2 , the present application is an online interview method using the online interview system based on AI large model and digital human in embodiment 1, comprising the steps of: S1, preparation before interview 1) Extract the candidate's resume from the system that needs to be interviewed online, and extract the basic information such as name, gender, age, education, and work-related information such as education, professional skills, work experience, and project experience. The parsing process is to convert the original pdf or docx format document to text format through conversion tools, and then use the resume parsing large model to extract the corresponding information in the resume according to the fixed data type, and finally form structured personal data. All extracted information is saved to the data module.

[0026] 2) Extract the photo attached to the candidate's resume, extract the corresponding facial features and save them. The facial features are used for face recognition matching at the beginning of the online video interview. If the resume does not attach a photo, the corresponding facial feature data will not be collected, and this step will be skipped at the beginning of the video. Mark the missing face authentication step and complete the information by other subsequent processes.

[0027] 3) Extract the position information of the candidate's interview from the position information, and generate the corresponding basic interview questions according to the position. The source of the interview questions has two parts, the basic part comes from the system's reserved interview questions for this type of position, which is obtained and saved by manual or historical interview records, and is updated regularly. The other part comes from the user's resume experience, calling the interview dialogue generation model to generate dynamic interview questions, and some of the content will be asked in depth.

[0028] 4) All prepared interview content will be saved in the interview question library with a unique ID as a marker, waiting to be used after the online interview starts. If the resume is updated before the online interview starts, the interview dialogue large model will be called to update the interview content and update the interview question library.

[0029] S2, start online video interview 1) Select a basic combination of appearance from the basic digital human matching library, the system has a default configuration of common business appearance, and select a background, which can customize the logo of the recruitment party and other information on the background, and then start the digital human.

[0030] The voice of the digital human defaults to the standard male and female voice of the system. When interviewing international positions, the language system is determined according to the language used by the candidate, supporting Chinese and English. The digital human's mouth movements are driven by the voice to perform real-time animation calculations. The facial expression of the digital human is controlled in a calm state with a slight smile, and the eyes randomly perform blinking animation playback. Half-body appearance, with professional business formal wear, to enhance the formal feeling of the interview.

[0031] The system will start the whole video recording after starting.

[0032] 2) First, make sure the online interview equipment is normal, including video, audio input and output. The system will first confirm that the user interaction layer audio and video playback signal is normal. Then it will read the candidate's camera and audio information to confirm that the candidate has turned on the camera and microphone.

[0033] If the camera is not turned on, the digital person will provide voice audio guidance according to the steps, first prompting the candidate to turn on the camera operation behavior, such as telling "please turn on the camera and keep the camera clear and visible", and waiting for the candidate to operate. Within a specified time, if the candidate still cannot turn on the camera, the digital person will tell the candidate that the online video interview is terminated due to the candidate's failure to turn on the camera.

[0034] Similarly, if the microphone is not turned on, the digital will remind to turn on the microphone, and then guide the candidate to turn on the microphone. Then wait for the candidate to operate. Within a specified time, if the candidate still cannot turn on the microphone, the digital person will tell the candidate that the online video interview is terminated due to the candidate's failure to turn on the microphone.

[0035] During this process, all feedback information of the candidate such as audio and video will be recorded, and the audio will be texted to make a text record. The result of the interview is not completed normally, waiting for subsequent manual review to review the specific reasons for the abnormal interview, and whether to provide the candidate with the opportunity to interview online again.

[0036] 3) After confirming that all equipment is normal, enter the candidate identity authentication step. The digital person prompts the candidate to look at the camera, such as telling "please look at the camera". If the candidate's face features are extracted from the photo, the system performs face feature matching to confirm whether it is the person. If there is no candidate's resume photo in the resume information extraction stage, the candidate's image record of looking straight at the camera is retained for subsequent use, and the subsequent face feature comparison step is skipped. The verification of this information is completed by the subsequent manual link.

[0037] In abnormal cases, when the face feature authentication does not match, the digital person will prompt the candidate that the matching fails, and tell the candidate like "face matching fails". The first matching failure will give the candidate a chance to retry, and the digital person will prompt the candidate to keep the camera clear and stable and look at the lens. If it fails again, the digital person will prompt the termination of this online interview. Then save all video process records and face feature authentication failure information, and wait for subsequent manual confirmation.

