Method, system and equipment for detecting interview auxiliary answering based on comparison of multiple models

Abstract

The invention provides an interview auxiliary answering detection method, system and device based on multi-model comparison, and the method comprises the steps: collecting an audio and video stream of a remote interview process, and carrying out the voice recognition of the audio and video stream, and obtaining a multi-round question and answer transcription text; calling at least two large language models based on each round of 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 suspicion score according to the model similarity score; obtaining a model consistency index based on the reference answer text set; performing negative correction on the model similarity score and the global suspicion score based on the model consistency index to obtain a corrected global suspicion score; and if the corrected global suspicious degree score is in a preset suspicious interval, generating a verification question-asking trap question and displaying the question to the interviewer for asking questions, updating the global suspicious degree score according to the verification answer of the candidate, and feeding back the updated global suspicious degree score to the interviewer.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence and online interview technology, and relates to a method, system and device for detecting interview-assisted responses based on multi-model comparison. Background Technology

[0002] In recent years, with the widespread adoption of remote work and online video interviews becoming the new normal in the recruitment process, the interview environment has changed significantly. Against this backdrop, a new challenge has emerged: candidates can, in real-time and covertly, use external artificial intelligence tools (especially large language models) to generate, organize, or refine their answers to interview questions. This behavior means that the knowledge level, logical thinking ability, and language expression skills presented by the candidate do not accurately reflect their true abilities, making it difficult for interviewers to make accurate assessments. This significantly increases the risk of incorrect hiring and potential human resource costs for companies.

[0003] To address the challenges posed by AI-generated text, the industry has proposed various detection technologies. However, these existing methods reveal significant limitations and shortcomings when faced with the specific scenario of real-time interactive interviews. Specifically, existing AI-generated text detection methods are mostly designed for static text and are not suitable for multi-turn dialogues, real-time detection, or closed-loop evaluation combined with verification questions. Furthermore, when "standard answer questions" exist (such as certain technical formulaic questions or textbook-style Q&A), simply relying on the similarity between the candidate's answer and the answer from a single model can easily lead to false positives. Summary of the Invention

[0004] This application provides a method, system, and device for detecting interview-assisted responses based on multi-model comparison, which addresses the problem that existing detection methods are difficult to apply to real-time interactive interview scenarios and are prone to high false positive rates.

[0005] Firstly, this application provides a method for detecting assisted responses in interviews based on multi-model comparison. The method includes: acquiring audio and video streams of the interviewer and candidate during a remote interview, and performing speech recognition on the audio and video streams to obtain multi-round question-and-answer transcribed text; the transcribed text includes interview question text and answer text; calling at least two large language models based on the interview question texts of each round to obtain a reference answer text set for each round of interview questions; the reference answer text set includes reference answer texts for each of the large language models; calculating and obtaining a model similarity score for the corresponding interview question based on each answer text and the corresponding reference answer text set, and obtaining a global suspicion score for the candidate based on the model similarity scores of each round of interview questions; and based on the interview questions of each round... Consistency calculation is performed on the reference answer text set to obtain the model consistency index among the reference answers of major language models in each round of interview questions; based on the consistency index among the reference answers of major language models in each round of interview questions, the model similarity score and global suspicion score of the current interview question are negatively corrected to obtain the corrected global suspicion score; based on the corrected global suspicion score and a preset threshold range, if the corrected global suspicion score is within the preset suspicion range, a verification follow-up question is generated and shown to the interviewer; based on the verification follow-up question, the candidate's verification answer is obtained, and the global suspicion score is updated according to the candidate's verification answer, and the updated global suspicion score is fed back to the interviewer.

[0006] In one implementation of the first aspect, calculating the model similarity score of the corresponding interview question based on each answer text and the corresponding reference answer text set, and obtaining the candidate's global suspicion score based on the model similarity scores of each round of interview questions includes: calculating the similarity value between each answer text and the corresponding reference answer text based on the semantic similarity calculation based on each answer text and the corresponding reference answer text set; calculating the model similarity score of the corresponding interview question based on the similarity value between each answer text and the corresponding reference answer text; and summarizing the model similarity scores of each round of interview questions to obtain the candidate's global suspicion score.

[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 one implementation of the first aspect, the model consistency index is calculated using the pairwise similarity mean or weighted mean of the reference answer texts for the current interview question. This index is introduced as a negative correction factor when calculating the model similarity score or global suspicion score for a single round of interview questions. The model similarity score and global suspicion score for the current interview question are negatively corrected based on the consistency index between the reference answers of various language models in each round of interview questions. Obtaining the corrected global suspicion score involves: negatively correcting the model similarity score for the current interview question based on the consistency index between the reference answers of various language models in each round of interview questions, and obtaining the corrected model similarity score for each interview question; and summarizing the corrected model similarity scores for each interview question to obtain the corrected global suspicion score.

[0011] In one implementation of the first aspect, the calculation method for the corrected model similarity score includes: S̃i = Si-λ·MCIi, or S̃i = Si·(1 -λ·MCIi), Where λ represents the preset or trained correction coefficient, Si represents the model similarity score of the i-th round of interview questions before correction, S̃i represents the model similarity score of the i-th round of interview questions after correction, and MCIi represents the model consistency index of the i-th round of interview questions.

[0012] In one implementation of the first aspect, generating and displaying verification follow-up trap questions based on the AI-initiated verification mode includes: identifying suspicious source types based on the corrected similarity scores, corresponding model consistency indices, and behavioral characteristics of the interview questions from each round; selecting target trap question templates from a preset trap question template library based on the suspicious source types, and generating a candidate trap question set by calling a large language model based on candidate information; the candidate information includes at least resume information and answer text entity information; calculating the 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 and displaying it to the interviewer; the verification follow-up trap question includes at least one candidate trap question.

