An interview content analysis method and system based on multi-agent cooperation

Abstract

The application is suitable for the technical field of artificial intelligence interview assistance, and provides an interview content analysis method and system based on multi-agent cooperation, which comprises the following steps: when the semantic deviation index of the target interviewee before and after the program is greater than the individualized deviation allowable threshold, a time opportunity correction factor is generated according to the exceeding amplitude, and the prompt generation starting length of the pre-generated agent of the streaming interview prompt is corrected according to the time opportunity correction factor. When the expression mode of the interviewee is different from the receiving logic of the interviewee, the application can avoid that the agent generates prompt information deviating from the final semantic direction based on the pre-program expression too early, and can also avoid that the prompt generation is delayed without difference, so that the real-time performance and the effective utilization rate are taken into account.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence interview assistance technology, and in particular relates to a method and system for analyzing interview content based on multi-agent collaboration. Background Technology

[0002] With the application of artificial intelligence technology in interview assistance scenarios, existing systems can typically invoke intelligent agents to perform real-time analysis based on streaming text recognition during the interviewer's questioning process, and generate prompts to assist the interviewee in responding. To ensure that the prompts are delivered before the interviewee's reply, existing systems generally do not wait for the entire interview question to finish before starting generation. Instead, they control the streaming interview prompt pre-generation intelligent agent to execute formal prompt generation when the identified content of the interview question reaches the default prompt generation start length.

[0003] However, in high-level interview scenarios such as management talent interviews, interviewers often use lengthy formats to ask follow-up questions. For the same type of interview question, different interviewers have different expression habits; some interviewers' expressions are relatively coherent, while others tend to first provide background information before concluding or changing the direction of the actual question. Correspondingly, different interviewees also have varying abilities to receive, understand, and summarize different interviewers' expression styles; that is, there is a degree of compatibility between the interviewer's expression style and the interviewee's receiving logic.

[0004] Existing interview assistance systems typically use a fixed default prompt generation start length, failing to consider the relationship between the semantic deviation between the target interviewer's initial and subsequent responses and the interviewee's ability to adapt to the prompts. When the initial and subsequent responses to a long question deviate significantly from each other, exceeding the interviewee's tolerance, the agent may prematurely generate prompts based on the initial responses. This results in the prompts being inconsistent with the final semantic direction of the complete interview question, thereby reducing the interviewee's effective adoption rate of the prompts. Therefore, there is an urgent need for an interview content analysis method that can individually adjust the prompt generation start length of the streaming interview prompt pre-generation agent by combining the characteristics of the interviewer's long question expression deviation and the interviewee's historical adoption. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for analyzing interview content based on multi-agent collaboration, aiming to solve the problems mentioned in the background art.

[0006] This invention is implemented as follows: a method for analyzing interview content based on multi-agent collaboration, the method comprising:

[0007] Real-time identification of questions asked by target visitors to respondents, and determination of the specified question type when the question is a preset long question;

[0008] From the historical interview data of the respondents, several first samples with the same question type as the specified questions are obtained, and the semantic deviation index of the front and back of each first sample and the effective adoption rate of the prompt information generated by the pre-generated intelligent agent of the respondent's live interview prompt are calculated.

[0009] Establish the correlation between the front-end and back-end semantic deviation index and the effective adoption rate. When the effective adoption rate first stabilizes and then decreases as the front-end and back-end semantic deviation index increases, the index corresponding to the trend inflection point is determined as the individualized deviation tolerance threshold for the respondents.

[0010] Obtain a second sample corresponding to the specified question type from the target visitor's historical visit data, and calculate the target visitor's forward and backward semantic deviation index based on the second sample;

[0011] When the semantic deviation index of the target visitor before and after the visit exceeds the individualized deviation tolerance threshold, a timing correction factor is generated based on the excess magnitude, and the prompt generation start length of the pre-generated agent for streaming interview prompts is corrected accordingly.

[0012] As a further limitation of the technical solution of the present invention, the preset long questions are determined in the following way: before the start of this interview, a set of preset question types corresponding to this interview is obtained, and the types of long questions that need to be put in long expression form are determined from the set of preset question types to form a set of preset long questions;

[0013] When real-time streaming identification of access questions raised by target visitors, the candidate question type corresponding to the access question is determined based on the streaming identification text of the access question within a preset start time period; when the candidate question type belongs to a preset long question set, the access question is determined to belong to the preset long question set, and the candidate question type is determined as the specified question type.

[0014] As a further limitation of the technical solution of the present invention, the acquisition of the first sample and the second sample also includes background screening; the background includes at least one of the following: interview topic, interview stage, interview objective, and related event type.

[0015] As a further limitation of the technical solution of this invention embodiment, the calculation process of the pre- and post-process semantic deviation index includes: obtaining the streaming recognition text of the access question in the sample and the default prompt generation start length of the streaming interview prompt pre-generated agent; taking the text formed after the current access question starts and the recognized content reaches the default prompt generation start length as the pre-process expression segment, and taking the text formed after the recognized content exceeds the default prompt generation start length until the end of the access question as the post-process expression segment; determining the semantic vectors of the pre-process expression segment and the post-process expression segment respectively, and determining the degree of semantic deviation based on the vector distance between the two semantic vectors; determining the expression length parameter according to the text length of the access question and / or the duration of the speech; and calculating the pre- and post-process semantic deviation index based on the degree of semantic deviation and the expression length parameter.

[0016] As a further limitation of the technical solution of this embodiment of the invention, the calculation process of the effective adoption rate includes: obtaining the prompt information generated by the streaming interview prompt pre-generation agent in the first sample, and the response text formed by the interviewee within the preset response time window after the prompt information is generated; splitting the prompt information into key points to obtain several prompt points, and determining the prompt semantic vector of each prompt point; performing semantic parsing on the response text to determine the response semantic segments in the response text corresponding to each prompt point; when the semantic similarity between the prompt point and the corresponding response semantic segment reaches the preset adoption threshold, and the corresponding response semantic segment matches the final semantic direction of the interview question, it is determined that the prompt point is effectively adopted; and calculating the effective adoption rate of the prompt information based on the ratio of the number of effectively adopted prompt points to the total number of prompt points.

[0017] As a further limitation of the technical solution of this invention, the step of establishing the correlation between the front-to-back semantic deviation index and the effective adoption rate, and determining the index corresponding to the trend inflection point as the individualized deviation tolerance threshold for the respondents when the effective adoption rate first stabilizes and then decreases as the front-to-back semantic deviation index increases, includes:

[0018] Several first samples are sorted from smallest to largest according to the semantic deviation index between the beginning and end of the process to obtain the first sample sequence. The effective adoption rate corresponding to each first sample is extracted according to the first sample sequence to form the effective adoption rate change sequence. The correlation between the semantic deviation index between the beginning and end of the process and the effective adoption rate is established based on the effective adoption rate change sequence.