[0038] S3, start the main process part of the interview 1) Digital human guides the candidate to introduce himself, and the process is a fixed process, and the digital human does not take special initiative. When the candidate introduces himself, the audio content will be converted into text information in real time and saved, and will be used as auxiliary input information in the subsequent interview questions.

[0039] 2) The interview dialogue model will understand the candidate's narration content in real time throughout the process, understand whether termination has occurred, and cooperate with the traditional waiting time to determine whether the candidate's answer is complete.

[0040] 3) According to the pre-fixed process of the interview, the digital human's mouth movements are driven by the speech content to perform real-time animation calculation. The facial expression of the digital human during the entire interview process is controlled to be calm and stable, the eyes randomly perform blinking animation, and the visual angle changes with the candidate's position captured by the camera.

[0041] 4) Start the formal interview session, the background service reads the pre-generated interview questions from the database and gives them to the digital human for content performance. The main function is to play the interview questions by voice and cooperate with simple facial expressions. The interview questions will be output by the digital human in sequence. After each question is played, it will enter a waiting state to wait for the candidate's answer. The total time of the interview is controlled, so the total number of questions and the candidate's answer time will be distributed to a certain extent to ensure that there is no big gap in time usage.

[0042] 5) If the candidate does not answer the question after listening to the interview question, but asks questions, the interview dialogue model will respond according to the specific content. Common situations include: not clear, ask for a second explanation; require further confirmation of the content in the interview question; the model will give the corresponding response and play the digital human's audio. However, if the inquiry is unrelated to the interview question, the interview dialogue model will respond that it cannot be answered at the moment, and the digital human will prompt the candidate to answer the interview question first, and then proceed to the subsequent free question section to ask related questions.

[0043] 6) Throughout the candidate interview, real-time speech recognition is used to analyze the text of the speech. The resulting text is then sent to the backend interview dialogue model. By default, the model is invoked to analyze the entire response after the candidate has finished speaking. If the backend AI model analyzes the current response and identifies areas for follow-up questions, these questions must be relevant to the main content of the job description. New questions are generated and sent to the backend service, where the digital human conducts follow-up questions. This process can be multi-round, allowing for follow-up questions based on the current question and answer, exploring multiple points within the current answer or delving deeper into a single knowledge point. The AI ​​model determines the specific approach based on previous information. If the candidate answers the question fluently and the feedback is deemed complete and accurate, no follow-up questions will be asked. Otherwise, further questions will be asked to determine the candidate's understanding of the content. The default frequency of follow-up questions for a single question is once, with a few cases involving two follow-up questions. The reason for using 1 to 2 levels is that, based on actual testing, knowledge points, after repeated refinement and in-depth analysis, can easily become sub-categorized and exceed the candidate's interview position requirements, which can easily lead to misjudgment. When asking follow-up questions, if there are multiple knowledge points, multiple follow-up questions will be generated. This means that the candidate's answer to the same question can generate different follow-up questions. This diversity of questions effectively improves the interview effect.

[0044] 7) After the follow-up questions are completed, the next question will be asked, and the above process will be repeated until all the interview questions are completed. Throughout the entire process, if the candidate takes too long to answer, the backend will call the digital human to remind the candidate to speed up and pay attention to time control.

[0045] 8) Complete the answers to all interview questions. All answers will be recorded in text form.

[0046] S4. Written examination 1) For some positions requiring a written exam, this process will be initiated. Written exam questions are randomly selected from a pre-prepared question bank. The Digital Human will provide process reminders at the start and throughout the written exam, guiding users through the preparation and answering of the exam.

[0047] 2) The written test is conducted through an online written test platform. After the written test begins, the digital human will send the web link to the candidate. After confirming that the web link has been successfully opened and the written test questions are displayed, the written test rules will be explained. The written test has a fixed time limit, and after the time limit expires, the current result will be automatically used as the final answer.

