Interview auxiliary answer detection method, system and device based on multi-model comparison

Abstract

The application provides an interview auxiliary answer detection method, system and device based on multi-large model comparison, which comprises the following steps: collecting audio and video streams of a remote interview process, performing speech recognition on the audio and video streams to obtain multi-turn question and answer transcription texts; calling at least two large language models based on each interview question text to obtain a reference answer text set; calculating and obtaining a model similarity score based on each answer text and the reference answer text set, and obtaining a global suspiciousness score according to the model similarity score; obtaining a model consistency index based on the reference answer text set; negatively correcting the model similarity score and the global suspiciousness score based on the model consistency index to obtain a corrected global suspiciousness score; if the corrected global suspiciousness score is in a preset suspicious interval, generating a verification follow-up trap question and displaying it to an interviewer for questioning, updating the global suspiciousness score according to the candidate's verification answer, and feeding back the updated global suspiciousness score to the interviewer.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and online interview, and relates to an interview auxiliary answer detection method, system and device based on comparison of multiple large models. BACKGROUND

[0002] In recent years, with the widespread adoption of remote working mode and the online video interview becoming the new normal of the recruitment process, the interview environment has changed significantly. Under this background, a new challenge has emerged: candidates can use external artificial intelligence tools (especially large language models) to generate, organize or polish their answers to interview questions in real time and covertly during the interview process. This behavior makes the knowledge level, logical thinking ability and language expression ability presented by the candidate not reflect their true ability, which makes the interviewer unable to make accurate assessment, and thus greatly increases the risk of misrecruitment and potential human cost for the enterprise.

[0003] To address the challenge of AI-generated text, the industry has proposed various detection technologies. However, these existing methods have obvious limitations and deficiencies when faced with the specific scenario of real-time interactive interviews, which are specifically manifested as follows: the existing AI-generated text detection methods are mostly for static text and are not suitable for multi-round dialogue, real-time detection and closed-loop evaluation combined with verification questions; at the same time, when there are "standard answer questions" (such as certain technical dogma questions and textbook-style answers), simply relying on the similarity between the candidate's answer and the single model answer is prone to false positives. SUMMARY

[0004] The application provides an interview auxiliary answer detection method, system and device based on comparison of multiple large models, which is used to solve the problem that the existing detection methods are difficult to apply to real-time interactive interview scenarios and are prone to high false positive rates.

[0005] In a first aspect, the application provides an interview auxiliary answer detection method based on multi-large model comparison, which comprises: collecting audio and video streams of an interviewer and a candidate during a remote interview, and performing speech recognition on the audio and video streams to obtain multi-round question and answer transcription texts; the question and answer transcription texts comprise interview question texts and answer texts; at least two large language models are called based on each round of interview question texts to obtain a reference answer text set for each round of interview question; the reference answer text set comprises reference answer texts of each large language model; the model similarity scores of the corresponding interview questions are calculated based on each answer text and the corresponding reference answer text set, and the global suspiciousness score of the candidate is obtained according to the model similarity scores of each round of interview questions; consistency calculation is performed based on the reference answer text set of each round of interview questions to obtain the model consistency index between the reference answers of each large language model in each round of interview questions; the model similarity score and the global suspiciousness score of the current interview question are negatively corrected based on the consistency index between the reference answers of each large language model in each round of interview questions to obtain the corrected global suspiciousness score; the corrected global suspiciousness score is compared with a preset threshold interval, and if the corrected global suspiciousness score is in the preset suspicious interval, a verification follow-up trap question is generated and displayed to the interviewer; the verification answer of the candidate is obtained based on the verification follow-up trap question, the global suspiciousness score is updated according to the verification answer of the candidate, and the updated global suspiciousness score is fed back to the interviewer.

[0006] In an implementation manner of the first aspect, the model similarity scores of the corresponding interview questions are calculated based on each answer text and the corresponding reference answer text set, and the global suspiciousness score of the candidate is obtained according to the model similarity scores of each round of interview questions, which comprises: semantic similarity calculation is performed based on each answer text and the corresponding reference answer text set to obtain the similarity values of each answer text and the corresponding reference answer text; the model similarity scores of the corresponding interview questions are calculated based on the similarity values of each answer text and the corresponding reference answer text; the global suspiciousness score of the candidate is obtained by summarizing the model similarity scores of each round of interview questions.

[0007] In one implementation of the first aspect, the semantic similarity calculation based on each answer text and the corresponding reference answer text set to obtain the similarity value between each answer text and the corresponding reference answer text includes: mapping each answer text and the corresponding reference answer text set to a vector space based on a semantic vectorization model to obtain the corresponding candidate answer text vector and reference answer text vector; performing semantic similarity calculation based on the candidate answer text vector and the reference answer text vector to obtain the similarity value between each answer text and the corresponding reference answer text; the similarity value is the cosine similarity between each answer text and the corresponding reference answer text, and the calculation method of the cosine similarity includes: sij=(e(ai)*e(rij)) / (||e(ai)||*||e(rij)||) where sij represents the cosine similarity between the candidate answer text of the i-th round and the reference answer text of the j-th large language model, ai represents the candidate answer text of the i-th round, rij represents the reference answer text of the j-th large language model of the i-th round, e(ai) represents the vector of the candidate answer text ai, and e(rij) represents the vector of the reference answer text rij.

[0008] In one implementation of the first aspect, the model similarity score of the interview question is obtained by weighted averaging of the weights of the major language models, including: Si = Σ(j=1→k) wj * sij, and Σ(j=1→k) wj = 1; where wj represents the weight of the j-th major language model of the current interview question, Si represents the model similarity score of the interview question in the i-th round, sij represents the cosine similarity between the candidate's answer text in the i-th round and the reference answer text of the j-th major language model, and k represents the number of major language models.

[0009] In one implementation of the first aspect, the candidate's global suspicion score is obtained by calculating the model similarity score of each round of interview questions through a nonlinear function; P=σ(α·S̄+β·Var(S)+γ·Fbehavior), where S={S1,S2,…,Sn} represents the model similarity score sequence of each round of interview questions, Si represents the model similarity score of the i-th round of interview questions, S̄ represents the mean of Si, Var(S) represents the variance of the similarity score sequence S, Fbehavior represents the numerical features extracted based on behavioral features such as answer time and number of pauses, σ(·) represents the Sigmoid function, α, β, and γ represent trainable parameters, and P represents the global suspicion score.

[0010] In an implementation form of the first aspect, the model consistency index is calculated by a mean or weighted mean of pairwise similarity of reference answer texts of the current interview question, and is introduced as a negative correction factor in calculating the model similarity score of the current interview question or the global suspiciousness score, and the model similarity score of the current interview question and the global suspiciousness score are negatively corrected based on the consistency index between reference answers of large language models in each interview question, to obtain the corrected global suspiciousness score, including: negatively correcting the model similarity score of the current interview question based on the consistency index between reference answers of large language models in each interview question, to obtain the corrected model similarity score of each interview question; and aggregating the corrected model similarity scores of the interview questions to obtain the corrected global suspiciousness score.

[0011] In an implementation form of the first aspect, the corrected model similarity score is calculated in the following manner:

[0012] S̃i = Si-λ·MCIi, or

[0013] S̃i = Si·(1 -λ·MCIi),

[0014] wherein λ represents a preset or trained correction coefficient, Si represents the model similarity score of the i-th interview question before correction, S̃i represents the model similarity score of the i-th interview question after correction, and MCIi represents the model consistency index of the i-th interview question.