[0013] Secondly, this application provides an interview-assisted answer detection system based on multi-model comparison. The system includes: a speech acquisition and transcription module configured to acquire audio and video streams of the interviewer and candidate during a remote interview, and to 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; a multi-model invocation module configured to invoke at least two large language models based on the interview question texts of each round, obtaining a reference answer text set for each round of interview questions; the reference answer text set includes reference answer texts for each of the large language models; a similarity calculation module configured to calculate and obtain a model similarity score for the corresponding interview question based on each answer text and the corresponding reference answer text set, and to obtain a global suspicion score for the candidate based on the model similarity score of each round of interview questions; and a model consistency analysis module configured to analyze the consistency of each interview question and the candidate's answers based on the model similarity scores of each round of interview questions; and a model consistency analysis module configured to perform analysis on each interview question and the candidate's answers based on the model similarity scores of each round of interview questions. The system performs consistency calculations on the reference answer text sets of each round of interview questions to obtain the model consistency index among the reference answers of various language models in each round of interview questions; a suspicion assessment module is configured to negatively correct the model similarity score and global suspicion score of the current interview question based on the consistency index among the reference answers of various language models in each round of interview questions, and obtain the corrected global suspicion score; a verification question generation module is configured to compare the corrected global suspicion score with a preset threshold range, and if the corrected global suspicion score is within the preset suspicion range, generate a verification follow-up question and display it to the interviewer; an interviewer user interface and prompt module is configured to obtain the candidate's verification answer based on the verification follow-up question, update the global suspicion score based on the candidate's verification answer, and feed back the updated global suspicion score to the interviewer.

[0014] Thirdly, this application provides an electronic device, which includes: a memory storing a computer program; and a processor communicatively connected to the memory, which executes the interview-assisted answer detection method based on multi-model comparison as described above when the computer program is invoked.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the aforementioned interview-assisted answer detection method based on multi-model comparison.

[0016] As described above, the interview-assisted response detection method, system, and device based on multi-model comparison described in this application have the following beneficial effects: This application improves the accuracy and practicality of judgment by introducing a multi-model consensus mechanism, negative correction of the model consistency index, and a verification mode. It can sensitively detect AI traces in open-ended questions and suppress false alarms in standard-answer questions. This application can provide real-time, quantitative prompts during the interview process to indicate whether candidates are suspected of using AI tools to assist in answering questions, and reduce the false alarm rate through MCI correction, thereby improving the credibility and fairness of the remote recruitment process. It solves the problem that existing detection methods are difficult to apply to real-time interactive interview scenarios and are prone to high false alarm rates.

[0017] The detection AI in this application works as an independent module that listens to voice and text in the background. When the detection result reaches a moderate level of suspicion, it automatically generates verification follow-up questions and prompts the interviewer to ask them through the interviewer's user interface (UI), realizing a collaborative working method of "AI generating questions and HR asking questions". Attached Figure Description

[0018] Figure 1 The diagram shown is a schematic representation of the hardware structure of the interview-assisted response detection method based on multi-model comparison as described in an embodiment of this application.

[0019] Figure 2 The diagram shown is a structural schematic of the remote interview system described in an embodiment of this application.

[0020] Figure 3 The diagram shown is a flowchart of the interview-assisted answer detection method based on multi-model comparison described in the embodiments of this application.

[0021] Figure 4 The diagram shown is a flowchart illustrating the verification mode and feedback loop described in the embodiments of this application.

[0022] Figure 5 The diagram shown is a partial detection result of a simulated interview as described in an embodiment of this application.

[0023] Figure 6 A schematic diagram showing the detection results of the interviewer interface described in the embodiments of this application.

[0024] Figure 7 This diagram shows the structure of the interview-assisted answer detection system based on multi-model comparison as described in an embodiment of this application.

[0025] Figure 8 This diagram shows a schematic representation of the electronic device described in an embodiment of this application.

[0026] Component designation explanation 100 terminal equipment 800 electronic devices 101 Processing unit 801 memory 102 storage unit 802 processor 103 Input devices 803 monitor 104 Output devices S301~S307 step 700 Interview Response Detection System Based on Multi-Model Comparison S401~S406 step 701 Voice Acquisition and Transcription Module 702 How large a model calls the module 703 Similarity calculation module 704 Model Consistency Analysis Module 705 Suspicion assessment module 706 Verification question generation module 707 Interviewer User Interface and Prompt Module Detailed Implementation

[0027] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0028] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0029] The following embodiments of this application provide a method, system, and device for detecting interview-assisted responses based on multi-model comparison, which solves the problems that existing detection methods are difficult to apply to real-time interactive interview scenarios and are prone to high false positive rates.

[0030] This application provides a method, system, and device for detecting interview-assisted responses based on multi-model comparison. This method allows for real-time, quantitative evaluation of whether candidates are using external AI tools without significantly disrupting the interview process. By introducing a multi-model consensus mechanism, negative correction using the Model Consistency Index (MCI), and a verification mode, it can sensitively detect AI traces in open-ended questions and suppress false positives in standard-answer questions, thereby improving accuracy and practicality. In the specific implementation, the interview questions and answers are still conducted directly between the interviewer and the candidate. The detection AI, as an independent module, listens to the audio and text in the background. When the detection result reaches a moderately suspicious level, it automatically generates verification follow-up questions and prompts the interviewer to ask them via the UI, achieving a collaborative working method of "AI generating questions, HR asking questions."

[0031] The interview-assisted response detection method based on multi-model comparison provided in this application can run on a terminal device. Figure 1 For example, Figure 1 This is a hardware block diagram of a terminal device for running the interview-assisted answer detection method based on multi-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 via a bus.

[0032] Storage unit 102 is the non-transitory computer-readable storage medium provided in this application. The memory stores instructions executable by at least one processor to cause the at least one processing unit 101 to execute the interview-assisted answer detection method based on multi-model comparison provided in this application. The non-transitory computer-readable storage medium of this application stores computer instructions for causing a computer to execute the interview-assisted answer detection method based on multi-model comparison provided in this application.

[0033] Storage unit 102 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function. The data storage area may store audio and video stream data required by the multi-model comparison-based interview-assisted response detection method, and data created based on the use of electronic devices determined by remote interview-assisted response detection. Furthermore, storage unit 102 may include high-speed random access memory and may also include non-transient memory, such as at least one disk storage device, flash memory device, or other non-transient solid-state storage device. In some embodiments, storage unit 102 may optionally include memory remotely located relative to processing unit 101. This remote memory can be connected via a network to the terminal device 100 determined by the multi-model comparison-based interview-assisted response detection method. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0034] The terminal device 100 also 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 streams of candidates and interviewers. These audio and video streams can be stored in the storage unit 102, so that the processing unit 101 can perform detection calculations for remote interview assistance based on these audio and video streams, and output the calculated detection results through the output device 104.

[0035] The input device 103 may include, but is not limited to, fisheye cameras and cameras. The output device 104 may include, but is not limited to, liquid crystal displays (LCDs), light-emitting diode displays (LEDs), plasma displays, and touch screens. This application does not limit the scope of the embodiments.