[0048] 3) The key process of the written test is to prevent the current popular AI assisted programming technology. At the beginning of the test, the digital human will prompt the user to share the computer screen, and start the dual screen recording, that is, the candidate's shared computer screen and the camera content. The content of the video will be saved for later review.

[0049] 4) During the entire test process, the digital human and the candidate will not have any other communication, and most of the candidate's input information at this time will be ignored.

[0050] 5) The candidate completes the test questions and clicks the submit button of the online test paper, which means the test is completed. At the same time, you can also submit the abandon button to abandon this test. In practice, there are also a large number of people who abandon the test. Through this way, it can also quickly filter out people who have good resumes but poor practical ability.

[0051] 6) After successfully completing the test, the digital human will guide the candidate to share his test ideas and record the corresponding information. The test system will automatically judge the score of the running result. The scoring model will score the test content and analyze the candidate's ideas to get a comprehensive score. The main purpose is to ensure that the test is indeed obtained through reasonable logical steps, rather than forced answers.

[0052] S5, free questions 1) At this stage, the candidate is provided with some routine inquiries about the position information, and all data come from the company's knowledge base for the corresponding position. The knowledge base contains various common questions in the form of QA, which will be manually edited and entered at the beginning, and then accumulated in the future. After inputting the question, the AI large model will first query the knowledge base to find matching content for output.

[0053] 2) If the content of the inquiry is not in the knowledge base, the AI large model will not be used to answer the question on the Internet. A general reply such as "specifically not clear, reply to you after confirmation" will be used. And these questions will be recorded in the candidate's record, and these questions will be extracted and supplemented to the new QA and added to the knowledge base during the subsequent manual review.

[0054] 3) This step takes a short time, controlled within 5 minutes, if the time exceeds, the digital human will kindly remind the candidate that this link is over.

[0055] 4) When the digital human and the candidate confirm that there is no problem, this link is over.

[0056] S6, end of interview 1), This step will prompt the candidate to thank the online interview and close the video connection, and end all video recording, and complete the real-time voice text content. In the transcribed text data, the user's privacy-related information such as the identity card information mentioned in the communication and the names of related personnel will be hidden. A unified scan will be conducted before saving the text content.

[0057] 2), After receiving the notification of the end of the online interview, the background service will call the scoring module to extract all the interview information, mainly text-based, to calculate the scores in each dimension. Finally, the total score is summarized and recorded in the database. Some links such as the written test will get the corresponding score after this step is completed. In this link, the corresponding score will be extracted and summarized into the total score.

[0058] 3), The full interview video will be used to analyze the candidate's attention concentration in the future. By monitoring the frequency and dwell time of eye movement, it can be determined whether the candidate is fully engaged in the interview. If the candidate is not focused during the interview and frequently appears with long-term visual stagnation or wandering, it may affect the authenticity of the interview. The system will make appropriate assessments and record the corresponding information. In addition, through voice tone analysis: identifying speech speed, pauses, volume changes, etc., to assess communication skills and adaptability (such as stress resistance).

[0059] 4), The scoring system uses a structured scoring method: according to the job requirements, preset evaluation dimensions such as communication skills, logical thinking, and professional skills, automatically score through AI large models combined with the candidate's comprehensive information, and generate a comprehensive evaluation report, giving the reasons for the score and the corresponding scoring basis.

[0060] S7, Review process 1), After the online interview is over, some of the candidates marked during the online interview will be manually reviewed according to their marked state during the interview. Those who lack photo information will be contacted to complete their recent photo and then perform facial feature comparison on the video again after obtaining the photo. Those who lack information in the free question section will also be manually supplemented.

[0061] 2), For the completed interview records, a portion will be randomly selected, with a proportion of about 1 / 8 to 1 / 10, and the entire interview video record will be reviewed in combination with the final score to confirm that the AI large model indeed gives a reasonable evaluation. When it is found that the score of the interview does not match the actual interview process, feedback will be given to the AI large model scoring module team to analyze and adjust the case data.