[0019] Perform trend detection on the effective adoption rate change sequence to determine whether the effective adoption rate shows a trend of first stabilizing and then decreasing as the semantic deviation exponent increases in the preceding and following stages.

[0020] If so, the semantic deviation index corresponding to the transition position from a stable state to a declining state will be determined as the individualized deviation tolerance threshold for the respondents.

[0021] If not, then individualized deviations from the allowable threshold will not be generated, and the default prompts for generating the startup length will be generated using streaming interviews to prompt the pre-generated agent.

[0022] As a further limitation of the technical solution of this invention, when the semantic deviation index of the target visitor before and after the event exceeds the individualized deviation tolerance threshold, the step of generating a timing correction factor based on the excess magnitude and correcting the prompt generation start length of the streaming interview prompt pre-generated agent accordingly includes:

[0023] Determine whether the semantic deviation index of the target visitor before and after the visit is greater than the individualized deviation tolerance threshold.

[0024] If not, when the identified content of the current access question reaches the default prompt generation start length, the control stream interview prompt pre-generation agent will execute the formal prompt generation;

[0025] If so, then a timing correction factor is generated based on the extent to which the target visitor's semantic deviation index exceeds the individualized deviation tolerance threshold.

[0026] The default prompt generation startup length is increased and corrected based on the timing correction factor to obtain the corrected prompt generation startup length;

[0027] When the identified content of the current access question reaches the length required to generate a correction prompt, the control flow interview prompt pre-generation agent performs formal prompt generation based on the identified content and generates interview auxiliary prompts.

[0028] As a further limitation of the technical solution of this embodiment of the invention, when increasing the startup length of the default prompt generation, a preset correction function is used, and the correction function is:

[0029] ;

[0030] in, This refers to shifting the startup length prompt. This refers to the default startup length prompt. This refers to the maximum possible value for the startup length that can be generated. This refers to the semantic deviation index between the target visitor's preceding and following journey. This refers to individual deviations from the permissible threshold. This refers to exceeding the allowable range. This refers to a preset correction strength coefficient, and satisfies... Greater than 0.

[0031] As a further limitation of the technical solution of the present invention, the prompt generation start length is at least one of the following: the length of the identified text, the duration of the identified speech, and the number of identified semantic units after the current access question begins; wherein, the default prompt generation start length, the corrected prompt generation start length, and the upper limit of the prompt generation start length adopt the same measurement type.

[0032] A multi-agent collaborative interview content analysis system, the system comprising:

[0033] The question identification module is used to identify the questions raised by the target visitor to the interviewee in real time, and to determine the specified question type when the question is a preset long question;

[0034] The sample analysis module is used to obtain several first samples with the same question type as the specified questions from the historical interview data of the respondents, and to calculate the semantic deviation index of the preceding and following stages of each first sample and the effective adoption rate of the prompt information generated by the pre-generated intelligent agent of the respondent's live interview prompts.

[0035] The threshold determination module is used to establish the correlation between the front-end and back-end semantic deviation index and the effective adoption rate. When the effective adoption rate first stabilizes and then decreases as the front-end and back-end semantic deviation index increases, the index corresponding to the trend inflection point is determined as the individualized deviation tolerance threshold for the respondents.

[0036] The target index calculation module is used to obtain a second sample corresponding to a specified question type from the historical access data of the target visitor, and calculate the semantic deviation index of the target visitor before and after the visit based on the second sample.

[0037] The startup length correction module is used to generate a timing correction factor based on the magnitude of the deviation when the semantic deviation index of the target visitor exceeds the individualized deviation tolerance threshold, and to correct the startup length of the prompt generation of the pre-generated agent for the streaming interview prompt accordingly.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] This invention addresses the differences in expression and reception adaptation between different visitors and respondents. It analyzes respondents' historical interview data to assess their adaptability to varying degrees of semantic deviation before and after the initial stages of similar long questions. An individualized deviation tolerance threshold is determined based on the turning point where the effective adoption rate shifts from stable to declining. Furthermore, the startup length of the prompt generation agent in the streaming interview is adjusted according to whether the target visitor's semantic deviation index before and after the initial stages exceeds this threshold. This prevents the agent from prematurely generating prompts that deviate from the final semantic direction based on the initial expression when the visitor's expression style and the respondent's reception logic are out of sync. It also avoids indiscriminately delaying prompt generation, balancing real-time performance and effective adoption rate. Attached Figure Description

[0040] Figure 1 A flowchart of the method provided in the embodiments of the present invention;

[0041] Figure 2 This is a flowchart illustrating the process of determining the individualized deviation tolerance threshold for respondents in the method provided in this embodiment of the invention;

[0042] Figure 3 This is a flowchart illustrating the process of correcting the startup length of the prompt generation in the method provided by this embodiment of the invention;

[0043] Figure 4 The application architecture diagram of the system provided in the embodiments of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0045] Figure 1 A flowchart of the method provided by an embodiment of the present invention is shown.

[0046] Specifically, a method for analyzing interview content based on multi-agent collaboration includes the following steps:

[0047] Step S100: Real-time identification of the interview questions posed by the target interviewer to the interviewee; when the interview question is a preset long question, a specified question type is determined. The multi-agent system in this invention may include a question identification agent, a sample analysis agent, a semantic deviation calculation agent, an effective adoption rate evaluation agent, a threshold determination agent, and a streaming interview prompt pre-generation agent. Each agent performs tasks such as question type identification, sample screening, index calculation, adoption rate evaluation, threshold determination, and prompt generation, and collaboratively completes interview content analysis through the system scheduling module.

[0048] The preset long questions are determined as follows: Before the start of the interview, a set of preset question types corresponding to the interview is obtained, and the types of long questions that need to be asked in a long expression form are determined from the preset question type set to form a preset long question set; when the interview questions raised by the target visitor are identified in real time, the candidate question types corresponding to the interview questions are determined based on the streaming identification text of the interview questions within a preset start time period; when the candidate question type belongs to the preset long question set, the interview questions are determined to belong to the preset long questions, and the candidate question type is determined as the specified question type.

[0049] In this embodiment of the invention, the target interviewer is the party who poses interview questions to the interviewee during the interview, and the interviewee is the party who receives interview assistance prompts and responds based on the interview questions. During the interview, the system collects and streams the target interviewer's voice data in real time to obtain streamed recognized text that is continuously updated with the interview questions, and performs real-time analysis on the interview questions posed by the target interviewer based on the streamed recognized text.