[0036] In the embodiments of this application, the aforementioned components of the terminal device 100 and Figure 1 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 1 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this application. Those skilled in the art can add or replace other components as needed.

[0037] Terminal device 100 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs. Terminal device 100 can also be a mobile or stationary server. This application does not limit the scope of the embodiments.

[0038] Figure 2 The diagram shown is a structural schematic of the remote interview system described in an embodiment of this application. Figure 2 As shown, 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 speech acquisition and transcription module, a multi-model calling module, a similarity calculation module, a model consistency analysis module, a suspiciousness assessment module, a verification question generation module, and an interviewer prompting module. During the actual interview, the candidate and interviewer establish an audio-visual connection via protocols such as WebRTC. The candidate's speech is sent in real-time to the speech recognition module for transcription into text. The interviewer's questions are transcribed into qi (interview questions), and the candidate's answers are transcribed into ai (interview questions). The detection AI does not directly participate in the question-and-answer process but acts as a side-channel listener, receiving the transcribed text and behavioral features for subsequent suspiciousness assessment and verification mode triggering.

[0039] In some embodiments, the model set configured by the large model calling module includes, but is not limited to: GPT series models, Claude series models, Gemini series models, and third-party large language model services such as DeepSeek, Grok, and Kimi. The system can select any two or more models from the above model set to participate in the deliberation based on business needs and actual availability; in some embodiments, 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.

[0040] The technical solutions in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0041] like Figure 3 As shown, this embodiment provides a method for detecting interview-assisted responses based on multi-model comparison, the method including the following steps S301 to S307.

[0042] Step S301: Collect the audio and video streams of the interviewer and the candidate during the remote interview, 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.

[0043] Specifically, the multi-round question-and-answer transcript consists of interview questions and candidate responses arranged in chronological order.

[0044] Step S302: Based on the interview question texts of each round, call at least two large language models to obtain the reference answer text set of each round of interview questions; the reference answer text set includes the reference answer text of each of the large language models.

[0045] In some embodiments, this application obtains the corresponding reference answer text set {ri1, ri2, …, rik} by inputting the interview question text qi of each round into large language models from at least two different providers.

[0046] In some embodiments, the large language model includes, but is not limited to, large language models such as the GPT model and the Claude model.

[0047] Step S303: 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 candidate's global suspiciousness score based on the model similarity score of each round of interview questions.

[0048] In one embodiment of this application, the calculation of the model similarity score of the corresponding interview question based on each answer text and the corresponding reference answer text set, and the acquisition of the candidate's global suspicion score based on the model similarity score of each round of interview questions, includes the following steps S3031 to S3033.

[0049] Step S3031: Calculate the semantic similarity between 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.

[0050] Step S3032: Calculate and obtain the model similarity score of the corresponding interview question based on the similarity value between each answer text and the corresponding reference answer text.

[0051] Step S3033: Summarize the model similarity scores based on the interview questions in each round to obtain the candidate's global suspiciousness score.

[0052] In one embodiment of this application, the process of calculating the semantic similarity between each answer text and the corresponding reference answer text to obtain the similarity value between each answer text and the corresponding reference answer text includes the following steps S30311 to S30312.

[0053] Step S30311: Based on the semantic vectorization model, map each answer text and the corresponding reference answer text set to the vector space to obtain the corresponding candidate answer text vector and reference answer text vector.

[0054] Step S30312: Calculate semantic similarity based on the candidate's answer text vector and the reference answer text vector to obtain the similarity value between each answer text and the corresponding reference answer text.

[0055] In some embodiments, this application uses a semantic vectorization model to map the candidate's answer text ai and each reference answer text rij to a vector space, calculates the semantic similarity between the two, and obtains a similarity value sij; constructs a similarity matrix based on the similarity values ​​of each round of questions and the model, calculates the AI ​​similarity score Si for each round of questions, and summarizes all questions to obtain the candidate's global suspiciousness score P.

[0056] Step S304: Perform consistency calculation based on the reference answer text set of each round of interview questions to obtain the model consistency index among the reference answers of major language models in each round of interview questions.

[0057] Step S305: Based on the consistency index among the reference answers of various language models in each round of interview questions, negatively correct the model similarity score and global suspicion score of the current interview question to obtain the corrected global suspicion score.

[0058] In some embodiments, this application calculates the consistency index MCIi between the reference answers of the multiple models in each round of questions, and uses MCIi as a negative correction factor to correct Si or P, so as to reduce false alarms for standard answer type questions.

[0059] Specifically, negative correction is a protective mechanism used to reduce false positive rates in standard-answer questions. Negative correction indicates that a candidate has answered a question with the standard answer, which is consistent with the reference answer of the AI ​​large language model, but it does not mean that the candidate has cheated using AI.

[0060] The main purpose of negative correction is to "reduce noise" and avoid misclassifying candidates who genuinely recite standard answers as AI-generated responses. Some questions inherently have only one or two typical answer formats, such as "What is the average time complexity of quicksort?" Whether it's top candidates or GPT / Claude, the answers are very similar. For these types of questions, simply looking at "the candidate's answer is highly similar to the model's answer" will generate many false positives. Therefore, MCI is used to negatively correct the suspicion score for these types of questions.

[0061] Secondly, negative correction can concentrate the system's "sensitivity" on more discriminative open-ended and experience-based questions, such as "What specific performance optimizations have you performed?" AI's answers to these questions are often templated, while real candidates provide details about their individual projects. Reducing the weight of high MCI (Multi-Cost Interpretation) standard questions allows more weight to be allocated to these types of questions.

[0062] In some embodiments, such as interviews containing many basic knowledge questions or formulaic questions (e.g., TCP three-way handshake, B+ tree characteristics, etc.), the answers of excellent candidates are often highly similar to the answers of large models. Without negative correction, each question significantly increases the suspicion level P', easily leading to a situation where "almost all candidates are marked as moderately or highly suspicious on the interface." This results in a decreased ability of the system to distinguish between "real candidates with good memory and thorough preparation" and "candidates who rely entirely on AI," a significant drop in precision, and a decrease in the interviewer's trust in the detection results. Existing technologies primarily check candidates' answers in remote interviews by searching articles and question banks for plagiarism, without a corresponding detection mechanism. Compared to existing technologies, this application reduces the false positive rate in standard answer-type questions through negative correction.