[0062] S8, Abnormal processing During the interview process, the system will wait when the network is unexpectedly interrupted. Each time, it will wait for a maximum of 3 minutes, and there can be a maximum of 3 waits. After the network is restored, the previous process will continue. If the network cannot be restored, the existing information of this interview will be recorded, and this interview will be recorded as an abnormal state. Subsequent manual follow-up can continue to interview again or terminate the interview process of the candidate.

[0063] If the candidate does not join the online interview at all or the candidate terminates the interview in advance, the corresponding abnormal state will be recorded. And in the subsequent step of manual review, confirm the reason, and decide whether to schedule an online interview again or not.

[0064] During the entire online interview process, each module of the interview system cooperates and provides its own capabilities. The implementation of the functions of each module is described as follows. In the user interaction layer, the video interview interface module: based on the pure front-end environment of the front-end and back-end separation architecture, using the Web page access method, so that it can be more convenient to support cross-platform use, providing audio and video access capabilities. Record the entire interview, and the recorded data will be compressed locally before being transmitted to the background. Support Picture-in-Picture (PiP) mode, the main screen is a shared screen + digital person small window.

[0065] Digital person display module: responsible for the configuration and rendering of the digital person's image, expression and background, as well as the driving of the digital person's voice and animation. The basic engine uses UE5's Meta Human, and at the same time, through the service layer, the obtained lip animation and expression animation are analyzed to perform the performance. The animation data of the lip is calculated in advance according to the text, and this part is completed by the service layer. The display layer accepts the animation data and then performs the performance. Basically, the service layer determines all the performance effects and pushes the corresponding data, and the display layer performs the corresponding performance. Voice driving is provided by the TTS service of the AI vendor, and diversified voice authorization can be obtained. The performance includes facial expression simulation, such as emotional expression, nodding, smiling, etc., as well as body movements such as using skeletal animation to control body movements, waving hands, blinking eyes, etc. The data display model supports high and low models, and can automatically adjust on devices with different performance to improve the smoothness of the performance. During the performance process, attention should be paid to the connection of the animation to avoid the situation of too dense and too fast animation.

[0066] Candidate action capture module: According to the picture data captured by the camera, the calculation is carried out by the front-end page locally, completed in the Web client, and can support the performance acceleration of independent GPU video card. The recognition of human face adopts the open source scheme of tensorflow, which has the advantages of high recognition rate and fast response speed compared with traditional opencv. The position of the candidate's head will be captured during the interview process, and a signal will be sent when the position information changes greatly. The digital human management module of the service layer will adjust the orientation and line of sight of the digital human's head accordingly to achieve the psychological feeling of looking at you. When the candidate leaves the screen or the face deviates from the direction of the camera for a long time, the corresponding reminder signal will be triggered. The digital human management module of the server will call the digital human display module to perform corresponding animation performance to remind the candidate.

[0067] In the service layer, the interview dialogue management module: responsible for outputting the process guide information in the whole process, and generating content according to the user's feedback information, so as to complete the multi-round interaction between the system and the candidate. The whole interview has a basic process, such as opening, professional skill interview, written test, and self-question and answer; these links have standard guide process and display process, and in some processes, such as the process of professional skill interview, the interaction with the candidate has no fixed process, because the interview questions and the scene may vary, at this time the AI big model of NLP understanding can respond flexibly according to the actual situation. After the initial guidance, the user will answer the questions given in the first step, at which time the interview dialogue management module will call the interview dialogue generation model to analyze and then generate the corresponding feedback. What needs to be controlled is that when the candidate's answer is judged as ambiguous or incomplete, the logic is not clear or contradictory, the description mentions a specific case implemented before, the question itself is complex and needs to understand from multiple angles, etc. The follow-up logic will be triggered. The feedback will be sent to the digital human management module, and the corresponding animation information will be generated after completion, and then the digital human display module will be driven to perform performance.

[0068] Candidate scoring module: in the final stage of the interview, all interview information is summarized, and then scoring model is called in each dimension to score and evaluate. This module itself only performs data summarization and scoring model calling, and does not implement specific scoring logic. After obtaining all the scoring data, it will be saved to the scoring data module.