[0050] In high-level interviews, management talent interviews, or other interview scenarios requiring immediate responses, an AI agent is typically integrated to assist interviewees in understanding the interview questions and organizing their responses. Specifically, existing systems can perform real-time parsing of the collected audio stream and generated streaming text during the interviewer's questioning process. Once the interview has progressed to a certain length, the streaming interview prompt pre-generation agent initiates the formal prompt generation process. After activation, the streaming interview prompt pre-generation agent combines the currently recognized interview question content, interview context, interviewee's historical interview data, interview topic, pre-set knowledge base, and other relevant data to generate interview assistance prompts to aid the interviewee's responses. These prompts are then provided to the interviewee via headphones, a display screen, a prompting terminal, or other output devices.

[0051] The reason for not waiting for the target interviewer to complete the entire interview question before performing overall analysis and prompt generation is that in real-time interview scenarios, there is usually a time-consuming process between the end of the interview question and the delivery of interview assistance prompts to the interviewee, including speech recognition, semantic parsing, context matching, agent reasoning, prompt generation, and front-end output. If the system only starts generating formal prompts after the interview question has completely ended, the interviewee often needs to answer immediately or has already started responding, causing the interview assistance prompts to be unable to be delivered within the effective time window, thus reducing the assistive role of the prompt information in the interviewee's immediate response. Especially in management talent interviews, interview questions are usually probing, continuous, and immediate, and interviewees need to respond within a short pause. Generating prompt information too late will cause the prompt content to lose its real-time assistive value. Therefore, existing systems usually start generating formal prompts for the streaming interview prompt pre-generation agent in advance, based on nodes where the identified content has reached a certain length, before the interview question has completely ended.

[0052] The node mentioned above, "starting formal prompt generation after the interview questions continue for a certain length," corresponds in this invention to the prompt generation start length of the streaming interview prompt pre-generation agent. The prompt generation start length can be understood as: after the current interview question begins, the system needs to accumulate the recognized text length, speech duration, or number of semantic units to reach the corresponding conditions before controlling the streaming interview prompt pre-generation agent to execute formal prompt generation. The setting and application of the prompt generation start length can be implemented using existing technologies. For example, existing systems can pre-set a default prompt generation start length based on the average length of the interview questions, the stability of speech recognition, the interviewee's required reaction time, the average generation time of the agent, and the output time of the prompt terminal; or, existing systems can use the corresponding recognized text length as the default prompt generation start length when they detect that the current interview question already contains a subject, predicate, core object, and initial question intent; or, existing systems can use historical interview data to statistically analyze how many words, how many seconds of speech duration, or how many semantic units are needed for similar interview questions to stably identify the question type, and use this statistical value as the default prompt generation start length.

[0053] However, researchers skilled in the art, through studying extensive access and response data, have discovered that different visitors exhibit different expression habits when asking long questions of the same type. Some visitors clearly indicate the final semantic direction of the question in the initial part of their long question, with the subsequent parts primarily serving as supplementary explanations; while others only provide background information, restate previous events, or make broad introductions in the initial part, only concluding with the true direction of the inquiry in the subsequent parts. Therefore, for the same specific question type, there may be varying degrees of semantic deviation between the initial and subsequent parts of the questions asked by different visitors.

[0054] For short questions or questions that can be clearly stated in a single sentence, there is usually no significant semantic deviation between the preceding and following segments of the expression. Alternatively, such questions may not yet have a long expression structure that can stably divide the preceding and following segments. Therefore, the semantic deviation index between the preceding and following segments described in this invention is not suitable for judgment. Furthermore, for excessively short access questions, the prompt generation start length of the streaming interview prompt pre-generated agent does not have significant adjustment value, as the system can usually complete recognition and prompt generation shortly after the question ends. Therefore, this invention primarily targets pre-set long questions, i.e., access questions that require the interviewer to pose a longer expression and that exhibit semantic deviation features between the preceding and following segments during streaming recognition.

[0055] Furthermore, even when the questions belong to the same specified question type, there are differences in expression reception adaptation between respondents and different interviewers. These differences refer to the respondent's understanding and adaptation to the interviewer's long question expression habits, semantic progression, and the degree of transition between different parts of the question. For respondents whose expression style is more compatible with the interviewer's, even if there is some semantic deviation in the preceding and following parts of the long question, the respondent can infer the final semantic direction of the question earlier based on the interviewer's expression habits. This makes it easier to convert the prompts generated by the streaming interview prompt pre-generation agent under the default prompt generation start length into usable response content.

[0056] Conversely, for respondents whose expression style is less adapted to the interviewer's, when there is a significant semantic deviation between the preceding and following parts of the interview question, it is difficult for them to infer the final semantic direction of the interview question based on the preceding parts. If the streaming interview prompt pre-generation agent still executes the formal prompt generation too early according to the default prompt generation start length, the identified content used to generate the prompt information may mainly come from the preceding parts of the expression, resulting in a deviation between the prompt information and the final semantic direction of the complete interview question. In this case, the respondent needs to filter, correct, or even discard the prompt information in subsequent responses, thus reducing the effective adoption rate of the prompt information.

[0057] Therefore, for a specific respondent, an initial increase in the front-to-back semantic deviation index does not necessarily lead to a decrease in effective adoption rate. Instead, there exists a permissible range related to their ability to understand and adapt to the respondent's expression. Within this permissible range, respondents can still effectively adopt the prompts based on their own understanding and adaptation of the respondent's expression. When the target respondent's front-to-back semantic deviation index exceeds this permissible range, the prompts generated at the default prompt generation start length are more likely to be inconsistent with the final semantic direction of the complete response question, thus causing the effective adoption rate to shift from a stable state to a declining state.

[0058] Based on the above understanding, this invention does not simply delay the generation of interview assistance prompts for all long questions. Instead, it only increases the prompt generation startup length of the streaming interview prompt pre-generation agent when the semantic deviation index of the current interviewer exceeds the interviewee's tolerance range. By appropriately increasing the prompt generation startup length, the streaming interview prompt pre-generation agent can collect more subsequent expressions in the current interview question before starting formal prompt generation, thereby improving the consistency between the semantic content on which it generates interview assistance prompts and the final semantic direction of the complete interview question. This avoids both premature startup causing the prompt information to deviate from the actual follow-up question and excessively late startup causing the interviewee to wait too long.

[0059] In step S100, the system first identifies the interview questions posed by the target interviewer to the interviewee in real time, and determines the designated question type when the interview questions belong to the preset long questions. Preset long questions are not simply determined retrospectively based on the final word count of the interview questions. Instead, before the interview begins, the system obtains a set of preset question types corresponding to the current interview based on the interview task, interview outline, interview topic, or interview template, and then filters out the question types that need to be presented in a long expression format from this set, forming a preset long question set. This preset long question set may include long questions related to project debriefing, organizational management, strategic explanation, risk management, performance attribution, and resource coordination, and can also be configured according to the actual interview business scenario.