[0015] In an implementation form of the first aspect, generating and displaying the verification follow-up trap question based on detection of the AI starting verification mode includes: identifying a suspicious source type based on the corrected similarity scores of the interview questions, corresponding model consistency indexes, and behavior characteristics; selecting a target trap question template from a preset trap question template library based on the suspicious source type, and generating a candidate trap question set based on the candidate information by calling a large language model; the candidate information at least includes resume information and answer text entity information; calculating a discrimination score of each candidate trap question based on the candidate trap question set and the candidate information; selecting the question with the highest discrimination score as the verification follow-up trap question based on the discrimination scores of the candidate trap questions, and displaying it to the interviewer; and the verification follow-up trap question includes at least one candidate trap question.

[0016] In a second aspect, the application provides an interview auxiliary answer detection system based on multi-large model comparison, which comprises: a voice collection and transcription module; a voice and video stream of an interviewer and a candidate in a remote interview process is collected and voice recognition is performed on the voice and video stream to obtain a multi-turn question and answer transcription text; the question and answer transcription text comprises an interview question text and an answer text; a multi-large model calling module; at least two large language models are called based on each interview question text to obtain a reference answer text set of each interview question; the reference answer text set comprises reference answer texts of each large language model; a similarity calculation module; a model similarity score of a corresponding interview question is calculated based on each answer text and the corresponding reference answer text set, and a global suspiciousness score of the candidate is obtained according to the model similarity scores of each interview question; a model consistency analysis module; consistency calculation is performed based on the reference answer text set of each interview question to obtain a model consistency index between reference answers of each large language model in each interview question; a suspiciousness evaluation module; the model similarity score and the global suspiciousness score of the current interview question are negatively corrected based on the consistency index between reference answers of each large language model in each interview question to obtain a corrected global suspiciousness score; a verification question generation module; the corrected global suspiciousness score is compared with a preset threshold interval, if the corrected global suspiciousness score is in a preset suspicious interval, a verification follow-up trap question is generated and displayed to the interviewer; an interviewer user interface and a prompt module is configured to obtain a verification answer of the candidate based on the verification follow-up trap question, update the global suspiciousness score according to the verification answer of the candidate, and feed back the updated global suspiciousness score to the interviewer.

[0017] In a third aspect, the application provides an electronic device, which comprises: a memory storing a computer program; a processor in communication with the memory, and calling the computer program to execute the interview auxiliary answer detection method based on multi-large model comparison.

[0018] In a fourth aspect, the application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the interview auxiliary answer detection method based on multi-large model comparison.

[0019] As described above, the interview auxiliary answer detection method, system and device based on multi-large model comparison have the following beneficial effects:

[0020] The application can improve the accuracy and practicability of judgment by introducing a multi-model consensus mechanism, a model consistency index negative correction, and a verification mode, which can sensitively capture AI traces in open questions and suppress false positives in standard answer questions; the application can real-time and quantitatively prompt the candidate whether to use AI tools to assist in answering during the interview process, and reduce the false positive rate through MCI correction, improve the credibility and fairness of the remote recruitment process, and solve the problems that the existing detection methods are difficult to apply to real-time interactive interview scenarios and are prone to high false positive rates.

[0021] The detection AI in the application works as an independent module in the background to bypass listen to voice and text, and automatically generates verification follow-up questions when the detection result reaches moderate suspicion and prompts the interviewer to ask through the interviewer user interface (UI), realizing the collaborative working mode of "AI setting questions and HR asking questions". BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 A hardware structure schematic diagram of the interview auxiliary answer detection method based on multi-model comparison described in the embodiments of the application is shown.

[0023] Figure 2 A structure schematic diagram of the remote interview system described in the embodiments of the application is shown.

[0024] Figure 3 A flowchart of the interview auxiliary answer detection method based on multi-model comparison described in the embodiments of the application is shown.

[0025] Figure 4 A flowchart of the verification mode and feedback closed loop described in the embodiments of the application is shown.

[0026] Figure 5 A part of the detection result schematic diagram of a certain simulated interview described in the embodiments of the application is shown.

[0027] Figure 6 A schematic diagram of the interviewer interface detection result described in the embodiments of the application is shown.

[0028] Figure 7 A structure schematic diagram of the interview auxiliary answer detection system based on multi-model comparison described in the embodiments of the application is shown.

[0029] Figure 8 A structure schematic diagram of the electronic device described in the embodiments of the application is shown.

[0030] ELEMENT NUMBER EXPLANATION

[0031] 100 Terminal device 800 Electronic device 101 Processing unit 801 Memory 102 Storage unit 802 Processor 103 Input device 803 Display 104 Output device S301~S307 Step 700 Interview auxiliary answer detection system based on multi-model comparison S401~S406 Step 701 Voice collection and transcription module 702 Multi-model calling module 703 Similarity calculation module 704 Model consistency analysis module 705 Suspicious degree evaluation module 706 Verification question generation module 707 Interviewer user interface and prompt module DETAILED DESCRIPTION

[0032] Following, the advantages and effects of the present application can be easily understood by those skilled in the art from the description. The present application can also be implemented or applied by different specific embodiments, and the details in the description can be modified or changed based on different views and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0033] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concepts of the present application in a schematic manner, and only the components related to the present application are shown in the diagrams, not the number, shape and size of the components when actually implemented. The actual implementation of each component may be a random change in type, number and proportion, and the layout type of the components may be more complex.

[0034] The following embodiments of the present application provide an interview auxiliary answer detection method, system and device based on multi-large model comparison, which solves the problems that the existing detection method is difficult to apply to real-time interactive interview scene and is prone to high false positive rate.

[0035] The following embodiments of the present application provide an interview auxiliary answer detection method, system and device based on multi-large model comparison, which is used to real-time and quantitative evaluate whether the candidate uses external AI tools without significantly interrupting the interview process. By introducing multi-large model deliberation mechanism, model consistency index (MCI) negative correction and verification mode, it can not only sensitively capture AI traces in open questions, but also suppress false positives in standard answer questions, thereby improving the accuracy and practicality of judgment. In specific implementation, the interview question and answer are still directly conducted by the interviewer and the candidate, and the detection AI is bypassed to listen to the voice and text in the background as an independent module. When the detection result reaches moderate suspiciousness, a verification follow-up question is automatically generated and the interviewer is prompted to ask through the UI, realizing the collaborative working mode of "AI setting questions and HR asking questions".

[0036] The interview auxiliary answer detection method based on multi-large model comparison provided by the embodiments of the present application can run in a terminal device. Taking a terminal device as an example, Figure 1 as an example, Figure 1 the hardware structure block diagram of the terminal device running the interview auxiliary answer detection method based on multi-large model comparison. The terminal device 100 includes but is not limited to a processing unit 101 and a storage unit 102. The processing unit 101 and the storage unit 102 are connected through a bus.

[0037] The storage unit 102 is a non-transitory computer readable storage medium provided by the present application. The storage unit stores instructions executable by at least one processor, so that the at least one processing unit 101 executes the interview auxiliary answer detection method based on multi-model comparison provided by the present application. The non-transitory computer readable storage medium of the present application stores computer instructions for causing a computer to execute the interview auxiliary answer detection method based on multi-model comparison provided by the present application.