[0069] The written test question management module provides a set of independent front-end pages, and the candidate can complete the written test part by accessing the page. The written test questions are manually entered into the system, and there are multiple sets of questions according to different business directions. At the same time, the written test questions will have multiple types of questions, mainly multiple choice questions and programming questions. The written test questions have matching answers, and the candidate can complete and submit the answers directly after completing the written test questions. The multiple choice questions are scored by matching the correct answers, and the programming questions will call the programming language engine to run the candidate's code and match the results with the answers. It should be noted that the programming questions can call the scoring model for process evaluation when the answer is correct to confirm that the programming approach is correct.

[0070] The digital human management module is responsible for driving the digital human in the foreground to perform corresponding performances. The input is the performance text, audio, control signal and other information, and the animation of the lip shape, the expression and the action animation of the digital human are generated. The main calling party is the interview dialogue management module, which will call the TTS service to obtain the corresponding audio data and generate performance animation after calculating the text feedback information. Then, the digital human display module in the interaction layer is called to perform the performance.

[0071] In the model layer, the resume analysis model supports pdf and docx file formats, obtains the user's original resume from the user data storage module before the interview, and extracts key information such as personal basic information, contact information, work experience, project experience, and skill keywords. The structured data is saved to the user data, and the corresponding tag field is added to facilitate subsequent use.

[0072] The interview dialogue generation model is based on the current mainstream open source model, such as Deepseek and Qwen, which can be used as the basis. LangGPT is used for structured prompt word construction on top of it, which aims to strengthen the AI's self-awareness and standardize all output formats to ensure the consistency of the upstream and downstream processing in the process. This model mainly completes two tasks: 1. Before the interview starts, the user's resume information and the recruitment position information are combined to prepare interview questions for each candidate in advance. 2. During the interview, feedback information is generated for each input information, such as ending the answer to the question or initiating a follow-up question. The model supports multiple languages, such as English or Chinese. Figure 3As shown, the service layer receives audio information in real time and converts it into text through ASR, and sends it to the model layer in real time. The model layer judges whether the input is complete, and analyzes the full text and generates feedback if the input is complete. The analysis results include scenarios such as the candidate's answer being judged as ambiguous or incomplete, the logic being unclear or contradictory, specific cases mentioned in the description, and the question itself being complex and requiring multiple perspectives. This information will be fed back to the service layer, which will decide to initiate follow-up questions based on this information.

[0073] Scoring model: responsible for scoring based on the input of comprehensive information, the candidate's interview question feedback information gives the score of the matching degree of the position information and gives the comprehensive score combined with the behavior data of motion capture. The scoring model mainly refers to the following dimensions: 1. Professional ability: the accurate grasp of professional knowledge related to the position; the ability to propose effective solutions to common problems; whether the answer is combined with actual cases and the logic is rigorous; whether there is a novel perspective or breakthrough suggestion. 2. Collaboration skills: whether the expression is clear, the logic is coherent, and whether the information can be effectively transmitted; whether the reaction is flexible when facing unexpected problems or pressure; whether it reflects a sense of cooperation and whether it can handle team conflicts; whether it remains calm under pressure and whether its emotions are controllable. 3. Enterprise culture adaptability: whether the candidate's career view and work attitude are consistent with the enterprise culture; whether it can adapt to the team's work style or the enterprise's change; whether the job motivation is clear and whether it is consistent with the long-term development of the post. 4. Comprehensive quality: whether it shows the ability to quickly learn new knowledge; whether it can objectively evaluate its strengths and weaknesses; whether it has a clear career goal and whether it is consistent with the development path of the post. Each dimension is scored based on the complete record of the interview, with a score of 10, the highest score being 10 and the lowest score being 0. Multiple dimensions are scored independently and then integrated.

[0074] In the data layer, the user data storage module: stores the candidate's resume, corresponding position information, and interview record text information. The resume and position information come from manual input data before the interview begins and are not changed during the subsequent interview process.

[0075] Interview record storage module: stores video data of the interview process, including video interview process and written test process video record, including candidate behavior data captured during the interview process such as human movement trajectory, expression label, etc., and stores abnormal video segments that need to be reviewed by humans.