[0060] When performing real-time streaming recognition of access questions posed by target visitors, the system can determine candidate question types based on the streaming text of the access question within a preset start time period. The preset start time period can be a preset time interval after the current access question begins, or it can be a segment after the current access question begins but before reaching a preset text length or a preset number of semantic units. The system can perform keyword recognition, semantic vector matching, intent classification, question type classification, or semantic judgment based on a large language model on the streaming text within this preset start time period to determine the candidate question type corresponding to the access question. When a candidate question type belongs to a preset set of long questions, the system determines that the access question belongs to the preset long question set and identifies the candidate question type as the specified question type.

[0061] The aforementioned real-time identification of access questions, determination of candidate question types based on streaming text within a preset start time period, and determination of whether candidate question types belong to a preset long question set can be implemented using mature speech recognition and natural language processing technologies. For example, the system can use a streaming automatic speech recognition model to convert the target visitor's speech into streaming text in real time, determine that the current speech belongs to the target visitor's question content through endpoint detection or speaker separation, determine candidate question types through a text classification model, semantic vector retrieval model, or intent recognition model, and determine whether the current access question belongs to the preset long question set by matching the candidate question types with the preset long question set. All of the above technical means can be implemented by existing streaming speech recognition, semantic classification, and question intent recognition technologies. The improvement focus of this invention is not on how to identify text from speech or how to perform general question classification, but on, after determining that the access question belongs to the preset long question set and obtaining the specified question type, further combining the interviewee's historical interview data and the target visitor's historical access data to specifically modify the prompt generation start length of the streaming interview prompt pre-generated intelligent agent.

[0062] Furthermore, the interview content analysis method based on multi-agent collaboration also includes the following steps:

[0063] Step S200: Obtain several first samples with the same question type from the respondents' historical interview data, and calculate the semantic deviation index between the preceding and following stages and the effective adoption rate of the respondents' feedback information generated by the pre-generated intelligent agent in the live interview prompts for each first sample.

[0064] The acquisition of the first sample also includes screening the interview background; the interview background includes at least one of the following: interview topic, interview stage, interview objective, and related event type.

[0065] The calculation process of the pre- and post-interview semantic deviation index includes: obtaining the streaming recognition text of the interview questions in the sample and the default prompt generation start length of the streaming interview prompt pre-generated agent; taking the text formed after the current interview question starts and the recognized content reaches the default prompt generation start length as the pre-interview expression segment, and taking the text formed after the recognized content exceeds the default prompt generation start length until the end of the interview question as the post-interview expression segment; determining the semantic vectors of the pre-interview expression segment and the post-interview expression segment respectively, and determining the degree of semantic deviation based on the vector distance between the two semantic vectors; determining the expression length parameter according to the text length of the interview question and / or the duration of the speech; and calculating the pre- and post-interview semantic deviation index based on the degree of semantic deviation and the expression length parameter.

[0066] The calculation process for the effective adoption rate includes: acquiring the prompt information generated by the pre-generated intelligent agent in the first sample of streaming interview prompts, and the response text formed by the interviewee within the preset response time window after the prompt information is generated; breaking down the prompt information into key points to obtain several key points, and determining the prompt semantic vector of each key point; performing semantic parsing on the response text to determine the response semantic segments in the response text corresponding to each key point; determining that the key point is effectively adopted when the semantic similarity between the key point and the corresponding response semantic segment reaches the preset adoption threshold, and the corresponding response semantic segment matches the final semantic direction of the interview question; and calculating the effective adoption rate of the prompt information based on the ratio of the number of effectively adopted key points to the total number of key points.

[0067] In this embodiment of the invention, the first sample is used to analyze how respondents adopted the agent's prompts when faced with similar long questions in historical interviews. Since different question types correspond to different response strategies, the system first selects samples from the respondents' historical interview data that are the same as the specified question type as the first sample.

[0068] Furthermore, even with the same question type, the interviewee's focus may differ if the interview topic, stage, objective, or related event type varies. For example, with project debriefing questions, mid-project interviews focus more on progress, while post-project interviews focus more on summarizing causes and analyzing responsibility. Therefore, this invention also filters the first sample based on the interview background, ensuring the first sample has a closer interview context to the current interview questions, reducing the influence of other factors on subsequent analysis results.

[0069] For each first sample, the system calculates a semantic deviation index between the preceding and following segments. This index represents the degree of semantic deviation between the preceding and following segments when the interviewer asks a long question. Specifically, the system acquires the streaming recognized text of the interview questions in the sample, as well as the default prompt generation start length of the streaming interview prompt pre-generated agent. After the current interview question begins, the text formed when the recognized content reaches the default prompt generation start length is taken as the preceding segment, and the text formed after exceeding the default prompt generation start length until the end of the interview question is taken as the following segment.

[0070] The system converts the preceding and following expression fragments into semantic vectors respectively, and calculates the vector distance between the two semantic vectors to obtain the degree of semantic deviation. This processing can be implemented using existing text vectorization models, sentence vector models, text embedding models, or large language model encoding methods. The larger the distance between the two semantic vectors, the more significant the semantic difference between the preceding and following expressions.

[0071] Simultaneously, the system determines the expression length parameter based on the text length and / or speech duration of the access question. The expression length parameter reflects the overall length of the access question's expression. Generally, when semantic deviations are similar, the longer the access question, the higher the likelihood of comprehension discrepancies between the preceding expression fragment and the complete question. Therefore, the system calculates a preceding-following semantic deviation index based on the degree of semantic deviation and the expression length parameter.

[0072] The significance of calculating the front-to-back semantic deviation index and the effective adoption rate lies in the following: the front-to-back semantic deviation index reflects whether there are significant jumps in the expression before and after a visitor asks a long question; the effective adoption rate reflects whether the respondent can effectively use the prompts generated by the agent under this expression mode. By obtaining these two parameters simultaneously in several first samples, the system can further determine whether the respondent's effective adoption rate of the prompts will decrease as the front-to-back semantic deviation index gradually increases, thus providing a basis for subsequently determining the individualized deviation tolerance threshold.

[0073] Furthermore, the interview content analysis method based on multi-agent collaboration also includes the following steps:

[0074] Step S300: Establish the correlation between the front-end and back-end semantic deviation index and the effective adoption rate. When the effective adoption rate first stabilizes and then decreases as the front-end and back-end semantic deviation index increases, the index corresponding to the trend inflection point is determined as the individualized deviation tolerance threshold for the respondents.

