Interview method and device based on ai primary and backup expert model

Abstract

The application provides an interview method and device based on an AI primary and backup expert model, and relates to the technical field of artificial intelligence. In the application, when determining the expert set corresponding to each interview feature information of a candidate object, a primary expert is determined based on affinity score sorting and a preset cumulative confidence threshold. When the number of primary experts is less than a preset minimum number of experts, a backup expert is determined through double random matrix transformation of the Sinkhorn algorithm. In this way, the cumulative confidence threshold drives dynamic expert activation, adaptive computing resource allocation is achieved, the computing efficiency is improved under the condition of ensuring the accuracy of the interview result output by the model, unnecessary computing overhead is avoided, and when the number of activated experts is insufficient, the Sinkhorn algorithm is used for double random matrix transformation to select experts with the most global balance as backup experts, so that the load distribution is balanced and experts are not idle.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to an interview method and device based on an AI primary and backup expert model. BACKGROUND

[0002] In recent years, with the rapid development of artificial intelligence (AI) in various fields, for example, the application of AI interviewers in the interview link of the recruitment process, especially the breakthrough of basic large models, the requirement for model capability is increasing. However, the traditional dense deep learning model (referred to as dense model) faces severe challenges when pursuing higher performance. Although the increase in model parameter quantity can improve the model capacity (that is, the ability of the model to learn and represent complex information), it will also cause the consumption of computing resources to grow exponentially. Figure 1 As shown in (a) and (b) of FIG. 1, the Dense model needs to activate all its parameters for calculation when processing each input, which makes it increasingly difficult to train and deploy larger-scale models, and even unsustainable in some scenarios. In addition, modern datasets are increasingly diverse and complex, often containing multi-modal data and complex structural relationships, and a single Dense model is difficult to effectively integrate heterogeneous and even conflicting knowledge. Figure 2 In order to cope with these challenges, the Mixture of Experts (MoE) model architecture has emerged and is receiving widespread attention again. The core idea of the MoE model is to adopt a “divide and conquer” strategy, which allows the model to maintain the calculation cost of each forward propagation (inference) within a controllable range while significantly increasing the total parameter quantity. The design motivation is to break the strong coupling relationship between the model capacity and the calculation cost, and achieve a better balance between model performance and efficiency. As shown in (b) and (c) of FIG. 1, the MoE model (belonging to a sparse model) processes each input by dynamically selecting and activating the most relevant part of the sub-model (i.e., “expert”), avoiding the huge calculation overhead of activating the entire network and reducing the delay. Research shows that the MoE architecture can significantly improve model performance and efficiency with less computing resources, especially in handling large-scale, multi-modal, and heterogeneous complex data.

[0003] Figure 1 Figure 3

[0004] ​​​However, the existing MoE model usually adopts a fixed Top-K expert routing strategy, that is, for each input token (i.e., the basic unit of text data, which can be a word, subword or single character), the routing mechanism (i.e., the gating network) calculates the affinity score of the input token with all N experts, selects the K experts with the highest scores for activation and combines the shared expert for calculation, and K is pre-set. However, this routing mechanism lacks flexibility and is difficult to achieve optimal allocation of computing resources. SUMMARY

[0005] The purpose of the present application is to provide an interview method and device based on an AI primary-backup expert model, to realize adaptive allocation of computing resources, improve computing efficiency, avoid unnecessary computing overhead, ensure balanced load distribution, and avoid idle experts while ensuring the accuracy of the interview results output by the model.

[0006] In a first aspect, the present application provides an interview method based on an AI primary-backup expert model, comprising:

[0007] obtaining at least one interview feature information of a candidate object;

[0008] calculating the affinity score between each interview feature information and a plurality of preset candidate experts;

[0009] determining the expert set corresponding to each interview feature information according to the affinity score of each interview feature information to each candidate expert; wherein the expert set includes a primary expert determined based on the ranking of the affinity score and a preset cumulative confidence threshold; when the number of primary experts is less than a preset minimum number of experts, the expert set further includes a backup expert determined through double random matrix transformation of the Sinkhorn algorithm;

[0010] calculating at least one interview feature information by activating the expert set corresponding to each interview feature information to obtain the interview result of the candidate object.

[0011] In an optional implementation, the affinity score between each interview feature information and a plurality of preset candidate experts is calculated, comprising:

[0012] for each interview feature information, calculating the original routing score of the interview feature information to each candidate expert according to the interview feature information and the expert gating vector of each candidate expert;

[0013] obtaining the affinity score of the interview feature information to each candidate expert by Sigmoid function normalization according to the original routing score of the interview feature information to each candidate expert.

[0014] In an optional implementation, the calculation formula of the original routing score is:

[0015] ;

[0016] wherein, denotes the original routing score of the i-th interview feature information to the j-th candidate expert, t denotes the i-th interview feature information, i denotes the expert gating vector of the j-th candidate expert, denotes the trainable bias term of the j-th candidate expert; t i i

[0017] The calculation formula of the affinity score is:

[0018] ;

[0019] wherein, denotes the affinity score of the i-th interview feature information to the j-th candidate expert, t denotes the application of the Sigmoid function to z. i In an optional implementation, according to the affinity score of each interview feature information to each candidate expert, the expert set corresponding to each interview feature information is determined, comprising:

[0020] For each interview feature information, the affinity scores of the interview feature information to each candidate expert are sorted in descending order to obtain a sorting result; according to the sorting result, the affinity scores are accumulated, and the candidate expert corresponding to the cumulative confidence threshold reached by the accumulated sum of the affinity scores, or the candidate expert corresponding to the maximum allowed number of experts reached by the number of candidate experts participating in the accumulation is determined as the main expert corresponding to the interview feature information;

[0021] The interview feature information with a number of main experts less than the minimum number of experts is determined as simple feature information; in the case of shielding the main expert, the normalized score matrix corresponding to the simple feature information is obtained through the double random matrix transformation of the Sinkhorn algorithm, and the normalized score matrix includes the normalized score of the simple feature information to each candidate expert; a number of candidate experts with the largest normalized scores in the normalized score matrix are determined as the backup experts corresponding to the simple feature information.