[0076] Scoring data module: records the summary of all candidate interview records and scoring results. At the same time, under the classification of position information, it records the scoring results of all candidates who have applied for this position information.

[0077] Interview question bank: Save the interview questions entered by the enterprise itself as the default interview questions for all positions. When the position matches, some questions will be selected as the default interview questions. The interview question bank is manually maintained and updated regularly.

Claims

1. An online interview system based on an AI large model and a digital person, characterized in that: Comprise a user interaction layer, a service layer, a model layer, and a data layer; wherein: The user interaction layer is based on an HTML standard web browser program, responsible for the interface display directly interacting with the user, and communicating with the service layer; the user interaction layer comprises a video interview interface module, a digital human display module, and a candidate motion capture module; the video interview interface module is used for candidate identity verification, camera / microphone permission management, video stream real-time display, and video recording of the interview process; the digital human display module is used for digital human image, expression, background configuration rendering, and digital human voice and animation driving; the candidate motion capture module is used for capturing the position of the candidate in the picture, assisting the corresponding attention feedback action of the digital human, including candidate eye tracking, attention analysis, expression recognition, and body language capture; The service layer coordinates the interaction of each module, processes the core business logic, calls the model layer capability, and exposes a unified interface to the outside; the service layer comprises an interview dialogue management module, a candidate scoring module, a written test question management module, and a digital human management module; the interview dialogue management module is used for generating interview questions according to position information, fixed questions and in-depth follow-up questions, adopting multi-round dialogue logic control, sequence adjustment, follow-up trigger conditions, real-time generation of feedback prompts, and abnormality processing; the candidate scoring module is responsible for calling the scoring model of the model layer, combining multi-modal data to generate structured scoring, and the multi-modal data includes voice, expression, and behavior; the written test question management module is responsible for calling the written test questions and scoring the written test questions; the digital human management module is responsible for driving the digital human in the foreground to perform corresponding performances and guide the user; The model layer encapsulates the AI large model capability and provides algorithm support; the model layer comprises a resume analysis model, an interview dialogue generation model, and a scoring model; the resume analysis model is used for extracting key information from the resume; the interview dialogue generation model generates basic interview questions according to position information and in-depth follow-up detail questions based on the candidate's experience; the scoring model is responsible for giving a scoring of the matching degree of position information according to the candidate's interview question feedback information and giving a comprehensive score combined with the behavior data captured by the action capture; The data layer stores and manages all data generated by the online interview system; the data layer comprises a user data storage module, an interview record storage module, a scoring data module, and an interview question bank; the user data storage module is used for storing candidate resumes, corresponding position information, and interview record text information; the interview record storage module stores video data of the interview process, including video interview process and written test process video records, candidate behavior data captured during the interview process, and abnormal video segments requiring manual review; the scoring data module records all candidate interview record summaries and scoring results, and records the scoring results of all candidates who have delivered the position information under the classified position information; the interview question bank is used to save the interview questions entered by the enterprise itself as the default interview questions for some positions. 2.The online interview system based on an AI large model and a digital human according to claim 1, wherein: In the user interaction layer, the video interview interface module uses a web page to access the pure front-end environment based on the front-end and back-end separation architecture. When recording the entire interview process, the recorded data is compressed locally and then transmitted to the background. The system also supports a picture-in-picture mode. The digital human display module uses Meta Human of UE5 for the basic engine. Based on the performance text content, the obtained lip animation and expression animation are analyzed to perform the performance. The data display model supports high and low modes, which can be dynamically adjusted on devices with different performance. The candidate motion capture module analyzes the camera-captured image data, calculates the local front-end page, and completes it on the web client. It also supports independent GPU performance acceleration. The face recognition uses the open source solution of TensorFlow. When the candidate leaves the screen or the face deviates from the camera direction for a long time, a corresponding reminder signal is triggered. The digital human management module in the service layer calls the digital human display module to perform the corresponding animation performance to remind the candidate. 3.The online interview system based on an AI large model and a digital human according to claim 1, wherein: In the service layer, the interview dialogue management module is responsible for outputting process guidance information throughout the process and generating content based on user feedback information to complete multiple rounds of interaction between the system and the candidate. The candidate scoring module summarizes all interview information at the end of the interview and then calls the scoring model in each dimension to evaluate the score. The written test question management module provides a separate front-end page. Candidates can complete the written test part by accessing this page. The written test questions are manually entered into the system. There are multiple sets of questions for different business directions. The main calling party of the digital human management module is the interview dialogue management module. After calculating the text feedback information, it calls the TTS service to obtain the corresponding audio data and generates performance animation, and then calls the digital human display module in the interaction layer to perform the performance. 4.The online interview system based on an AI large model and a digital human according to claim 1, wherein: In the model layer, the resume analysis model supports pdf and docx file formats. Before the interview, the user's original resume is obtained from the user data storage module, and the key information in the resume is extracted. The data is unified into structured data. The key information includes personal basic information, contact information, work experience, project experience, and skill keywords. The structured data is saved to the user data and the corresponding tag field is added. The interview dialogue generation model prepares interview questions for each candidate in advance before the interview, based on the user's resume information and the job position information, combined with the interview question bank. During the interview, feedback information is generated for each input information, such as ending the answer to the question or initiating a follow-up question. The system supports multiple languages during the dialogue. The scoring model's reference dimensions include professional ability, collaboration skills, enterprise culture adaptability, and comprehensive quality. Each dimension is scored based on the complete recorded information. The scores from multiple dimensions are integrated. 5.The online interview system based on an AI large model and a digital human according to claim 1, wherein: In the data layer, the user data storage module is used to store the candidate resume, corresponding position information, and interview record text information. The resume and position information come from manual input data before the interview starts and are not changed during the subsequent interview process. The interview question bank saves the interview questions entered by the enterprise as default interview questions regardless of the position. When the position matches, some questions will be selected from the question bank as default interview questions. The interview question bank is manually operated and updated regularly.