[0075] Specifically, Figure 2 A flowchart is shown to determine the individualized deviation tolerance threshold for respondents.

[0076] The process of establishing the correlation between the front-to-back semantic deviation index and the effective adoption rate, and determining the index corresponding to the trend inflection point as the individualized deviation tolerance threshold for respondents when the effective adoption rate first stabilizes and then declines with the increase of the front-to-back semantic deviation index, specifically includes the following steps:

[0077] Step S301: Sort several first samples in ascending order of semantic deviation index before and after the process to obtain a first sample sequence. Extract the effective adoption rate corresponding to each first sample according to the first sample sequence to form an effective adoption rate change sequence. Establish the correlation between the semantic deviation index before and after the process and the effective adoption rate based on the effective adoption rate change sequence.

[0078] Step S302: Perform trend detection on the effective adoption rate change sequence to determine whether the effective adoption rate shows a trend of first stabilizing and then decreasing as the semantic deviation exponent increases in the preceding and following stages.

[0079] Step S303: If so, the semantic deviation index of the transition from the stable state to the declining state is determined as the individualized deviation tolerance threshold of the respondent.

[0080] Step S304: If not, do not generate individualized deviation tolerance thresholds, and use streaming interview prompts to generate the default prompts for the pre-generated agent to generate the start-up length.

[0081] In this embodiment of the invention, step S300 is used to determine whether there is a pattern of change in the adoption of prompts related to the semantic deviation index between the beginning and end of a long question of a specified question type. That is, this step is not simply to calculate a statistical value, but to determine whether the respondent can still stably adopt the prompts generated by the pre-generated intelligent agent of the streaming interview prompt as the degree of jump in the respondent's expression of the long question gradually increases; if the effective adoption rate begins to decline after a certain position, it indicates that the respondent has an individualized tolerance range for this type of expression deviation.

[0082] In step S301, the system sorts several first samples from smallest to largest according to the semantic deviation index between the preceding and following stages, obtaining a first sample sequence. The purpose of this processing is to arrange the samples according to the degree of deviation in the visitor's long question expression from lowest to highest, making it easier to observe the change in effective adoption rate as the semantic deviation index between the preceding and following stages increases. Subsequently, the system extracts the effective adoption rate corresponding to each first sample according to the first sample sequence, forming an effective adoption rate change sequence. This process can be implemented using conventional data sorting, sequence construction, and statistical analysis methods.

[0083] In step S302, the system performs trend detection on the effective adoption rate change sequence to determine whether the effective adoption rate shows a trend of first stabilizing and then decreasing as the semantic deviation index increases. Specifically, existing data analysis methods such as sliding window mean, piecewise linear fitting, inflection point detection, or rate of change detection can be used. For example, the system can first determine whether the effective adoption rate fluctuation in the previous sample is within a preset stable range, and then determine whether the effective adoption rate in the subsequent sample shows a decreasing trend; when both conditions are met, the effective adoption rate change sequence is considered to conform to a trend of first stabilizing and then decreasing.

[0084] In step S303, if the aforementioned trend is detected, the system designates the point where the effective adoption rate transitions from a stable state to a declining state as the turning point, and determines the preceding-to-follow-end semantic deviation index corresponding to this turning point as the individualized deviation tolerance threshold for the respondent. This threshold indicates that for the respondent, under a specified question type, when the preceding-to-follow-end semantic deviation index of the respondent's long question does not exceed this threshold, the respondent can generally still adopt the prompts generated by the agent relatively stably; when it exceeds this threshold, the prompts generated under the default prompt generation start length are more likely to be inconsistent with the final semantic direction of the complete access question, thus leading to a decrease in the effective adoption rate.

[0085] In step S300, the system can determine the acceptable range of semantic deviation between the respondent's prior and subsequent responses under a specified question type from the respondent's historical interview data. The significance of this individualized deviation tolerance threshold lies in providing a quantitative basis for subsequently judging whether the target interviewer's expression style will affect the respondent's effective use of prompting information. Only when the target interviewer's prior and subsequent semantic deviation index exceeds this threshold will the system adjust the prompt generation startup length of the streaming interview prompt pre-generated agent, thereby balancing the real-time nature of prompt generation and the effectiveness of the prompt content.

[0086] Furthermore, the interview content analysis method based on multi-agent collaboration also includes the following steps:

[0087] Step S400: Obtain a second sample corresponding to the specified question type from the target visitor's historical access data, and calculate the target visitor's forward and backward semantic deviation index based on the second sample.

[0088] The acquisition of the second sample also includes screening the interview background; the interview background includes at least one of the following: interview topic, interview stage, interview objective, and related event type.

[0089] In this embodiment of the invention, the second sample is used to characterize the target interviewer's expression habits when asking similar long questions in historical interviews. The first sample is mainly used to analyze the interviewee's use of prompting information, while the second sample is mainly used to determine the target interviewer's semantic deviation index before and after the specified question type.

[0090] Since target visitors may express themselves differently under different question types, the system selects a second sample corresponding to the specified question type from the target visitor's historical visit data. Furthermore, to avoid interference caused by differences in interview backgrounds, the acquisition of the second sample also includes interview background screening, that is, prioritizing historical samples whose interview topic, interview stage, interview goal, or related event type matches the current interview question.

[0091] After obtaining the second sample, the system calculates the preceding and following semantic deviation index corresponding to the second sample according to the aforementioned calculation method. When there are multiple second samples, the preceding and following semantic deviation indices of multiple second samples can be averaged or weighted to obtain the preceding and following semantic deviation index of the target visitor.

[0092] The significance of identifying the second sample lies in providing a visitor-side basis for subsequent judgments on whether the prompt generation startup length needs to be corrected. Step S300 yields the individualized deviation tolerance threshold for the respondent, while step S400 yields the semantic deviation index of the target visitor's preceding and following steps. By comparing the two, it can be determined whether the target visitor's long question expression exceeds the respondent's adaptation range.

[0093] Furthermore, the interview content analysis method based on multi-agent collaboration also includes the following steps:

[0094] Step S500: When the semantic deviation index of the target visitor before and after the visit is greater than the individualized deviation tolerance threshold, a timing correction factor is generated based on the excess magnitude, and the prompt generation start length of the pre-generated agent for the streaming interview prompt is corrected accordingly.

[0095] Specifically, Figure 3 A flowchart is shown to correct the startup length generated by the prompt.

[0096] Specifically, when the semantic deviation index of the target visitor exceeds the individualized deviation tolerance threshold, a timing correction factor is generated based on the excess magnitude, and the prompt generation start length of the pre-generated agent for streaming interview prompts is corrected accordingly. This includes the following steps:

[0097] Step S501: Determine whether the semantic deviation index of the target visitor before and after the visit is greater than the individualized deviation tolerance threshold.