[0038] The storage unit 102 can include a program storage area and a data storage area, wherein the program storage area can store an operating system and application programs required by at least one function; the data storage area can store audio and video stream data required by the interview auxiliary answer detection method based on multi-model comparison and data created by use of the electronic device determined according to the remote interview auxiliary answer detection, etc. In addition, the storage unit 102 can include a high-speed random access memory, and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the storage unit 102 can optionally include a memory remotely arranged with respect to the processing unit 101, and these remote memories can be connected to the terminal device 100 determined based on the interview auxiliary answer detection method based on multi-model comparison through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0039] The terminal device 100 further includes an input device 103 and an output device 104, the input device 103 can receive input audio and video stream data, such as audio and video stream of the candidate and the interviewer, which can be stored in the storage unit 102, so that the processing unit 101 performs remote interview auxiliary answer detection calculation according to the audio and video stream, and outputs the calculated detection result through the output device 104.

[0040] The input device 103 can include but is not limited to a fisheye camera, a camera, etc. The output device 104 can include but is not limited to a liquid crystal display (LCD), a light emitting diode (LED) display, a plasma display, and a touch screen. The embodiments of the present application are not limited.

[0041] In the embodiments of the present application, the above-mentioned components of the terminal device 100 and other components not shown in the Figure 1 may be connected to each other, for example, through a bus. It should be understood that Figure 1 The computing device structure block diagram shown is only for the purpose of example, and is not a limitation on the scope of the present application. Those skilled in the art can add or replace other components as needed.

[0042] The terminal device 100 can be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smartphone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other type of mobile device, or a stationary computing device such as a desktop computer or PC. The terminal device 100 can also be a mobile or stationary server. Embodiments of the present application are not limited.

[0043] Figure 2 A structure diagram of a remote interview system described in embodiments of the present application is shown as Figure 2 The remote interview system includes a candidate terminal, an interviewer terminal, and a data processing server deployed in the cloud or locally. The server side includes, in sequence, a voice collection and transcription module, a multi-large model calling module, a similarity calculation module, a model consistency analysis module, a suspicious degree evaluation module, a verification question generation module, and an interviewer prompting module, etc. In the actual interview process, the candidate and the interviewer establish an audio and video connection through WebRTC or the like, and the candidate's voice is sent into the speech recognition module in real time to transcribe it into text, and the interviewer's questions are transcribed into qi, and the candidate's answers are transcribed into ai. The detection AI does not directly participate in the question and answer, but as a bypass listener receives the transcribed text and behavior characteristics for subsequent suspicious degree evaluation and verification mode triggering.

[0044] In some embodiments, the model set configured by the multi-large model calling module includes but is not limited to GPT series models, Claude series models, Gemini series models, and DeepSeek, Grok, Kimi, etc. third-party large language model services. The system can select any two or more models from the above model set according to business needs and actual availability to participate in deliberation; in some implementations, self-developed large language models or open source models deployed internally by the enterprise (such as variants based on LLaMA, Qwen, etc.) can also be added to the set to improve privacy and controllability.

[0045] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application.

[0046] As shown in Figure 3 The present embodiment provides an interview auxiliary answer detection method based on multi-large model comparison, which includes the following steps S301 to S307.

[0047] Step S301, collect the audio and video streams of the interviewer and the candidate in the remote interview process, and perform speech recognition on the audio and video streams to obtain multi-round question and answer transcription texts; the question and answer transcription texts include interview question texts and answer texts.

[0048] Specifically, the multi-round question and answer transcription text is an interview question text and a candidate answer text arranged in chronological order.

[0049] In step S302, at least two large language models are called based on each round of interview question text to obtain a reference answer text set of each round of interview question; the reference answer text set includes reference answer texts of each large language model.

[0050] In some embodiments, the present application inputs at least two or more large language models of different providers for each round of interview question text qi to obtain a corresponding reference answer text set {ri1, ri2, …, rik}.

[0051] In some embodiments, the large language model includes, but is not limited to, a GPT model, a Claude model, and other large language models.

[0052] In step S303, a model similarity score of the corresponding interview question is calculated based on each answer text and the corresponding reference answer text set, and a global suspiciousness score of the candidate is obtained according to the model similarity scores of each round of interview question.

[0053] In an embodiment of the present application, calculating a model similarity score of the corresponding interview question based on each answer text and the corresponding reference answer text set, and obtaining a global suspiciousness score of the candidate according to the model similarity scores of each round of interview question includes steps S3031 to S3033.

[0054] In step S3031, semantic similarity calculation is performed based on each answer text and the corresponding reference answer text set to obtain a similarity value of each answer text and the corresponding reference answer text.

[0055] In step S3032, a model similarity score of the corresponding interview question is calculated based on the similarity value of each answer text and the corresponding reference answer text.

[0056] In step S3033, the global suspiciousness score of the candidate is obtained by summarizing the model similarity scores of each round of interview question.

[0057] In an embodiment of the present application, the similarity value of each answer text and the corresponding reference answer text is obtained by performing semantic similarity calculation based on each answer text and the corresponding reference answer text set, which includes steps S30311 to S30312.

[0058] In step S30311, each answer text and the corresponding reference answer text set are mapped to a vector space based on a semantic vectorization model to obtain a corresponding candidate answer text vector and a reference answer text vector.

[0059] Step S30312, based on the candidate answer text vector and the reference answer text vector, semantic similarity calculation is performed to obtain the similarity value of each answer text and the corresponding reference answer text.

[0060] In some embodiments, based on the semantic vectorization model, the candidate answer text ai and each reference answer text rij are mapped to a vector space, the semantic similarity of the two is calculated to obtain the similarity value sij; the similarity matrix is constructed according to the similarity value of each round of questions and large language model, the AI similarity score Si is calculated for each round of questions, and the global suspiciousness score P of the candidate is obtained by summarizing all the questions.

[0061] Step S304, based on the reference answer text set of each round of interview questions, consistency calculation is performed to obtain the model consistency index between the reference answers of each large language model in each round of interview questions.

[0062] Step S305, based on the consistency index between the reference answers of each large language model in each round of interview questions, the model similarity score and the global suspiciousness score of the current interview question are negatively corrected to obtain the corrected global suspiciousness score.

[0063] In some embodiments, the application calculates the consistency index MCIi between the reference answers of each large language model in each round of questions, and uses MCIi as a negative correction factor to modify Si or P to reduce false positives in standard answer type questions.

[0064] Specifically, negative correction is a protection mechanism for reducing false positives in standard answer type questions. Negative correction means that the candidate answered a standard answer to a question, and the standard answer is consistent with the reference answer of the AI large language model, but it does not mean that the candidate cheated with AI.

[0065] The main purpose of negative correction is to "reduce noise" and avoid misjudging candidates who normally recite standard answers as AI answers. Some questions themselves have only one or two typical answer forms, such as "What is the average time complexity of quicksort?", and the answers of both excellent candidates and GPT / Claude are very similar. For such questions, if only "the similarity between the candidate's answer and the model's answer is high" is considered, there will be more false positives, so the suspicious score of such questions is negatively corrected by MCI.

[0066] Secondly, negative correction can make the "sensitivity" of the system focus on more distinguishable open questions and experience questions, such as "What specific performance optimization have you done". In these questions, the answers of AI are often templated, while real candidates will have personal project details. By reducing the weight of high MCI standard questions, more weight can be left for such questions.

[0067] In some embodiments, for example, the interview contains more basic knowledge questions or "eight essays" (such as TCP three-way handshake, B+ tree features, etc.), the answers of excellent candidates are often highly similar to the answers of large models. If no negative correction is made at this time, each question will significantly increase the suspiciousness P' at this time, and the situation of "almost all candidates are marked as moderate or high suspiciousness on the interface" is easy to appear, which leads to poor differentiation between "real candidates with good memory and sufficient preparation" and "candidates relying on AI throughout the process", and the precision (accuracy) is significantly reduced, and the trust of the interviewer on the detection result is reduced. The existing technology mainly checks the content of the candidate's answer through checking articles and checking the question bank for duplication in the remote interview process, and there is no corresponding detection mechanism. Compared with the existing technology, the negative correction of the present application can reduce the false positive rate in standard answer type questions.