[0063] In some embodiments, this application performs negative correction for scenarios such as standard answer questions and fixed questions. For example, in scenarios such as standard knowledge point questions, algorithm questions, and definition questions, the answers to these questions are highly consistent across multiple models, with an MCI value close to 1. Negative correction can reduce their contribution to the overall suspiciousness. In scenarios such as questions related to official processes or company general specifications, the company documents or public wikis have already fixed these questions, and these questions are more suitable as knowledge point check questions rather than the main basis for AI cheating judgment. In scenarios such as question bank entries that are explicitly marked as "basic knowledge check questions" by the business side, negative correction based on MCI is enabled by default in these scenarios.

[0064] In some embodiments, there are scenarios where negative correction can be weakened or not used. For example, in open-ended behavioral interview questions and personal experience questions, the answers to these questions vary greatly among multiple models, and the MCI itself is low, so the impact on the score is limited. In such scenarios, it is not necessary to introduce a large amount of negative correction. In scenarios that emphasize writing style or long-form writing, if the business goal is to capture AI participation as much as possible rather than protect standard answer questions, MCI negative correction can be used only in knowledge point questions, and the mechanism can be weakened or turned off in advanced writing questions. In resource-constrained initial versions, if only a single model is connected and it is not possible to calculate the MCI of multiple models, negative correction can be temporarily not performed, and the MCI module can be used as an optional feature, which can be enabled later when it is expanded to multi-model collaboration.

[0065] Step S306: Based on the comparison between the corrected global suspicion score and the preset threshold range, if the corrected global suspicion score is within the preset suspicion range, a verification follow-up question trap is generated and shown to the interviewer.

[0066] In some embodiments, the preset threshold range includes a normal range, a moderately suspicious range, and a highly suspicious range, wherein the preset suspicious range is a moderately suspicious range.

[0067] Step S307: Obtain the candidate's verification answer based on the verification follow-up question trap, update the global suspicion score according to the candidate's verification answer, and feed back the updated global suspicion score to the interviewer.

[0068] In some embodiments, this application compares the modified global suspicion score P' with a preset threshold range, dividing it into a normal range, a moderately suspicious range, and a highly suspicious range, and provides the suspicion level to the interviewer in real time. When P' falls into the moderately suspicious range, the detection AI automatically initiates the verification mode, generates at least one verification follow-up "trap" question based on the candidate's resume and previous answers, and displays it to the interviewer through the interviewer's interface, allowing the interviewer to ask the candidate in the dialogue. Steps S302–S305 are repeated for the candidate's verification answer to update the global suspicion score P'', and the updated suspicion level is fed back to the interviewer.

[0069] This application collects audio and video recordings of interviewers and candidates during remote interviews and obtains multi-round question-and-answer transcripts using a speech-to-text module. For each round of interview questions, questions are simultaneously sent to multiple large language models to obtain reference answer texts. Semantic similarity is calculated between the candidate's answers and each reference answer, constructing a similarity matrix across multiple models and rounds of questions, and calculating a suspicion score. Furthermore, the consistency index (MCI) among the reference answers of multiple models is calculated, introducing negative correction into the overall score to suppress false positives for "standard answer" questions. During normal question-and-answer sessions, a detection AI continuously listens to the interview dialogue in a side-channel manner and calculates global suspicion. When the suspicion score enters the moderate suspicion range, a verification mode is automatically activated. The detection AI generates verification "trap" follow-up questions and pushes them to the interviewer, who then asks the candidate these questions in the dialogue. The verification answers are re-evaluated, and the suspicion score is updated, forming a feedback loop. This application can provide real-time, quantitative alerts during the interview process regarding whether candidates are suspected of using AI tools to assist in answering questions, and reduces the false positive rate through MCI correction, thereby improving the credibility and fairness of the remote recruitment process.

[0070] In one embodiment of this application, the similarity value between each answer text and its corresponding reference answer text is the cosine similarity between each answer text and its corresponding reference answer text. The method for calculating the cosine similarity includes: sij=(e(ai)*e(rij)) / (||e(ai)||*||e(rij)||) 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.

[0071] 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: Si =Σ(j=1→k) wj * sij, and Σ(j=1→k) wj=1; 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.

[0072] 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: P =σ(α·S̄ +β·Var(S) +γ·F behavior) 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.

[0073] 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.

[0074] 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.

[0075] Step S3051: Based on the consistency index among the reference answers of various language models in each round of interview questions, negatively correct the model similarity score of the current interview question, and obtain the corrected model similarity score for each interview question.

[0076] Step S3052: Summarize the corrected model similarity scores for each interview question to obtain the corrected global suspicion score.

[0077] In one embodiment of this application, the calculation method for the corrected model similarity score includes: S̃i = Si-λ·MCIi, or S̃i = Si·(1 -λ·MCIi) Where λ is the preset or trained correction coefficient, Si represents the model similarity score of the i-th round of interview questions before correction, S̃i represents the model similarity score of the i-th round of interview questions after correction, and MCIi represents the model consistency index of the i-th round of interview questions.

[0078] In some embodiments, this application provides the calculation and practical examples of the Model Consistency Index (MCI). In this embodiment, in order to suppress false positives of "standard answer questions," this 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-like answers, then the question is a standard answer question. In this case, even if the candidate's answer is highly similar to the model's answer, it should not be directly judged as AI cheating, but its contribution weight in the global suspiciousness should be appropriately reduced.

[0079] For the i-th question, suppose there are k large language models, and the reference answer texts {ri1, ri2, …,rik} are obtained, which are vectorized into e(rij). The model consistency index MCIi for this question can be calculated as follows: MCIi = 2 / (k·(k-1))·Σ(j1 <j2) cos_sim( e(ri j1), e(ri j2) )。

[0080] When MCIi is close to 1, it indicates that there is a high degree of consistency between different models, and the question is a standard answer question; when MCIi is low, it indicates that the answers given by different models are quite different, and it is closer to an open-ended question.

[0081] This application illustrates MCI negative correction using an example. Suppose there are two questions, Q1 and Q2, in an interview, which are: Q1: "Please write down the average time complexity of the quicksort algorithm" (typical standard answer question).

[0082] Q2: "Based on a past project, please explain how you handled a sudden online outage" (open-ended behavioral interview question).