[0098] Step S502: If not, when the identified content of the current access question reaches the default prompt generation start length, the control stream interview prompt pre-generation agent executes formal prompt generation.

[0099] Step S503, if yes, then generate a timing correction factor based on the extent to which the target visitor's semantic deviation index exceeds the individualized deviation tolerance threshold.

[0100] Step S504: Adjust the default prompt generation startup length according to the timing correction factor to obtain the corrected prompt generation startup length;

[0101] In step S505, when the identified content of the current access question reaches the starting length for generating correction prompts, the control stream interview prompt pre-generation agent performs formal prompt generation based on the identified content and generates interview auxiliary prompts.

[0102] When increasing or correcting the default startup length, a preset correction function is used, which is:

[0103] ;

[0104] in, This refers to shifting the startup length prompt. This refers to the default startup length prompt. This refers to the maximum possible value for the startup length that can be generated. This refers to the semantic deviation index between the target visitor's preceding and following journey. This refers to individual deviations from the permissible threshold. This refers to exceeding the allowable range. This refers to a preset correction strength coefficient, and satisfies... Greater than 0.

[0105] The prompt generation start length is at least one of the following: the length of identified text, the duration of identified speech, and the number of identified semantic units since the start of the current access question; wherein, the default prompt generation start length, the corrected prompt generation start length, and the upper limit of the prompt generation start length all use the same measurement type.

[0106] In this embodiment of the invention, step S500 is used to adjust the prompt generation start-up length of the streaming interview prompt pre-generated agent based on the comparison result between the target visitor's forward and backward semantic deviation index and the interviewee's individualized deviation tolerance threshold. This step is a key step in forming closed-loop control in this invention. Specifically, it first determines the individualized deviation tolerance threshold that the interviewee can adapt to using the first historical sample, then determines the target visitor's forward and backward semantic deviation index using the second sample, and finally decides whether to adjust the prompt generation start-up length based on the difference between the two.

[0107] When the semantic deviation index of the target interviewer before and after the interview does not exceed the individualized deviation tolerance threshold, it indicates that the deviation of the target interviewer's expression before and after the interview under the specified question type is still within the range that the interviewee can adapt to. At this time, the streaming interview prompt pre-generation agent can continue to use the default prompt generation start length, without increasing the waiting length of the identified content, so as to ensure that the interview auxiliary prompts can be generated early and provided to the interviewee in a timely manner.

[0108] When the semantic deviation index of the target interviewer's preceding and following segments exceeds the individualized deviation tolerance threshold, it indicates that the target interviewer's long question expression style has exceeded the interviewee's adaptation range for this type of semantic deviation. If the formal prompt generation is still initiated at the default prompt generation start length, the content relied upon by the agent may be too concentrated on the preceding expression fragments, leading to a deviation between the interview assistance prompts and the final semantic direction of the complete interview questions. Therefore, this invention generates a timing correction factor based on the extent to which the target interviewer's semantic deviation index exceeds the individualized deviation tolerance threshold, and accordingly increases the prompt generation start length.

[0109] The significance of employing an overshoot generation timing correction factor lies in making the adjustment of the prompt generation initiation length proportional and limited. That is, it doesn't drastically delay agent initiation simply because the individualized deviation tolerance threshold is exceeded; instead, the increase in prompt generation initiation length is determined by the degree of exceedance. When the exceedance is small, only a small increase in prompt generation initiation length is made; when the exceedance is large, the increase is correspondingly larger. This allows the agent to collect more subsequent expression content, reducing the risk of semantic bias caused by prematurely generated prompts, while also avoiding excessively delaying formal prompt generation and preventing the provision of interview-supporting prompts too late, thus affecting the respondent's normal adoption.

[0110] In one specific implementation, a linear correction method can be used to determine the corrected prompt generation start length. Specifically, first, the proportion by which the target visitor's semantic deviation index exceeds the individualized deviation tolerance threshold is calculated. Then, this proportion is multiplied by a preset correction intensity coefficient to obtain a timing correction factor. The timing correction factor is then used to increase the default prompt generation start length, resulting in the corrected prompt generation start length. To avoid the corrected prompt generation start length being too large, a maximum allowed prompt generation start length can be set; when the calculated corrected prompt generation start length exceeds the maximum allowed prompt generation start length, the maximum allowed prompt generation start length is taken as the final corrected prompt generation start length.

[0111] The preset correction strength coefficient is used to control the sensitivity to increases in the prompt generation start length. This coefficient can be predetermined based on historical interview data, such as by statistically analyzing the correlation between "excess magnitude" and "effective adoption rate decline" in multiple historical samples; alternatively, it can be preset by the system based on different interview scenarios, the average generation time of the agent, the prompt terminal output time, and the acceptable prompt waiting range for the interviewee. Generally, the higher the real-time requirement, the smaller the preset correction strength coefficient can be set; the higher the requirement for prompt accuracy, the larger the preset correction strength coefficient can be set.

[0112] In addition to the linear correction method mentioned above, piecewise correction, table lookup correction, or model prediction methods can also be used. For example, the excess range can be divided into three intervals: low, medium, and high, each corresponding to a different increase in the prompt generation start length; a mapping table between the excess range and the corrected prompt generation start length can also be established based on historical data; or a regression model or classification model can be used to predict a suitable corrected prompt generation start length based on the target visitor's semantic deviation index before and after the interview, the interviewee's individualized deviation tolerance threshold, the default prompt generation start length, and interview scenario parameters.

[0113] For example, in a management talent interview, the current question was identified as a long-form question related to project debriefing. The system retrieved multiple first samples of the same type from the respondent's historical interview data and calculated the corresponding semantic deviation indices for the preceding and following stages to be 0.25, 0.34, 0.43, 0.52, 0.63, and 0.74, with corresponding effective adoption rates of 82%, 81%, 80%, 79%, 70%, and 62%. After trend detection of the effective adoption rate sequence, the system found that the effective adoption rate remained relatively stable before the semantic deviation index reached 0.52; after exceeding 0.52, the effective adoption rate began to decline. Therefore, the system determined 0.52 as the individualized deviation tolerance threshold for this respondent in the long-form question related to project debriefing.