[0068] In some embodiments, the present application performs negative correction for standard answer questions, fixed questions, etc. For example, standard knowledge point questions, algorithm questions, definition questions, etc. The answers to such questions are highly consistent among multiple models, and the MCI value is close to 1. Through negative correction, the contribution of the MCI to the overall suspiciousness can be reduced. Questions related to official processes or company general specifications, etc. In such scenarios, company documents or public Wiki are fixed, and such questions are more suitable as knowledge point check questions rather than the main basis for AI cheating judgment. Question bank items explicitly marked as "basic knowledge check questions" by business parties, etc. In such scenarios, the negative correction based on MCI is enabled by default.

[0069] In some embodiments, there are some scenarios where negative correction can be weakened or not used, for example, open behavioral interview questions and personal experience questions, etc. The answers to such questions differ greatly among multiple models, and the MCI itself is low, so the impact on the score is limited. In such scenarios, it may not be necessary to introduce a large amount of negative correction. Scenarios that emphasize writing style or long writing, if the business goal is to capture AI participation as much as possible, rather than to protect standard answer questions, then MCI negative correction can only be used in knowledge point questions, and the mechanism can be weakened or closed in high-level writing questions. In the initial version of the resource-limited, if only a single model is accessed and the MCI of multiple models cannot be calculated, the negative correction can be temporarily not done, and the MCI module can be used as an optional function, and the function can be enabled when the system is expanded to multiple models.

[0070] Step S306, comparing the corrected global suspiciousness score with a preset threshold interval, if the corrected global suspiciousness score is in the preset suspicious interval, generating a verification follow-up trap question and displaying it to the interviewer.

[0071] In some embodiments, the preset threshold interval includes a normal interval, a moderately suspicious interval, and a highly suspicious interval, and the preset suspicious interval is the moderately suspicious interval.

[0072] In step S307, the verification answer of the candidate is obtained based on the verification trap question, the global suspicious score is updated according to the verification answer of the candidate, and the updated global suspicious score is fed back to the interviewer.

[0073] In some embodiments, the present application divides the modified global suspicious score P' into a normal interval, a moderately suspicious interval, and a highly suspicious interval by comparing the modified global suspicious score P' with the preset threshold interval, and real-time prompts the suspicious level to the interviewer. When P' falls into the moderately suspicious interval, the verification mode is automatically started by the detection AI, at least one verification trap question is generated based on the candidate's resume and the previous answer, and is displayed to the interviewer through the interviewer interface, and the interviewer asks the candidate in the conversation. The verification answer of the candidate is repeated in steps S302-S305, the global suspicious score P'' is updated, and the updated suspicious level is fed back to the interviewer.

[0074] The present application collects the audio and video of the interviewer and the candidate during the remote interview process, and obtains the multi-round question and answer transcription text through the speech-to-text module. For each round of interview questions, the question is sent to multiple large language models to obtain reference answer texts. The semantic similarity between the candidate's answer and each reference answer is calculated, a similarity matrix of multiple models and multiple rounds of questions is constructed, and a suspicious score is calculated. Further, the consistency index MCI between multiple large model reference answers is calculated, and a negative correction is introduced into the overall score to suppress false positives of "standard answer questions". During the normal question and answer process, the detection AI continuously monitors the interview conversation in a bypass mode and calculates the global suspicious degree. When the suspicious degree enters the moderately suspicious interval, the verification mode is automatically started, the detection AI generates verification trap questions and pushes them to the interviewer, the interviewer asks the candidate in the conversation, and the suspicious score is updated again based on the verification answer. A feedback loop is formed. The present application can real-time and quantitatively prompt whether the candidate is suspected of using AI tools to assist in answering during the interview process, and reduce the false positive rate through MCI correction to improve the credibility and fairness of the remote recruitment process.

[0075] In an embodiment of the present application, the similarity value between each answer text and the corresponding reference answer text is the cosine similarity between each answer text and the corresponding reference answer text, and the calculation method of the cosine similarity includes:

[0076] sij=(e(ai)*e(rij)) / (||e(ai)||*||e(rij)||)

[0077] Where sij represents the cosine similarity between the candidate's answer text in the i-th round and the reference answer text of the j-th large language model, ai represents the candidate's answer text in the i-th round, rij represents the reference answer text of the j-th large language model in the i-th round, e(ai) represents the vector of the candidate's answer text ai, and e(rij) represents the vector of the reference answer text rij.

[0078] In one embodiment of this application, the model similarity score of the interview question is obtained by weighted averaging of the weights of various language models, including:

[0079] Si =Σ(j=1→k) wj * sij, and Σ(j=1→k) wj=1;

[0080] Where wj represents the weight of the j-th large language model for the current interview question, Si represents the model similarity score of the interview question in the i-th round, sij represents the cosine similarity between the candidate's answer text in the i-th round and the reference answer text of the j-th large language model, and k represents the number of large language models.

[0081] In one embodiment of this application, the candidate's global suspiciousness score is obtained by calculating the model similarity score of each round of interview questions using a nonlinear function, including:

[0082] P =σ(α·S̄ +β·Var(S) +γ·F behavior)

[0083] Where S={S1,S2,…,Sn} represents the model similarity score sequence of each round of interview questions, Si represents the model similarity score of the i-th round of interview questions, S̄ represents the mean of Si, Var(S) represents the variance of the similarity score sequence S, F represents the numerical features extracted based on behavioral features such as answer time and number of pauses, σ(·) represents the Sigmoid function, α, β, and γ represent trainable parameters, and P represents the global suspiciousness score.

[0084] In one embodiment of this application, the model consistency index is calculated by the pairwise similarity mean or weighted mean of each reference answer text of the current interview question, and is introduced as a negative correction factor when calculating the model similarity score or global suspicion score of a single round of interview questions.

[0085] In one embodiment of this application, the model similarity score and global suspicion score of the current interview question are negatively corrected based on the consistency index among the reference answers of various language models in each round of interview questions, and the corrected global suspicion score is obtained by the following steps S3051 to S3052.

[0086] Step S3051, based on the consistency index between the reference answers of each large language model in each round of interview question, the model similarity score of the current interview question is negatively corrected to obtain the corrected model similarity score of each interview question.

[0087] Step S3052, based on the corrected model similarity score of each interview question, the corrected global suspiciousness score is obtained by summarizing.

[0088] In an embodiment of the present application, the calculation method of the corrected model similarity score includes:

[0089] S̃i = Si-λ·MCIi, or

[0090] S̃i = Si·(1 -λ·MCIi)

[0091] Wherein, λ is a preset or trained correction coefficient, Si represents the model similarity score of the i-th round of interview question before correction, S̃i represents the model similarity score of the i-th round of interview question after correction, and MCIi represents the model consistency index of the i-th round of interview question.

[0092] In some embodiments, the present application provides the calculation and actual example of the model consistency index (MCI). In this embodiment, in order to suppress the false positives of "standard answer questions", the present application introduces the model consistency index MCI. MCI is used to characterize whether the reference answers given by multiple large language models on the same question are highly consistent. Intuitively, if different models give almost the same textbook answers, then the question is a standard answer question. At this time, even if the candidate's answer is highly similar to the model answer, it should not be directly judged as AI cheating, but the contribution weight in the global suspiciousness should be appropriately reduced.