[0083] The system simultaneously uses three major models—GPT, Claude, and DeepSeek—as references: For Q1, the reference answers for the three models are all focused on "O(n log n)" and its standard interpretation. The pairwise cosine similarity between the models is approximately 0.96, 0.95, and 0.97, so MCI1≈0.96; For Q2, the three models provide significantly different cases and descriptions due to different predicted scenarios, with pairwise similarities of approximately 0.55, 0.48, and 0.52, respectively. Therefore, MCI2 ≈ 0.52. Further assuming a weighted similarity score between the candidate's answer and the three models: S1 = 0.90, S2 = 0.82. If a linear negative correction is used: S̃i = Si -λ·MCIi, where λ = 0.4, then: S̃1 = 0.90 - 0.4·0.96 ≈ 0.52 (Standard answer question, actual validity is less likely).

[0084] S̃2 = 0.82 - 0.4·0.52 ≈ 0.61 (Open-ended question, still with a high suspicion score).

[0085] In summary, this application, through MCI negative correction, can significantly suppress false alarms for standard answer questions without sacrificing sensitivity to open-ended questions.

[0086] In some embodiments, the sample code for similarity calculation and MCI correction provided in this application is as follows: import numpy as np # Import the NumPy numerical computing library for vector operations and norm calculations. def cosine_similarity(v1, v2): v1 = np.array(v1) # Convert the input vector v1 into a NumPy array v2 = np.array(v2) # Convert the input vector v2 into a NumPy array # Calculate cosine similarity: (v1 · v2) / (||v1|| * ||v2||), and add 1e-8 to prevent division by zero. return float(v1 @ v2 / (np.linalg.norm(v1) * np.linalg.norm(v2) +1e-8)) def compute_scores_with_mci(question, answer, llm_clients, embed_fn,weights, lambda_mci=0.4): # 1. Generate multi-model reference answers # For each large language model client, call its generate_answer(question) function to get the model's answer to the question. refs = [client.generate_answer(question) for client in llm_clients] # 2. Text Vectorization Encoding # Encode the candidate's answer as a vector e_ans e_ans = embed_fn(answer) # Encode each reference answer text into a vector list e_refs e_refs = [embed_fn(r) for r in refs] # 3. Calculate the similarity between candidate answers and the reference answers of each model. # For each reference answer vector e_r, calculate the cosine similarity between e_ans and e_r to obtain a list of sims. sims = [cosine_similarity(e_ans, e_r) for e_r in e_refs] # Calculate the weighted sum of the similarity scores of each model according to their respective weights to obtain the raw score S for a single round. S = sum(w * s for w, s in zip(weights, sims)) # 4. Calculate the Model Consistency Index (MCI) based on pairwise similarity between models. pair_sims = [] # Used to store the pairwise similarity between all model answer vectors for i in range(len(e_refs)): # Iterate through the first model index i for j in range(i + 1, len(e_refs)): # Iterate through the second model index j (only taking values ​​after i to avoid duplication) # Calculate the cosine similarity between the i-th and j-th reference answer vectors and add it to pair_sims pair_sims.append(cosine_similarity(e_refs[i], e_refs[j])) # Take the average of all pairs of models as the MCI; if the list is empty (insufficient number of models), then the MCI is 0.0. MCI = float(np.mean(pair_sims)) if pair_sims else 0.0 # 5. Use MCI to negatively correct the original similarity score. # S_tilde = S - lambda_mci * MCI, the higher the MCI (the more the question resembles a standard answer), the lower the final score will be. S_tilde = S - lambda_mci *MCI # Returns a dictionary containing the original score, MCI, corrected score, and detailed similarity for each model. return { "S_raw": S, # Raw weighted similarity score without MCI correction "MCI": MCI, # Model Consistency Index (average of pairwise similarity between multiple model answers) "S_corrected": S_tilde, # Final score after negative correction "sims_per_model": sims, # A list of similarities for each model } Figure 4 The diagram shows a schematic representation of the verification mode and feedback loop described in the embodiments of this application. Figure 4 As shown, the verification mode and feedback loop process specifically includes the following steps: Step S401: Determine whether the corrected global suspicion score P' is in the medium suspicion interval.

[0087] Step S402: The detection AI automatically starts the verification mode and generates at least one verification follow-up question, a "trap" question, based on the candidate's resume and previous answers.

[0088] Step S403: Display at least one confirmatory follow-up question ("trap") to the interviewer through the interviewer's interface, and have the interviewer ask the candidate in the conversation.

[0089] Step S404: Based on the candidate's verification answers, obtain the reference answer text set of the major language models and the model consistency index MCI_ver between the reference answers of the major language models.

[0090] Step S405. Calculate and obtain the corrected similarity score for the current interview question based on the model consistency index between the reference answer text set and the reference answers of each large language model in each round of interview questions.

[0091] Step S406. Update the global suspicion score according to the corrected similarity score, obtain the corrected global suspicion score P'', and feedback the updated suspicion level to the interviewer.

[0092] In some embodiments, the present application provides a schematic example of simulation data for multiple rounds of questions. Figure 5 Shown is a schematic diagram of partial detection results of a certain simulated interview described in an embodiment of the present application. As Figure 5 shown, Round Q1 (project introduction question): S1 = 0.682, MCI1 = 0.60, S̃1 ≈ 0.38; Round Q2 (open question on technical details): S2 = 0.88, MCI2 = 0.45, S̃2 ≈ 0.70; Round Q3 (standard algorithm question): S3 = 0.91, MCI3 = 0.94, S̃3 ≈ 0.53; Verification question Q_ver (on-site fault handling combined with resume): S_ver = 0.93, MCI_ver = 0.40, S̃_ver≈ 0.77.

[0093] When the system calculates the global suspicion P, it uses the corrected scoring sequence {S̃1, S̃2, S̃3, S̃_ver}, combines behavioral characteristics such as answering duration and pauses, and outputs P' through a non-linear 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".

[0094] In some embodiments, the global suspicion score P’ is set between 0 and 1 (the higher, the more suspicious). The division method of the preset threshold interval is as follows: Normal interval: 0 ≤ P’<T1; Moderately suspicious interval: T1 ≤ P’<T2; Highly suspicious interval: T2 ≤ P’ ≤1; Among them, 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.

[0095] Figure 6 Shown is a schematic diagram of the detection results of the interviewer interface described in an embodiment of the present application. As Figure 6As shown, the speech acquisition and transcription module performs speech recognition on the audio and video streams of the interviewer and the candidate, obtaining the interview question text and the candidate's answer text arranged in chronological order. The AI ​​recognition control panel is used to provide real-time prompts to the interviewer regarding the suspicion level, the global suspicion score P', and the similarity score of recent corrections and verifications. The detection AI automatically initiates the verification mode to generate at least one verification follow-up "trap" question, which is displayed to the interviewer through the interviewer's interface and asked to the candidate. The global suspicion score P'' is updated based on the candidate's verification answer, and the updated suspicion level is fed back to the interviewer.