[0114] Subsequently, the system retrieves a second sample corresponding to the long question in the project review category from the target visitor's historical access data and calculates the target visitor's semantic deviation index before and after the process to be 0.65. Since 0.65 is greater than 0.52, it indicates that the target visitor's expression of the long question has exceeded the individualized deviation tolerance threshold for that respondent. At this point, the system calculates the excess ratio, which is 0.65 minus 0.52 and then divided by 0.52, resulting in approximately 0.25. If the preset correction intensity coefficient is 0.6, then the timing correction factor is 0.15. Assuming the default prompt generation start length is 40 recognized characters, then the corrected prompt generation start length is 40 multiplied by 1.15, which is 46 recognized characters. If the maximum allowed prompt generation start length is 60 recognized characters, then 46 does not exceed the upper limit, therefore, 46 recognized characters are ultimately used as the corrected prompt generation start length.

[0115] After the current access question begins, if the default prompt generation start length is used, the streaming interview prompt pre-generation agent will initiate formal prompt generation when 40 characters of content have been identified. However, with this invention, the system will control the streaming interview prompt pre-generation agent to execute formal prompt generation only when 46 characters of content have been identified. This allows the agent to acquire more subsequent expression content than the default method, making its generated interview assistance prompts closer to the final semantic direction of the complete access question. Furthermore, since the corrected prompt generation start length only increases from 40 to 46 characters and is limited by the maximum allowed prompt generation start length, it does not significantly delay the generation time of the interview assistance prompts.

[0116] Through the above steps, this invention can individually adjust the prompt generation start length of the pre-generated agent for streaming interview prompts, taking into account the differences in expression reception adaptation between different respondents and different target interviewers. Compared to always using a fixed default prompt generation start length, this invention can appropriately increase the length of the identified content before the agent begins formal prompt generation when the target interviewer's expression deviates significantly from its initial state. This reduces the likelihood of the prompt information deviating from the final semantic direction of the complete interview question due to premature generation.

[0117] The beneficial effects of this invention are as follows: Firstly, by limiting the individualized deviation tolerance threshold to determine whether the prompt generation start length needs to be corrected, it avoids indiscriminately delaying all long questions. Secondly, by controlling the degree of correction through the excess range, it ensures that the increase in the prompt generation start length matches the deviation of the target interviewer's expression, thereby improving the effectiveness of interview-aided prompts while minimizing real-time loss due to late start. Therefore, this invention effectively addresses the aforementioned core research point: in probing long question scenarios, it balances the timeliness of agent prompt generation with the consistency between the prompt information and the final semantic direction of the complete interview question.

[0118] This invention can be applied to management talent interviews, expert interviews, cadre interviews, corporate executive surveys, project debriefing interviews, media interview assistance, and other scenarios requiring real-time generation of interview assistance prompts. Especially in scenarios with lengthy interview questions, significant differences in interviewers' expression styles, and interviewees needing to organize their responses immediately, this invention can specifically adjust the prompt generation startup length based on historical interview and interviewee data, improving the stability and adaptability of the interview assistance system.

[0119] Furthermore, Figure 4 An application architecture diagram of the system provided in an embodiment of the present invention is shown.

[0120] In another preferred embodiment of the present invention, an interview content analysis system based on multi-agent collaboration includes:

[0121] The question identification module 100 is used to identify in real time the questions raised by the target visitor to the interviewee, and to determine the specified question type when the question is a preset long question.

[0122] Furthermore, the interview content analysis system based on multi-agent collaboration also includes:

[0123] The sample analysis module 200 is used to obtain several first samples with the same question type as the respondents from the respondents' historical interview data, calculate the semantic deviation index between the front and back ends of each first sample and the effective adoption rate of the respondents' feedback information generated by the pre-generated intelligent agent in the live interview.

[0124] Furthermore, the interview content analysis system based on multi-agent collaboration also includes:

[0125] The threshold determination module 300 is used to establish the correlation between the front-end and back-end semantic deviation index and the effective adoption rate. When the effective adoption rate first stabilizes and then decreases as the front-end and back-end semantic deviation index increases, the index corresponding to the trend inflection point is determined as the individualized deviation tolerance threshold for the respondents.

[0126] Furthermore, the interview content analysis system based on multi-agent collaboration also includes:

[0127] The target index calculation module 400 is used to obtain a second sample corresponding to a specified question type from the target visitor's historical access data, and calculate the target visitor's forward and backward semantic deviation index based on the second sample.

[0128] Furthermore, the interview content analysis system based on multi-agent collaboration also includes:

[0129] The startup length correction module 500 is used to generate a timing correction factor based on the excess magnitude when the semantic deviation index of the target visitor before and after the interview exceeds the individualized deviation tolerance threshold, and to correct the startup length of the prompt generation of the pre-generated agent for the streaming interview prompt accordingly.

[0130] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0131] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0132] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0133] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0134] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for analyzing interview content based on multi-agent collaboration, characterized in that, The method includes: Real-time identification of questions asked by target visitors to respondents, and determination of the specified question type when the question is a preset long question; Several first samples with the same question type as the specified questions are obtained from the historical interview data of the respondents. The semantic deviation index of the front and back ends and the effective adoption rate of the prompt information generated by the pre-generated intelligent agent of the respondent's live interview prompt are calculated for each first sample. The calculation process of the effective adoption rate includes: acquiring the prompt information generated by the pre-generated intelligent agent of the streaming interview prompt in the first sample, and the response text formed by the interviewee within the preset response time window after the prompt information is generated; breaking down the prompt information into key points to obtain several key points, and determining the prompt semantic vector of each key point; performing semantic parsing on the response text to determine the response semantic segments in the response text corresponding to each key point; when the semantic similarity between the key point and the corresponding response semantic segment reaches the preset adoption threshold, and the corresponding response semantic segment matches the final semantic direction of the interview question, the key point is determined to be effectively adopted; and calculating the effective adoption rate of the prompt information based on the ratio of the number of effectively adopted key points to the total number of key points. Establish the correlation between the front-end and back-end semantic deviation index and the effective adoption rate. When the effective adoption rate first stabilizes and then decreases as the front-end and back-end semantic deviation index increases, the index corresponding to the trend inflection point is determined as the individualized deviation tolerance threshold for the respondents. The steps include: Several first samples are sorted from smallest to largest according to the semantic deviation index between the beginning and end of the process to obtain the first sample sequence. The effective adoption rate corresponding to each first sample is extracted according to the first sample sequence to form the effective adoption rate change sequence. The correlation between the semantic deviation index between the beginning and end of the process and the effective adoption rate is established based on the effective adoption rate change sequence. Perform trend detection on the effective adoption rate change sequence to determine whether the effective adoption rate shows a trend of first stabilizing and then decreasing as the semantic deviation exponent increases in the preceding and following stages. If so, the semantic deviation index corresponding to the transition position from a stable state to a declining state will be determined as the individualized deviation tolerance threshold for the respondents. If not, then individualized deviations from the allowable threshold will not be generated, and the default prompts for generating the startup length will be generated using streaming interviews to prompt the pre-generated agent. Obtain a second sample corresponding to the specified question type from the target visitor's historical visit data, and calculate the target visitor's forward and backward semantic deviation index based on the second sample; When the semantic deviation index of the target visitor before and after the visit exceeds the individualized deviation tolerance threshold, a timing correction factor is generated based on the excess magnitude, and the prompt generation start length of the pre-generated agent for streaming interview prompts is corrected accordingly.