[0093] For the i-th question, suppose there are k large language models, and the reference answer texts {ri1, ri2, …, rik} are obtained, which are respectively vectorized as e(rij). The model consistency index MCIi of the question can be calculated as follows:

[0094] MCIi = 2 / (k·(k-1))·Σ(j1<j2) cos_sim( e(ri j1), e(ri j2) ).

[0095] When MCIi is close to 1, it means that different models are highly consistent, and the question belongs to a standard answer question; when MCIi is low, it means that the answers given by each model are quite different, which is closer to an open question.

[0096] In an embodiment of the present application, the MCI negative correction is described. Suppose there are two questions Q1 and Q2 in a certain interview, which are:

[0097] Q1: "Please write the average time complexity of the quicksort algorithm" (typical standard answer question).

[0098] Q2: "Please describe how you handled an online outage in one of your past projects" (open behavioral interview question).

[0099] The system simultaneously calls GPT, Claude, and DeepSeek three large models as references:

[0100] For Q1, the reference answers of the three models are concentrated on "O(n log n)" and its standard explanation, and the cosine similarity between each pair of models is about 0.96, 0.95, and 0.97, so MCI1≈0.96;

[0101] For Q2, the three models give different cases and descriptions due to different speculated scenarios, and the similarity between each pair is about 0.55, 0.48, and 0.52, so MCI2≈0.52. Further assuming that the similarity between the candidate's answer and the three models after weighting is: S1 = 0.90, S2 = 0.82. If linear negative correction is used: S̃i = Si -λ·MCIi, let λ= 0.4, then:

[0102] S̃1 = 0.90 - 0.4·0.96 ≈ 0.52 (standard answer question, actual effective suspicion is reduced).

[0103] S̃2 = 0.82 - 0.4·0.52 ≈ 0.61 (open question, still maintains a high suspicion score).

[0104] In summary, by MCI negative correction, the present application can significantly suppress false positives of standard answer questions without losing sensitivity of open questions.

[0105] In some embodiments, the sample code for similarity calculation and MCI correction provided by the present application is as follows:

[0106] import numpy as np # import the numerical calculation library numpy for vector operation and norm calculation

[0107] def cosine_similarity(v1, v2):

[0108] v1 = np.array(v1) # convert input vector v1 to numpy array

[0109] v2 = np.array(v2) # convert input vector v2 to numpy array

[0110] # Compute cosine similarity: (v1 · v2) / (||v1|| * ||v2||), plus 1e-8 to prevent division by zero

[0111] return float(v1 @ v2 / (np.linalg.norm(v1) * np.linalg.norm(v2) +1e-8))

[0112] def compute_scores_with_mci(question, answer, llm_clients, embed_fn,weights, lambda_mci=0.4):

[0113] # 1. Generate multi-model reference answers

[0114] # For each large language model client, call its generate_answer(question) to get the model's answer to the question

[0115] refs = [client.generate_answer(question) for client in llm_clients]

[0116] # 2. Text vectorization encoding

[0117] # Encode the candidate answer answer into a vector e_ans

[0118] e_ans = embed_fn(answer)

[0119] # Encode each reference answer text into a list of vectors e_refs

[0120] e_refs = [embed_fn(r) for r in refs]

[0121] # 3. Compute similarity between candidate answer and each model reference answer

[0122] # For each reference answer vector e_r, compute the cosine similarity between e_ans and e_r, resulting in a list of sims

[0123] sims = [cosine_similarity(e_ans, e_r) for e_r in e_refs]

[0124] # Weighted sum of similarity scores of each model by weights

[0125] S = sum(w * s for w, s in zip(weights, sims))

[0126] # 4. Calculate model consistency index MCI based on pairwise similarity scores

[0127] pair_sims = [] # To store pairwise similarity scores of all model answer vectors

[0128] for i in range(len(e_refs)): # Iterate through the first model index i

[0129] for j in range(i + 1, len(e_refs)): # Iterate through the second model index j (only take i after, to avoid duplication)

[0130] # Calculate the cosine similarity between the ith and jth reference answer vectors and add it to pair_sims

[0131] pair_sims.append(cosine_similarity(e_refs[i], e_refs[j]))

[0132] # Take the average of all pair_sims as MCI; if the list is empty (insufficient models), MCI takes 0.0

[0133] MCI = float(np.mean(pair_sims)) if pair_sims else 0.0

[0134] # 5. Negatively correct the original similarity score using MCI

[0135] # S_tilde = S - lambda_mci * MCI, the higher MCI (the more similar the question is to the standard answer question), the lower the final score

[0136] S_tilde = S - lambda_mci * MCI

[0137] # Return a dictionary containing the original score, MCI, corrected score, and detailed model similarity

[0138] return {

[0139] "S_raw": S, # Original weighted similarity score without MCI correction

[0140] "MCI": MCI, # Model consistency index (average of pairwise similarity between multi-model answers)

[0141] "S_corrected": S_tilde, # Final score after negative correction

[0142] "sims_per_model": sims, # Similarity list corresponding to each model

[0143] }

[0144] Figure 4 A flowchart of the verification mode and feedback loop described in the embodiments of the present application is shown. As shown in the figure, the flow of the verification mode and feedback loop specifically includes the following steps: Figure 4

[0145] Step S401, judge whether the corrected global suspiciousness score P' is in the moderate suspiciousness interval.

[0146] Step S402, automatically start the verification mode by the detection AI, and generate at least one verification "trap" question based on the candidate's resume and previous answers.

[0147] Step S403, display at least one verification "trap" question to the interviewer through the interviewer interface, and ask the candidate in the conversation.

[0148] Step S404, obtain the reference answer text set of large language models and the model consistency index MCI_ver between the reference answers of each large language model according to the verification answers of the candidate.

[0149] Step S405, calculate the corrected similarity score of the current interview question according to the reference answer text set and the model consistency index between the reference answers of each large language model in each round of interview question.

[0150] Step S406, update the global suspiciousness score according to the corrected similarity score, obtain the corrected global suspiciousness score P'', and feed back the updated suspiciousness level to the interviewer.

[0151] In some embodiments, the present application provides a simulated data schematic example of multi-round questions. Figure 5 A part of the detection result schematic diagram of a simulated interview described in the embodiments of the present application is shown. As shown in the figure, Figure 5 ​​

[0152] Round Q1 (project introduction question): S1 = 0.62, MCI1 = 0.60, S̃1 ≈ 0.38;

[0153] Round Q2 (technical details open question): S2 = 0.88, MCI2 = 0.45, S̃2 ≈ 0.70;

[0154] Round Q3 (standard algorithm question): S3 = 0.91, MCI3 = 0.94, S̃3 ≈ 0.53;

[0155] Verification question Q_ver (on-site troubleshooting combined with resume): S_ver = 0.93, MCI_ver = 0.40, S̃_ver ≈ 0.77.

[0156] When calculating the global suspiciousness P, the system uses the corrected score sequence {S̃1, S̃2, S̃3, S̃_ver}, combines the answer time, pause and other behavior characteristics, and outputs P' through a nonlinear function. If P' is higher than the high threshold T2, it is marked as "highly suspicious"; if P' is between T1 and T2, it is marked as "moderately suspicious".

[0157] In some embodiments, the global suspiciousness score P' is set between 0 and 1 (the higher the more suspicious). The preset threshold interval is divided as follows:

[0158] Normal interval: 0 ≤ P' < T1; moderately suspicious interval: T1 ≤ P' < T2; highly suspicious interval: T2 ≤ P' ≤ 1;

[0159] Wherein, T1 represents the boundary point between the normal interval and the moderately suspicious interval, and T2 represents the boundary point between the moderately suspicious interval and the highly suspicious interval.