[0096] In one embodiment of this application, generating a verification probing trap question based on the detection AI activation verification mode and presenting it to the interviewer includes the following steps S3061 to S3064.

[0097] Step S3061: Identify and obtain suspicious source types based on the corrected similarity scores of each round of interview questions, the corresponding model consistency index, and behavioral features.

[0098] Step S3062: Select a target trap question template from the preset trap question template library based on the suspected source type, and generate a candidate trap question set by calling the large language model based on the candidate information; the candidate information includes at least resume information and answer text entity information.

[0099] In some embodiments, the preset trap question template library includes several trap question models, such as "requiring specific project details," "requiring quantifiable metrics," and "requiring a retelling from another perspective." The target trap question template is a trap question model selected based on the type of suspicious source and candidate information (e.g., resume information, answer text entity information). The answer text entity information includes, but is not limited to, data such as project name, company name, technology stack, time, and metrics.

[0100] Step S3063: Calculate the discrimination score of each candidate trap question based on the candidate trap question set and the candidate information.

[0101] Step S3064: Select the question with the highest discrimination score as the confirmatory follow-up question based on the discrimination score of each candidate trap question, and present it to the interviewer; the confirmatory follow-up question includes at least one candidate trap question.

[0102] In some embodiments, the verification follow-up questions in the verification mode of this application are automatically generated by the detection AI and displayed through the interviewer's interface, which the interviewer then selects and asks to the candidate.

[0103] In some embodiments, this application provides an example of triggering the verification mode and generating AI-generated trap questions. In this embodiment, the AI, as an independent service, receives transcribed text and behavioral characteristics during the interview process in real time. Without changing the original human-computer dialogue structure, a corrected score S̃i is calculated for each round of responses, and the current suspiciousness P' is calculated based on a sliding window or global statistics. When P' moves from the normal range to the moderately suspicious range, the verification question generation module is automatically triggered. First, the verification question generation module reads the candidate's resume summary and the content summary of the recent rounds of questions and answers.

[0104] Second, design one or more "trap-style" verification question templates based on the main sources of suspicion (e.g., certain questions are highly similar to LLM answers). For example, ask candidates to provide more detailed implementation steps or failure cases for the projects mentioned above.

[0105] Third, after checking the length and comprehensibility of the generated verification follow-up questions, the question text is pushed to the "AI Suggested Questions" area of ​​the interviewer's UI, where the interviewer can select one of the questions and ask the candidate in their own words or by reading it directly.

[0106] Fourth, after the candidate answers, the system calculates the similarity and MCI of the answer to the verification question with the example described above, obtaining a corrected score S̃_ver for the verification question, and then updates the global suspiciousness P'' according to a preset weight. Through this method, the system achieves a feedback loop of "AI monitoring + AI question generation + HR questioning + AI re-judgment," effectively improving the ability to recognize AI-assisted answers without depriving the interviewer of their control.

[0107] In some embodiments, the algorithm for automatically generating and quantitatively filtering confirmatory "trap" follow-up questions in this application is as follows: When the global suspicion level P' falls into the moderate suspicion range, the system enters automatic verification mode, and the detection AI automatically completes the following steps to generate and filter verification "trap" follow-up questions: First, suspicious source analysis is required: the detection AI identifies suspicious source types based on the corrected score S_tilde_i from multiple rounds of responses, the corresponding MCI_i, and behavioral characteristics, for example: Type A: The answer has a high similarity to the reference answer of the model and a high MCI, which is suspected to be "standard answer-style AI answer".

[0108] Type B: The answer is inconsistent with the project experience in the candidate's resume or is too vague.

[0109] Type C: Multi-turn responses exhibit a highly consistent and textbook-like language style, along with abnormal behavioral characteristics.

[0110] Next, candidate questions are generated: based on the type of suspicious source, the corresponding trap question template is selected (e.g., "requiring specific project details" or "requiring a retelling from another perspective"), and combined with entity information in the candidate's resume and dialogue history (e.g., project name, company name, technology stack, etc.), a set of candidate trap questions is generated by calling the large language model: Q = {q_1, q_2, …, q_K}.

[0111] Quantitative scoring metrics for candidate trap problems: For each candidate problem q_k, the following metrics are calculated, and the discrimination score D(q_k) of the problem is further calculated based on these metrics, which is used to screen the optimal trap problem.

[0112] Resume relevance score R_k: Extracts entity information set E_cv (including but not limited to company, project, position, etc.) from candidate resumes and historical answers, and entity set E(q_k) from candidate questions q_k. R_k is calculated as follows: R_k = |E(q_k)∩E_cv| / ( |E(q_k)| + epsilon ) Here, epsilon is a constant to prevent division by zero. The larger R_k is, the more closely the question matches the candidate's personal experience, and the less friendly it is to AI large language models.

[0113] Model consistency score M_k: Candidate questions q_k are input into multiple large language models to obtain a set of reference answers {u_k1, …, u_km}, which is vectorized into e(u_kj). The cosine similarity is calculated pairwise. The calculation of M_k is as follows: M_k = ( 2 / ( m * (m - 1) ) ) *Σ_{j1 <j2} cos_sim( e(u_kj1), e(u_kj2) ) The larger M_k is, the more consistent the answers of different models are to the question, and the closer they are to the "standard answer question". This is not conducive to distinguishing AI from real experience, so it is used as a negative factor in the overall score.

[0114] Suspicious content targeting score C_k: Let the set of suspicious rounds be I_susp, where the semantic vector of each round's response is h_i. The semantic vector g_k is calculated for candidate questions q_k, and C_k is calculated as follows: C_k=max_{i∈I_susp} cos_sim( g_k, h_i ) The larger C_k is, the more semantically "aligned" the question is with the previous highly suspicious answer.

[0115] Detail requirement score L_k (optional metric): Several "detail slots" are predefined in the problem template, such as whether specific time, quantifiable metrics, number of key steps, etc., are required. The number of activated detail slots s_k in q_k is counted. Let the maximum number of slots be S_max. L_k is calculated as follows: L_k = s_k / S_max The larger L_k is, the higher the requirements for the details of the answer, which is more conducive to distinguishing real experience from generalized answers.