6. An online interview method using the online interview system based on the AI large model and the digital person according to any one of claims 1-5, characterized in that, The steps include: S1, preparation before the interview; Extract the candidate's resume from the system that needs to be interviewed online, and extract the personal basic information from it. Save all the extracted information to the data module; Extract the photo attached to the candidate's resume, extract the corresponding facial features and save them; Extract the position information of the candidate's interview from the position information, and generate the corresponding basic interview questions according to the position; All prepared interview content will be saved in the interview question bank with a unique ID as a marker, waiting to be used after the online interview starts. If the resume is updated before the online interview starts, the interview content will be updated by the interview dialogue large model and updated to the interview question bank; S2, start online video interview Select a basic combination of appearance from the basic digital human combination library, and select a background, then start the digital human. After the system starts, start the whole video recording; First, make sure that the online interview equipment is normal, including video and audio input and output; After confirming that all devices are normal, enter the candidate identity authentication step; S3, start the main process part of the interview The digital human guides the candidate to introduce himself. When the candidate introduces himself, the audio content will be converted into text information in real time and saved, and will be used as auxiliary input information in the subsequent interview questions; During the whole process, the interview dialogue model understands the candidate's narration content according to the real-time information, understands whether the termination has occurred, and cooperates with the traditional waiting time to judge whether the candidate's answer is complete; According to the pre-fixed process, the digital human's mouth movements are driven by the voice to perform real-time animation calculation; Start the formal interview session. The background service reads the pre-generated interview questions from the database and gives them to the digital human for content performance; If the candidate does not answer the question itself after listening to the interview question, but asks a question, the interview dialogue model will respond accordingly according to the specific content; During the whole interview process of the candidate, real-time speech recognition is used for speech text analysis, and the obtained text is sent to the back-end interview dialogue model. After the candidate finishes speaking, the model is called to analyze the whole block of answers; After the follow-up is completed, the next question is entered, and the above process is repeated until all the interview questions are completed; After completing all the answers to the interview questions, all the answers will be recorded in text form; S4, written test section For some positions with written tests, this process will be entered. The written test questions are randomly selected from the pre-prepared question bank. The digital human will remind the user to complete the preparation work and answer the questions during the written test. The written test is conducted online. After the test starts, the digital human will send a web link to the candidate. Once the candidate confirms that the test has been opened and the test questions are visible, the rules of the test will be explained. At the beginning of the test, the digital human will prompt the user to share the computer screen and start the dual-screen recording. The video will be saved for later review. During the entire test process, the digital human and the candidate will not have any further communication, and the candidate's input information will be ignored. After the candidate completes the test questions and clicks the submit button on the online test paper, the test is completed. The candidate can also submit the abandon button to abandon the test. After successfully completing the test, the digital human will guide the candidate to share his test ideas and record the corresponding information. The scoring model will conduct a comprehensive score based on the content of the test and the candidate's thought analysis. S5, Free Questions The candidate will ask some routine job information questions. The answer data comes from the company's knowledge base for the corresponding position. The knowledge base is edited and entered by artificial intelligence. The AI large model will first query the knowledge base and output the matching content after inputting