[0160] Figure 6 A schematic diagram showing the interviewee interface detection result according to the embodiments of the present application is shown. As shown in the figure, the voice collection and transcription module performs voice recognition on the audio and video streams of the interviewer and the candidate, obtains the interview question texts and candidate answer texts arranged in chronological order, the AI recognition control panel is used to real-time prompt the suspiciousness level, the global suspiciousness score P', and the similarity score of recent correction verification to the interviewer, and the detection AI automatically starts the verification mode to generate at least one verification follow-up question "trap" question, which is displayed to the interviewer through the interviewer interface to ask the candidate; the global suspiciousness score P'' is updated according to the candidate's verification answer, and the updated suspiciousness level is fed back to the interviewer. Figure 6

[0161] ​In an embodiment of the present application, generating and presenting the verification follow-up trap question based on the detection of the AI starting verification mode includes the following steps S3061-S3064.

[0162] Step S3061, based on the corrected similarity score of each round of interview question, the corresponding model consistency index and the behavior characteristics, the suspicious source type is obtained.

[0163] Step S3062, based on the suspicious source type, a target trap question template is selected from a preset trap question template library, and a large language model is called to generate a candidate trap question set based on the candidate information; the candidate information at least includes resume information and answer text entity information.

[0164] In some embodiments, the preset trap question template library includes several trap question models, such as “require specific project details” “require quantitative indicators” “require restatement from another perspective” and the like. The target trap question template is a trap question model selected according to the suspicious source type, candidate information (e.g., resume information, answer text entity information). The answer text entity information includes but is not limited to project name, company name, technology stack, time, indicator and the like.

[0165] Step S3063, based on the candidate trap question set and the candidate information, the discrimination score of each candidate trap question is calculated.

[0166] Step S3064, based on the discrimination score of each candidate trap question, the question with the highest discrimination score is selected as the verification follow-up trap question, and is presented to the interviewer; the verification follow-up trap question includes at least one candidate trap question.

[0167] In some embodiments, the verification follow-up question in the verification mode of the present application is automatically generated by the detection AI, and is displayed through the interviewer interface, and is asked to the candidate after being selected by the interviewer.

[0168] In some embodiments, the present application provides an embodiment of triggering the verification mode and AI-generated trap questions, in which the detection AI receives the transcribed text and behavior characteristics in real time during the interview as an independent service, calculates the modified score S'i for each round of answer without changing the original human-computer dialogue structure, and calculates the current suspicious degree P' based on the sliding window or global statistics. When P' enters the moderate suspicious interval from the normal interval, the verification question generation module is automatically triggered:

[0169] First, the verification question generation module reads the candidate's resume abstract and the content abstract of the last several rounds of questions and answers.

[0170] Second, design one or several "trick" verification question templates according to the main source of suspicion (for example, some questions are highly similar to the LLM answer), such as asking the candidate to give more detailed implementation steps or failure cases for the project mentioned above.

[0171] Third, after checking the length and understandability of the generated verification questions, the question text is pushed to the "AI suggested questions" area of the interviewer's end UI, and the interviewer selects one of the questions and asks the candidate in his own language or directly reads it.

[0172] Fourth, after the candidate answers, the system calculates the similarity and MCI of the verification question answer as described in the above embodiment to obtain the modified score S̃_ver of the verification question, and then updates the global suspiciousness P'' according to the preset weight. Through the above way, the system realizes the feedback loop of "AI monitoring + AI question generation + HR asking + AI re-judging", which effectively improves the recognition ability of AI assisted answers without depriving the interviewer of the dominant right.

[0173] In some embodiments, the generation and quantitative screening algorithm of the verification "trick" follow-up questions of the present application is as follows:

[0174] When the global suspiciousness P' falls into the moderate suspiciousness interval, the system enters the automatic verification mode, and the detection AI automatically completes the following steps to generate and screen verification "trick" follow-up questions:

[0175] First, suspicious source analysis is needed: the detection AI identifies the suspicious source type based on the modified score S_tilde_i of the multi-round answer, the corresponding MCI_i, and the behavior characteristics, for example:

[0176] Type A: the answer has a high similarity to the model reference answer and a high MCI, suspected "standard answer AI answer".

[0177] Type B: the answer is inconsistent with the project experience in the candidate's resume or is too general.

[0178] Type C: the language style of multi-round answers is highly uniform and textbook-like, and the behavior characteristics are abnormal.

[0179] Second, candidate question generation: according to the suspicious source type, select the corresponding trap question template (for example, "require specific project details" "require restatement from another perspective" etc.), and combine entity information (for example, project name, company name, technology stack, etc.) in the candidate's resume and dialogue history, call the large language model to generate a set of candidate trap questions: Q = {q_1, q_2, …, q_K}.

[0180] Quantitative scoring indicators of candidate trap questions: Calculate the following indicators for each candidate question q_k, and further calculate the discrimination score D(q_k) of the question based on these indicators for screening the optimal trap question.

[0181] Resume relevance score R_k: Extract entity information set E_cv (including but not limited to company, project, position, etc.) from candidate resume and historical answers, extract entity set E(q_k) from candidate question q_k, where R_k is calculated as follows:

[0182] R_k = |E(q_k)∩E_cv| / ( |E(q_k)| + epsilon )

[0183] Where epsilon is a constant to prevent division by zero, and the larger R_k, the more the question fits the candidate's personal experience, and the less friendly it is to AI large language models.

[0184] Model consistency score M_k: Input the candidate question q_k into multiple large language models to obtain the reference answer set {u_k1, …, u_km}, vectorize as e(u_kj), and calculate the cosine similarity between each other, M_k is calculated as follows:

[0185] M_k = ( 2 / ( m * (m - 1) ) ) *Σ_{j1<j2} cos_sim( e(u_kj1), e(u_kj2) )

[0186] Where M_k is larger, indicating that different models answer the question more consistently, closer to the "standard answer question", and less likely to distinguish AI from real experience, so it is used as a negative factor in the comprehensive score.

[0187] Suspicious content targeting score C_k: Let the suspicious round set be I_susp, where each round of answer semantic vector is h_i. Calculate the semantic vector g_k for the candidate question q_k, and C_k is calculated as follows:

[0188] C_k=max_{i∈I_susp} cos_sim( g_k, h_i )

[0189] Where C_k is larger, indicating that the question is more "aligned" with the previous high-suspicious answer in semantics.

[0190] Detail requirement score L_k (optional indicator): Predefine several "detail slots" in the question template, such as whether to require specific time, quantitative indicators, number of key steps, etc. Count the number of activated detail slots s_k in q_k, set the maximum slot number S_max, and L_k is calculated as follows:

[0191] L_k = s_k / S_max

[0192] The larger L_k is, the higher the requirement for the details of the answer to the problem is, and the more beneficial it is to distinguish between real experience and generalized answers.

[0193] Based on the above indicators, a discrimination score D(q_k) is defined for each candidate question q_k:

[0194] D(q_k) = sigmoid(α* R_k +β* C_k +γ* L_k -δ* M_k )

[0195] wherein sigmoid(x) is a Sigmoid function or other monotonically increasing normalization function, used to map the linear combination in the parentheses to the interval [0, 1], so that D(q_k) falls between 0 and 1; and α, β, γ, δ>0 are configurable weights, used to balance the influence of each component.

[0196] The system calculates D(q_k) for all candidate questions, and selects the question or questions with the highest discrimination score as the final verification trap question, for example:

[0197] q* = argmax_{q_k∈Q}D(q_k)

[0198] Or the top N questions that satisfy D(q_k)≥T_D are selected and pushed to the "AI suggested question" area on the interviewer's side, and the interviewer asks the candidate during the interview. After the candidate answers, the system calculates the similarity and suspiciousness of the question again to update the global suspiciousness P''.