2. The interview content analysis method based on multi-agent collaboration according to claim 1, characterized in that, The preset long questions are determined as follows: Before the start of this interview, a set of preset question types corresponding to this interview is obtained, and the types of long questions that need to be asked in long expression form are determined from the set of preset question types to form a set of preset long questions; When identifying access questions raised by target visitors in real-time streaming, the candidate question type corresponding to the access question is determined based on the streaming identification text of the access question within a preset start time period; When a candidate question type belongs to a preset long question set, the access question is determined to belong to the preset long question set, and the candidate question type is determined as the specified question type.

3. The interview content analysis method based on multi-agent collaboration according to claim 1, characterized in that, The acquisition of the first and second samples also includes screening the interview background; the interview background includes at least one of the following: interview topic, interview stage, interview objective, and related event type.

4. The interview content analysis method based on multi-agent collaboration according to claim 1, characterized in that, The calculation process of the pre- and post-interview semantic deviation index includes: obtaining the streaming recognition text of the interview questions in the sample and the default prompt generation start length of the streaming interview prompt pre-generated agent; taking the text formed after the current interview question starts and the recognized content reaches the default prompt generation start length as the pre-interview expression segment, and taking the text formed after the recognized content exceeds the default prompt generation start length until the end of the interview question as the post-interview expression segment; determining the semantic vectors of the pre-interview expression segment and the post-interview expression segment respectively, and determining the degree of semantic deviation based on the vector distance between the two semantic vectors; determining the expression length parameter according to the text length of the interview question and / or the duration of the speech; and calculating the pre- and post-interview semantic deviation index based on the degree of semantic deviation and the expression length parameter.

5. The interview content analysis method based on multi-agent collaboration according to claim 1, characterized in that, When the semantic deviation index of the target visitor's preceding and following journey exceeds the individualized deviation tolerance threshold, the steps for adjusting the prompt generation start-up length of the pre-generated agent for streaming interview prompts based on the excess magnitude and the generation timing correction factor include: Determine whether the semantic deviation index of the target visitor before and after the visit is greater than the individualized deviation tolerance threshold. If not, when the identified content of the current access question reaches the default prompt generation start length, the control stream interview prompt pre-generation agent will execute the formal prompt generation; If so, then a timing correction factor is generated based on the extent to which the target visitor's semantic deviation index exceeds the individualized deviation tolerance threshold. The default prompt generation startup length is increased and corrected based on the timing correction factor to obtain the corrected prompt generation startup length; When the identified content of the current access question reaches the starting length for corrective prompt generation, the control flow interview prompt pre-generation agent performs formal prompt generation based on the identified content and generates interview auxiliary prompts.

6. The interview content analysis method based on multi-agent collaboration according to claim 5, characterized in that, When increasing or correcting the default startup length, a preset correction function is used, which is: ； in, This refers to shifting the startup length prompt. This refers to the default startup length prompt. This refers to the maximum possible value for the startup length that can be generated. This refers to the semantic deviation index between the target visitor's preceding and following journey. This refers to individual deviations from the permissible threshold. This refers to exceeding the allowable range. This refers to a preset correction strength coefficient, and satisfies... Greater than 0.

7. The interview content analysis method based on multi-agent collaboration according to claim 6, characterized in that, The prompt generation start length is at least one of the following: the length of identified text, the duration of identified speech, and the number of identified semantic units since the start of the current access question; wherein, the default prompt generation start length, the corrected prompt generation start length, and the upper limit of the prompt generation start length all use the same measurement type.

8. A multi-agent collaborative interview content analysis system, characterized in that, The system includes: The question identification module is used to identify the questions raised by the target visitor to the interviewee in real time, and to determine the specified question type when the question is a preset long question; The sample analysis module is used to obtain several first samples with the same question type as the respondents from the respondents' historical interview data, calculate the semantic deviation index before and after the interview for each first sample, and the effective adoption rate of the respondents' responses to the prompts generated by the pre-generated intelligent agent for streaming interview prompts. The calculation process of the effective adoption rate includes: obtaining the prompts generated by the pre-generated intelligent agent for streaming interview prompts in the first samples, and the response text formed by the respondents within the preset response time window after the prompts are generated; breaking down the prompts into key points to obtain several key points, and determining the prompt semantic vector of each key point; performing semantic parsing on the response text to determine the response semantic segments in the response text that correspond to each key point; when the semantic similarity between the key point and the corresponding response semantic segment reaches the preset adoption threshold, and the corresponding response semantic segment matches the final semantic direction of the interview question, the key point is determined to be effectively adopted; based on the ratio of the number of effectively adopted key points to the total number of key points, the effective adoption rate of the prompt information is calculated. The threshold determination module is used to establish the correlation between the front-end and back-end semantic deviation index and the effective adoption rate. When the effective adoption rate first stabilizes and then decreases as the front-end and back-end semantic deviation index increases, the index corresponding to the trend inflection point is determined as the individualized deviation tolerance threshold for the respondents. Several first samples are sorted from smallest to largest according to the semantic deviation index between the beginning and end of the process to obtain the first sample sequence. The effective adoption rate corresponding to each first sample is extracted according to the first sample sequence to form the effective adoption rate change sequence. The correlation between the semantic deviation index between the beginning and end of the process and the effective adoption rate is established based on the effective adoption rate change sequence. Perform trend detection on the effective adoption rate change sequence to determine whether the effective adoption rate shows a trend of first stabilizing and then decreasing as the semantic deviation exponent increases in the preceding and following stages. If so, the semantic deviation index corresponding to the transition position from a stable state to a declining state will be determined as the individualized deviation tolerance threshold for the respondents. If not, then individualized deviations from the allowable threshold will not be generated, and the default prompts for generating the startup length will be generated using streaming interviews to prompt the pre-generated agent. The target index calculation module is used to obtain a second sample corresponding to a specified question type from the historical access data of the target visitor, and calculate the semantic deviation index of the target visitor before and after the visit based on the second sample. The startup length correction module is used to generate a timing correction factor based on the magnitude of the deviation when the semantic deviation index of the target visitor exceeds the individualized deviation tolerance threshold, and to correct the startup length of the prompt generation of the pre-generated agent for the streaming interview prompt accordingly.

Images