[0116] Based on the above indicators, a discrimination score D(q_k) is defined for each candidate question q_k: D(q_k) = sigmoid(α* R_k +β* C_k +γ* L_k -δ* M_k ) Where: sigmoid(x) is the Sigmoid function or other monotonically increasing normalization function, used to map the linear combination within the parentheses to the interval [0, 1], so that the value of D(q_k) falls between 0 and 1; α, β, γ, δ>0 are configurable weights used to balance the influence of each component.

[0117] The system calculates D(q_k) for all candidate questions and selects the question with the highest discriminative power or several questions as the final confirmatory trap questions, for example: q* = argmax_{q_k∈Q}D(q_k) Alternatively, the system can select the top N questions that satisfy D(q_k)≥T_D and push them to the "AI Suggested Questions" area on the interviewer's end, allowing the interviewer to ask the candidates during the interview. After the candidate answers, the system recalculates the similarity and suspiciousness of the question to update the global suspiciousness P''.

[0118] Through the aforementioned quantitative indicators and comprehensive scoring, the verification trap questions in this application are not simply follow-up questions based on insufficient scores, but are specifically designed for the "suspected AI answering" mode. The ability to distinguish between different types of questions is quantitatively evaluated through indicators such as R_k, C_k, L_k, and M_k, thereby enhancing the recognition effect of external AI-assisted answers at the algorithm level.

[0119] The scope of protection of the interview-assisted answer detection method based on multi-model comparison described in this application is not limited to the execution order of the steps listed in this embodiment. Any solution implemented by adding, subtracting, or replacing steps in the prior art based on the principles of this application is included within the scope of protection of this application.

[0120] This application also provides an interview-assisted answer detection system based on multi-model comparison. The interview-assisted answer detection system based on multi-model comparison can implement the interview-assisted answer detection method based on multi-model comparison described in this application. However, the implementation device of the interview-assisted answer detection method based on multi-model comparison described in this application includes, but is not limited to, the structure of the interview-assisted answer detection system based on multi-model comparison listed in this embodiment. All structural modifications and substitutions of the prior art made according to the principles of this application are included within the protection scope of this application.

[0121] like Figure 7 As shown in the figure, this embodiment provides an interview-assisted answer detection system based on multi-model comparison. The system 700 includes: a speech acquisition and transcription module 701, a multi-model calling module 702, a similarity calculation module 703, a model consistency analysis module 704, a suspiciousness assessment module 705, a verification question generation module 706, and an interviewer user interface and prompt module 707.

[0122] The speech acquisition and transcription module 701 is configured to acquire audio and video streams between interviewers and candidates during remote interviews, and to 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.

[0123] The large model calling module 702 is configured to call at least two large language models based on the text of each round of interview questions to obtain a set of reference answer texts for each round of interview questions; the set of reference answer texts includes the reference answer texts of each of the large language models.

[0124] The similarity calculation module 703 is 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 to obtain the candidate's global suspiciousness score based on the model similarity score of each round of interview questions.

[0125] The model consistency analysis module 704 is configured to perform consistency calculations based on the reference answer text set of each round of interview questions, and obtain the model consistency index between the reference answers of major language models in each round of interview questions.

[0126] The suspicion assessment module 705 is configured to negatively correct the model similarity score and global suspicion score of the current interview question based on the consistency index between the reference answers of various language models in each round of interview questions, and obtain the corrected global suspicion score.

[0127] The verification question generation module 706 is configured to compare the corrected global suspicion score with a preset threshold range. If the corrected global suspicion score is within the preset suspicion range, a verification follow-up question trap is generated and displayed to the interviewer.

[0128] The interviewer's user interface and prompting module 707 are configured to obtain the candidate's verification answer based on the verification follow-up question trap, update the global suspicion score based on the candidate's verification answer, and feed back the updated global suspicion score to the interviewer.

[0129] It should be noted that the functions or operations of the speech acquisition and transcription module 701, the multi-model calling module 702, the similarity calculation module 703, the model consistency analysis module 704, the suspiciousness assessment module 705, the verification question generation module 706, and the interviewer user interface and prompt module 707 described in this embodiment correspond one-to-one with the steps in the interview auxiliary answer detection method based on multi-model comparison described above, so they will not be repeated here.

[0130] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, or methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules / units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or units may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or modules or units may be electrical, mechanical, or other forms.

[0131] The modules / units described as separate components may or may not be physically separate. The components shown as modules / units may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules / units can be selected to achieve the objectives of the embodiments of this application, depending on actual needs. For example, the functional modules / units in the various embodiments of this application may be integrated into one processing module, or each module / unit may exist physically separately, or two or more modules / units may be integrated into one module / unit.

[0132] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those 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 this application.

[0133] Figure 8 The diagram shown is a structural schematic of the electronic device 800 described in an embodiment of this application. Figure 8 As shown, in this embodiment, the electronic device 800 includes a memory 801 and a processor 802.

[0134] The memory 801 is used to store computer programs; preferably, the memory 801 includes various media that can store program code, such as ROM, RAM, magnetic disk, USB flash drive, memory card or optical disk.

[0135] Specifically, memory 801 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. Electronic device 800 may further include other removable / non-removable, volatile / non-volatile computer system storage media. Memory 801 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of this application. It is understood that memory 801 may be volatile memory or non-volatile memory, or both. Non-volatile memory may be read-only memory (ROM) or programmable read-only memory (PROM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM) and synchronous static random access memory (SSRAM). The memories described in the embodiments of this application are intended to include, but are not limited to, these and any other suitable categories of memories.

[0136] The processor 802 is connected to the memory 801 and is used to execute the computer program stored in the memory 801 so that the electronic device 800 executes the interview-assisted answer detection method based on multi-model comparison as described in any embodiment of this application.

[0137] Optionally, the processor 802 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0138] Optionally, in this embodiment, the electronic device 800 may further include a display 803. The display 803 is communicatively connected to the memory 801 and the processor 802, and is used to display the relevant graphical user interface (GUI) of the interview-assisted answer detection method based on multi-model comparison described in this application embodiment.

[0139] This application also provides a computer-readable storage medium. Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing a processor. The program can be stored in a computer-readable storage medium, which is a non-transitory medium, such as random access memory, read-only memory, flash memory, hard disk, solid-state drive, magnetic tape, floppy disk, optical disk, and any combination thereof. The storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video disc (DVD)), or a semiconductor medium (e.g., solid-state drive (SSD)).