[0199] Through the above quantitative indicators and comprehensive scores, the verification trap question in the present application is not a simple follow-up question based on insufficient scores, but is specially designed for the "suspected AI answering" mode, and the discrimination ability is quantitatively evaluated through R_k, C_k, L_k, M_k, etc. so as to enhance the recognition effect of external AI auxiliary answering at the algorithm level.

[0200] The protection scope of the interview auxiliary answering detection method based on multi-model comparison described in the embodiments of the present application is not limited to the execution order of the steps listed in the embodiments. Any scheme realized by adding, replacing or replacing steps of the prior art according to the principle of the present application is included in the protection scope of the present application.

[0201] The embodiment of the present application also provides an interview auxiliary answer detection system based on multi-large model comparison. The interview auxiliary answer detection system based on multi-large model comparison can realize the interview auxiliary answer detection method based on multi-large model comparison. However, the implementation device of the interview auxiliary answer detection method based on multi-large model comparison includes but is not limited to the structure of the interview auxiliary answer detection system based on multi-large model comparison listed in the embodiment. Any structural deformation and replacement of the prior art according to the principle of the present application is included in the protection scope of the present application.

[0202] As shown in Figure 7 The embodiment provides an interview auxiliary answer detection system based on multi-large model comparison. The system 700 includes a voice collection and transcription module 701, a multi-large model calling module 702, a similarity calculation module 703, a model consistency analysis module 704, a suspicious degree evaluation module 705, a verification question generation module 706, and an interviewer user interface and prompt module 707.

[0203] The voice collection and transcription module 701 is configured to collect the audio and video streams of the interviewer and the candidate during the remote interview process, and perform voice recognition on the audio and video streams to obtain multi-round question and answer transcription texts. The question and answer transcription texts include interview question texts and answer texts.

[0204] The multi-large model calling module 702 is configured to call at least two large language models based on each round of interview question texts to obtain a reference answer text set for each round of interview question. The reference answer text set includes reference answer texts of each large language model.

[0205] The similarity calculation module 703 is configured to calculate the model similarity scores of the corresponding interview questions based on each answer text and the corresponding reference answer text set, and obtain the global suspicious degree score of the candidate according to the model similarity scores of each round of interview questions.

[0206] The model consistency analysis module 704 is configured to perform consistency calculation based on the reference answer text set of each round of interview questions to obtain the model consistency index between the reference answers of each large language model in each round of interview questions.

[0207] The suspicious degree evaluation module 705 is configured to negatively correct the model similarity score and the global suspicious degree score of the current interview question based on the consistency index between the reference answers of each large language model in each round of interview questions, and obtain the corrected global suspicious degree score.

[0208] The verification question generation module 706 is configured to compare the corrected global suspiciousness score with a preset threshold interval, and if the corrected global suspiciousness score is in a preset suspicious interval, generate a verification follow-up trap question and display it to the interviewer.

[0209] The interviewer user interface and prompt module 707 is configured to obtain a verification answer of the candidate based on the verification follow-up trap question, update the global suspiciousness score according to the verification answer of the candidate, and feed back the updated global suspiciousness score to the interviewer.

[0210] It should be noted that the functions or operations of the voice collection and transcription module 701, the multi-model calling module 702, the similarity calculation module 703, the model consistency analysis module 704, the suspiciousness evaluation module 705, the verification question generation module 706, and the interviewer user interface and prompt module 707 described in the embodiments of the present disclosure correspond one by one to the steps in the interview auxiliary answer detection method based on multi-model comparison described above, and therefore will not be repeated here.

[0211] In several embodiments provided in the present application, it should be understood that the disclosed system, device or method can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the modules / units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of modules or units can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed ones can be indirect coupling or communication connection through some interfaces, devices or modules or units, and can be electrical, mechanical or other forms.

[0212] The modules / units described as separate components can or can not be physically separated, and the components displayed as modules / units can or can not be physical modules, i.e., can be located in one place, or can be distributed on a plurality of network units. Part or all of the modules / units can be selected according to actual needs to achieve the purpose of the embodiments of the present application. For example, the functional modules / units in each embodiment of the present application can be integrated in one processing module, or each module / unit can be physically separated, or two or more modules / units can be integrated in one module / unit.

[0213] Those skilled in the art should further understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software, or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been described in the above description in a general manner. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0214] Figure 8 A structural schematic diagram of an electronic device 800 described in embodiments of the present application is shown. As shown, the electronic device 800 in the embodiments includes a memory 801 and a processor 802. Figure 8

[0215] The memory 801 is used to store a computer program; preferably, the memory 801 includes: ROM, RAM, a disk, a U disk, a memory card, or an optical disk, and various media that can store program codes.

[0216] Specifically, the memory 801 can include a computer system readable medium in the form of a volatile memory, such as a random access memory (RAM) and / or a cache memory. The electronic device 800 can further include other removable / non-removable, volatile / non-volatile computer system storage media. The memory 801 can include at least one program product having a set (for example, at least one) program modules configured to perform the functions of the embodiments of the present application. It can be understood that the memory 801 can be a volatile memory or a non-volatile memory, and can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (Read Only Memory, ROM) and a programmable read-only memory (Programmable Read-Only Memory, PROM), which is used as an external cache. By way of example but not limitation, many forms of RAM can be used, such as static random access memory (Static Random Access Memory, SRAM), synchronous static random access memory (Synchronous Static Random Access Memory, SSRAM). The memory described in the embodiments of the present application is intended to include but not limited to these and any other suitable categories of memory.

[0217] ​The processor 802 is connected with the memory 801, and is configured to execute a computer program stored in the memory 801, so that the electronic device 800 executes the interview auxiliary answer detection method based on multi-model comparison according to any of the embodiments of the present application.

[0218] Optionally, the processor 802 can be a general processor, including a central processing unit (CPU), a network processor (NP), etc.; and can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0219] Optionally, the electronic device 800 in the embodiment can further include a display 803. The display 803 is connected with the memory 801 and the processor 802 in communication, and is configured to display a related graphical user interface (GUI) interactive interface of the interview auxiliary answer detection method based on multi-model comparison according to the embodiments of the present application.

[0220] The embodiments of the present application further provide a computer readable storage medium. Those skilled in the art can understand that all or part of the steps of the methods described in the above embodiments can be completed by a program instructing a processor, and the program can be stored in a computer readable storage medium, which is a non-transitory medium, such as a random access memory, a read only memory, a flash memory, a hard disk, a solid state disk, a magnetic tape, a floppy disk, an optical disc and any combination thereof. The storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, a data center, etc. integrated with one or more available medium sets. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a digital video disc (DVD)) or a semiconductor medium (for example, a solid state disk (SSD)) etc.

[0221] The description of the flow or structure corresponding to each of the above figures has its own emphasis, and the parts not described in detail in a certain flow or structure can be referred to the related description of other flows or structures.

[0222] The above embodiments are only illustrative of the principles and effects of the present application, and are not intended to limit the present application. Any modification or change made by those skilled in the art without departing from the spirit and scope of the present application shall be covered by the claims of the present application.