the question. If the content of the inquiry is not in the knowledge base, a general reply will be used, and these questions will be recorded in the candidate's record. These questions will be extracted and added to the knowledge base during the subsequent manual review. The duration of the question is set. If the set time is exceeded or the digital human and the candidate confirm that there are no problems, the process is completed. S6, End of Interview The digital human will prompt the end of the online interview, thank the candidate, and close the video connection. At the same time, all video recordings are completed, and the real-time voice text content is completed. After receiving the online interview end notification, the background service calls the scoring module to extract all the interview information for dimension score calculation. Finally, the total score is calculated and recorded in the database. The full interview video will be used to analyze the candidate's attention concentration level in the future. The scoring system uses a structured scoring method. According to the job requirements, the evaluation dimensions are preset. The AI large model automatically scores based on the candidate's comprehensive information and generates a comprehensive evaluation report, providing the scoring reason and corresponding scoring basis. S7, Review Process After the online interview, some candidates marked during the online interview will be manually reviewed based on their interview process. Candidates who lack photo information will be contacted to complete their recent photo and then perform face feature comparison again after receiving the photo. Candidates who lack information in the free question section will also have their information supplemented manually. S8, Abnormal Handling During the interview, if the network is unexpectedly interrupted, the system will wait until the network is restored. If the network cannot be restored, the existing information from the interview will be recorded, and the interview will be marked as abnormal. The subsequent manual review can continue the interview or terminate the candidate's interview process. If the candidate does not join the online interview or the candidate terminates the interview in advance, the corresponding abnormal state will be recorded, and the reason will be confirmed in the subsequent manual review step, and it will be decided whether to arrange the online interview again. 7.The online interview method based on an AI large model and a digital human according to claim 6, wherein, In step S1, the resume analysis process is to convert the original pdf or docx format document to text format through conversion tools, and then use the resume analysis large model to extract the corresponding information in the resume according to the fixed data type, and finally form structured personal data; When extracting the photo attached to the candidate's resume, if the resume does not attach a photo, the corresponding face feature data will not be collected, and this step will be skipped at the beginning of the video, and the face authentication link will be marked as missing, and the information will be supplemented by other subsequent processes; The source of the interview questions has two parts, one of which is from the system's reserved interview questions for this type of position, which is obtained and saved by manual or historical interview records, and is updated regularly; The other part comes from the dynamic interview questions generated by the interview dialogue generation model according to the user's resume experience, and some of the content will be further questioned. 8.The online interview method based on an AI large model and a digital human according to claim 6, wherein, In step S6, when analyzing the candidate's attention concentration, it is judged whether the candidate is fully engaged in the interview. If the candidate's attention is not concentrated during the interview, and there are frequent long periods of eye stagnation or wandering, the system will evaluate the authenticity of the interview and record the corresponding information. In addition, the communication ability and adaptability are evaluated through voice tone analysis. 9.The online interview method based on an AI large model and a digital human according to claim 6, wherein, In step S7, for the interview records with completed scores, some are randomly selected, the video records of the entire interview are referred to, and the final scores are reviewed to confirm whether the AI large model gives a reasonable evaluation in this process; When it is found that there is no match between the score of the interview and the actual interview process, the AI large model scoring module team is fed back, and the case data is analyzed and adjusted.

Images