[0140] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0141] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for detecting interview-assisted responses based on multi-model comparison, characterized in that, include: The system collects audio and video streams of interviewers and candidates during remote interviews, and performs speech recognition on the audio and video streams to obtain multi-round question-and-answer transcribed text. The transcribed question-and-answer text includes the interview questions and answers. Based on the text of each round of interview questions, at least two large language models are invoked to obtain a set of reference answer texts for each round of interview questions; the set of reference answer texts includes the reference answer texts of each of the large language models; The model similarity score of the corresponding interview question is calculated based on each answer text and the corresponding reference answer text set, and the global suspicion score of the candidate is obtained based on the model similarity score of the interview questions in each round. Consistency calculations are performed based on the reference answer text sets of interview questions in each round to obtain the model consistency index among the reference answers of major language models in each round of interview questions. Based on the consistency index among the reference answers of various language models in each round of interview questions, the model similarity score and global suspicion score of the current interview question are negatively corrected to obtain the corrected global suspicion score; Based on the comparison between the corrected global suspicion score and the preset threshold range, if the corrected global suspicion score is within the preset suspicion range, a verification follow-up question trap is generated and shown to the interviewer. Based on the aforementioned confirmatory follow-up questions, the candidate's verification answers are obtained, and the global suspicion score is updated according to the candidate's verification answers. The updated global suspicion score is then fed back to the interviewer.

2. The interview-assisted answer detection method based on multi-model comparison according to claim 1, characterized in that, The model similarity score for each interview question is calculated based on the text of each answer and the corresponding set of reference answers. Then, the candidate's global suspiciousness score is obtained based on the model similarity scores of each round of interview questions, including: Semantic similarity is calculated based on each answer text and its corresponding reference answer text set to obtain the similarity value between each answer text and its corresponding reference answer text; The model similarity score for the corresponding interview question is calculated based on the similarity value between each answer text and the corresponding reference answer text. The global suspiciousness score of the candidate is obtained by summarizing the model similarity scores based on the interview questions in each round.

3. The interview-assisted response detection method based on multi-model comparison according to claim 2, characterized in that, Semantic similarity is calculated based on each answer text and its corresponding reference answer text set to obtain the similarity value between each answer text and its corresponding reference answer text, including: Based on the semantic vectorization model, each answer text and the corresponding reference answer text set are mapped to a vector space to obtain the corresponding candidate answer text vector and reference answer text vector; Semantic similarity is calculated based on the candidate's 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. sij=(e(ai)*e(rij)) / (||e(ai)||*||e(rij)||) 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.

4. The interview-assisted response detection method based on multi-model comparison according to claim 2, characterized in that, The model similarity score for the interview questions is obtained by weighted averaging of the weights of various 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 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.

5. The interview-assisted response detection method based on multi-model comparison according to claim 2, characterized in that, The candidate's global suspiciousness score is calculated using a non-linear function based on the model similarity scores of the interview questions in each round, including: P =σ(α·S̄ +β·Var(S) +γ·F behavior), 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.

6. The interview-assisted answer detection method based on multi-model comparison according to claim 1, characterized in that, 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; Based on the consistency index among the reference answers of various language models in each round of interview questions, the model similarity score and global suspicion score of the current interview question are negatively corrected. The corrected global suspicion score includes: Based on the consistency index among the reference answers of various language models in each round of interview questions, the model similarity score of the current interview question is negatively corrected to obtain the corrected model similarity score of each interview question; The corrected global suspicion score is obtained by summarizing the model similarity scores for each interview question.

7. The interview-assisted response detection method based on multi-model comparison according to claim 6, characterized in that, The calculation methods for the corrected model similarity score include: S̃i = Si-λ·MCIi, or S̃i = Si·(1 -λ·MCIi), Where λ represents the preset or trained correction coefficient, Si represents the model similarity score of the i-th round of interview questions before correction, S̃i represents the model similarity score of the i-th round of interview questions after correction, and MCIi represents the model consistency index of the i-th round of interview questions.

8. The interview-assisted response detection method based on multi-model comparison according to claim 1, characterized in that, Based on the detection AI-driven verification mode, confirmatory follow-up questions are generated and presented to the interviewer, including: Based on the corrected similarity scores of each round of interview questions, the corresponding model consistency index, and behavioral characteristics, suspicious source types are identified. Based on the suspected source type, a target trap question template is selected from a preset trap question template library, and a candidate trap question set is generated by calling a large language model based on the candidate information; the candidate information includes at least resume information and answer text entity information; Calculate the discrimination score for each candidate trap question based on the candidate trap question set and the candidate information; Based on the discrimination scores of each candidate trap question, the question with the highest discrimination score is selected as the confirmatory follow-up trap question and presented to the interviewer; the confirmatory follow-up trap question includes at least one candidate trap question.

9. A multi-model comparison-based interview response assistance detection system, characterized in that, include: The speech acquisition and transcription module is configured to acquire audio and video streams between interviewers and candidates during remote interviews, and to perform speech recognition on the audio and video streams to obtain multi-round question-and-answer transcribed text. The transcribed question-and-answer text includes the interview questions and answers. The multi-model invocation module is configured to invoke at least two large language models based on the text of each round of interview questions to obtain a reference answer text set for each round of interview questions; the reference answer text set includes the reference answer text of each of the large language models; The similarity calculation module is configured to calculate the model similarity score of the corresponding interview questions based on each answer text and the corresponding reference answer text set, and to obtain the candidate's global suspiciousness score based on the model similarity score of each round of interview questions; The model consistency analysis module is configured to perform consistency calculations based on the reference answer text set of each round of interview questions, and obtain the model consistency index between the reference answers of major language models in each round of interview questions. The suspicion assessment module is configured to negatively correct the model similarity score and global suspicion score of the current interview question based on the consistency index between the reference answers of various language models in each round of interview questions, and obtain the corrected global suspicion score; The verification question generation module is configured to compare the corrected global suspicion score with a preset threshold range. If the corrected global suspicion score is within the preset suspicion range, a verification follow-up question trap is generated and displayed to the interviewer. The interviewer's user interface and prompt module are configured to obtain the candidate's verification answer based on the verification follow-up question trap, update the global suspicion score based on the candidate's verification answer, and feed back the updated global suspicion score to the interviewer.

10. An electronic device, characterized in that, include: A memory that stores a computer program; The processor, which is communicatively connected to the memory, executes the interview-assisted answer detection method based on multi-model comparison as described in any one of claims 1 to 8 when calling the computer program.

Images