Claims

1. An interview auxiliary answer detection method based on multi-large model comparison, characterized in that, The method comprises the following steps: Collecting audio and video streams of an interviewer and a candidate during a remote interview process, and performing speech recognition on the audio and video streams to obtain a multi-round question and answer transcription text; The question and answer transcription text includes interview question text and answer text; Based on each round of interview question text, at least two large language models are called to obtain a reference answer text set for each round of interview question; the reference answer text set includes reference answer texts of each large language model; Based on each answer text and the corresponding reference answer text set, the model similarity score of the corresponding interview question is calculated, and the global suspiciousness score of the candidate is obtained according to the model similarity score of each round of interview question; Based on the reference answer text set of each round of interview question, consistency calculation is performed to obtain the model consistency index between the reference answers of each large language model in each round of interview question; Based on the consistency index between the reference answers of each large language model in each round of interview question, the model similarity score and the global suspiciousness score of the current interview question are negatively corrected to obtain a corrected global suspiciousness score; Based on the comparison between the corrected global suspiciousness score and the preset threshold interval, if the corrected global suspiciousness score is in the preset suspicious interval, a verification follow-up trap question is generated and displayed to the interviewer; Based on the verification follow-up trap question, the verification answer of the candidate is obtained, the global suspiciousness score is updated according to the verification answer of the candidate, and the updated global suspiciousness score is fed back to the interviewer.

2. The interview answer detection method based on multi-model contrast according to claim 1, characterized in that, Based on each answer text and the corresponding reference answer text set, the model similarity score of the corresponding interview question is calculated, and the global suspiciousness score of the candidate is obtained according to the model similarity score of each round of interview question, which comprises: Based on each answer text and the corresponding reference answer text set, semantic similarity calculation is performed to obtain the similarity value of each answer text and the corresponding reference answer text; Based on the similarity value of each answer text and the corresponding reference answer text, the model similarity score of the corresponding interview question is calculated; Based on the model similarity score of each round of interview question, the global suspiciousness score of the candidate is obtained by summarizing.

3. The interview answer detection method based on multi-model contrast according to claim 2, characterized in that, Based on each answer text and the corresponding reference answer text set, semantic similarity calculation is performed to obtain the similarity value of each answer text and the corresponding reference answer text, which comprises: Based on a semantic vectorization model, each answer text and the corresponding reference answer text set are mapped to a vector space to obtain a candidate answer text vector and a reference answer text vector; Based on the candidate answer text vector and the reference answer text vector, semantic similarity calculation is performed to obtain the similarity value of each answer text and the corresponding reference answer text; the similarity value is the cosine similarity of each answer text and the corresponding reference answer text; sij=(e(ai)*e(rij)) / (||e(ai)||*||e(rij)||) wherein sij represents the cosine similarity between the ith round of candidate answer text and the jth reference answer text of the large language model, ai represents the ith round of candidate answer text, rij represents the jth reference answer text of the ith round of large language model, e(ai) represents the vector of the candidate answer text ai, and e(rij) represents the vector of the reference answer text rij.

4. The interview answer detection method based on multi-model contrast according to claim 2, characterized in that, The model similarity score of the interview question is obtained by weighted average of the weights of each large language model, including: Si = Σ(j=1→k) wj * sij, and Σ(j=1→k) wj = 1; wherein wj represents the weight of the jth large language model of the current interview question, Si represents the model similarity score of the ith round of interview question, sij represents the cosine similarity between the ith round of candidate answer text and the jth reference answer text of the large language model, and k represents the number of large language models.

5. The interview answer detection method based on multi-model contrast according to claim 2, characterized in that, The global suspiciousness score of the candidate is obtained by calculating the model similarity score of each round of interview question through a nonlinear function, including: P = σ(α·S̄ + β·Var(S) + γ·Fbehavior), wherein S = {S1, S2, …, Sn} represents the sequence of model similarity scores of each round of interview question, Si represents the model similarity score of the ith round of interview question, S̄ represents the mean of Si, Var(S) represents the variance of the similarity score sequence S, Fbehavior represents the numerical features extracted according to the answering time and pause frequency behavior characteristics, σ(·) represents the Sigmoid function, α, β, γ represent trainable parameters, and P represents the global suspiciousness score.

6. The interview answer detection method based on multi-model contrast according to claim 1, characterized in that, The model consistency index is calculated by the mean or weighted mean of the pairwise similarity of each reference answer text of the current interview question, and is introduced as a negative correction factor when calculating the model similarity score of a single round of interview question or the global suspiciousness score; The model similarity score of the current interview question is negatively corrected based on the consistency index between the reference answers of each large language model in each round of interview question, and the corrected global suspiciousness score is obtained, including: The model similarity score of each interview question is negatively corrected based on the consistency index between the reference answers of each large language model in each round of interview question, and the corrected model similarity score of each interview question is obtained. The corrected global suspiciousness score is obtained by summarizing the corrected model similarity scores of each interview question.

7. The interview answer detection method based on multi-model contrast according to claim 6, characterized in that, The calculation method of the corrected model similarity score includes: S̃i = Si - λ·MCIi, or S̃i = Si·(1 - λ·MCIi), wherein λ represents a preset or trained correction coefficient, Si represents the model similarity score of the ith round of interview question before correction, S̃i represents the model similarity score of the ith round of interview question after correction, and MCIi represents the model consistency index of the ith round of interview question.

8. The interview answer detection method based on multi-model contrast according to claim 1, characterized in that, Generating and displaying a verification follow-up question trap question to the interviewer based on the detection of AI starting the verification mode includes: Identify based on the corrected similarity score of each round of interview question, the corresponding model consistency index and the behavior characteristics, obtain the suspicious source type; Based on the suspicious source type, select the target trap question template from the preset trap question template library, and generate the candidate trap question set based on the large language model based on the candidate information; The candidate information at least includes resume information, answer text entity information; Based on the candidate trap question set and the candidate information, calculate the discrimination score of each candidate trap question; Based on the discrimination score of each candidate trap question, select the question with the highest discrimination score as the verification follow-up trap question, and display it to the interviewer; The verification follow-up trap question includes at least one candidate trap question.

9. An interview auxiliary answer detection system based on multi-large model comparison, characterized in that, Comprise: Voice collection and transcription module, configured to collect the audio and video stream of the interviewer and the candidate in the remote interview process, and perform voice recognition on the audio and video stream to obtain multi-round question and answer transcription text; The question and answer transcription text includes interview question text and answer text; Multi-model calling module, configured to call at least two large language models based on each round of interview question text to obtain a set of reference answer texts for each round of interview question; The reference answer text set includes the reference answer text of each large language model; Similarity calculation module, configured to calculate the model similarity score of the corresponding interview question based on each answer text and the corresponding reference answer text set, and obtain the global suspiciousness score of the candidate according to the model similarity score of each round of interview question; Model consistency analysis module, configured to calculate the consistency of each round of interview question based on the reference answer text set, and obtain the model consistency index between the reference answers of each large language model in each round of interview question; Suspiciousness evaluation module, configured to correct the model similarity score and the global suspiciousness score of the current interview question based on the consistency index between the reference answers of each large language model in each round of interview question, and obtain the corrected global suspiciousness score; Verification question generation module, configured to compare the corrected global suspiciousness score with the preset threshold interval, if the corrected global suspiciousness score is in the preset suspicious interval, generate the verification follow-up trap question and display it to the interviewer; The interviewer user interface and prompt module is configured to obtain the verification answer of the candidate based on the verification follow-up trap question, update the global suspiciousness score according to the verification answer of the candidate, and feed back the updated global suspiciousness score to the interviewer.

10. An electronic device, comprising: Comprise: Memory, storing a computer program; Processor, connected in communication with the memory, calling the computer program to execute the interview auxiliary answer detection method based on the multi-large model comparison in any one of claims 1 to 8.

Images