[0021] In an optional implementation, according to the sorting result, the affinity scores are accumulated, and the candidate expert corresponding to the cumulative confidence threshold reached by the accumulated sum of the affinity scores, or the candidate expert corresponding to the maximum allowed number of experts reached by the number of candidate experts participating in the accumulation is determined as the main expert, comprising:

[0022]

[0023] In an optional implementation, according to the sorting result, the affinity scores are accumulated, and the candidate expert corresponding to the cumulative confidence threshold reached by the accumulated sum of the affinity scores, or the candidate expert corresponding to the maximum allowed number of experts reached by the number of candidate experts participating in the accumulation is determined as the main expert, comprising: ​​​​​

[0024] taking out the affinity scores in sequence according to the sorting result;

[0025] judging whether the accumulation of the current taken affinity scores reaches a cumulative confidence threshold value;

[0026] if the cumulative confidence threshold value is not reached, judging whether the number of candidate experts corresponding to the current taken affinity scores reaches a maximum allowed number of experts;

[0027] if the maximum allowed number of experts is not reached, re-executing the step of taking out the affinity scores in sequence according to the sorting result;

[0028] if the cumulative confidence threshold value is reached or the maximum allowed number of experts is reached, determining the candidate expert corresponding to the current taken affinity scores as the main expert.

[0029] In an optional embodiment, in the case of shielding the main expert, the normalized score matrix corresponding to the simple feature information is obtained through double random matrix transformation of the Sinkhorn algorithm, including:

[0030] obtaining the original routing scores of the simple feature information to each candidate expert, and resetting the original routing score corresponding to the main expert to negative infinity;

[0031] transforming the original routing scores of the simple feature information to each candidate expert into an exponential transformation score matrix through exponential transformation;

[0032] transforming the exponential transformation score matrix into a normalized score matrix through double random matrix transformation of the Sinkhorn algorithm.

[0033] In an optional embodiment, the interview result of the candidate object is obtained by calculating at least one interview feature information through activating the expert set corresponding to each interview feature information, including:

[0034] for each interview feature information, according to the affinity score of the interview feature information to each activated expert in the corresponding expert set, the gating weight of the interview feature information on each activated expert is obtained through normalization processing;

[0035] inputting the interview feature information into the corresponding expert set for processing to obtain the expert result output by each activated expert;

[0036] weighting and aggregating the expert result output by each activated expert according to the gating weight of the interview feature information on each activated expert to obtain the target output result of the interview feature information;

[0037] generating the interview result of the candidate object according to the target output result of each interview feature information.

[0038] In a second aspect, the present application provides an interview device based on an AI primary and backup expert model, comprising:

[0039] a feature acquisition module configured to acquire at least one interview feature information of a candidate object;

[0040] a score calculation module configured to calculate an affinity score between each interview feature information and a plurality of preset candidate experts;

[0041] an expert set determination module configured to determine an expert set corresponding to each interview feature information according to the affinity scores of each interview feature information to each candidate expert; wherein the expert set comprises a primary expert determined based on the ordering of the affinity scores and a preset cumulative confidence threshold; when the number of primary experts is less than a preset minimum number of experts, the expert set further comprises a backup expert determined through double random matrix transformation of the Sinkhorn algorithm;

[0042] a result determination module configured to calculate at least one interview feature information by activating the expert set corresponding to each interview feature information, and obtain an interview result of the candidate object.

[0043] In a third aspect, the present application provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program capable of running on the processor, and the processor executes the computer program to implement the interview method based on the AI primary and backup expert model of any one of the preceding embodiments.

[0044] In a fourth aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is run by a processor to execute the interview method based on the AI primary and backup expert model of any one of the preceding embodiments.

[0045] This invention provides an interview method and apparatus based on an AI master-slave expert model. The method includes: acquiring at least one interview feature information of a candidate; calculating the affinity score between each interview feature information and multiple preset candidate experts; determining the expert set corresponding to each interview feature information based on the affinity score of each interview feature information to each candidate expert; wherein the expert set includes master experts determined based on the affinity score ranking and a preset cumulative confidence threshold; when the number of master experts is less than a preset minimum number of experts, the expert set also includes backup experts determined by a double random matrix transformation using the Sinkhorn algorithm; by activating the expert set corresponding to each interview feature information, calculating at least one interview feature information to obtain the interview result of the candidate. This dynamic expert activation driven by the cumulative confidence threshold achieves adaptive allocation of computational resources, improving computational efficiency and avoiding unnecessary computational overhead while ensuring the accuracy of the interview results output by the model; and when the number of activated experts is insufficient, the double random matrix transformation using the Sinkhorn algorithm selects the most globally balanced expert as a backup expert, ensuring balanced load distribution and avoiding expert idleness. Attached Figure Description

[0046] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0047] Figure 1 Here are schematic diagrams of the network structures of the Dense model and the MoE model;

[0048] Figure 2 This is a schematic diagram of the architecture of the Dense model;

[0049] Figure 3 This is a schematic diagram of the architecture of a sparse model;

[0050] Figure 4 This is a schematic diagram of the MoE model architecture;

[0051] Figure 5 A flowchart illustrating an interview method based on an AI master-slave expert model provided in an embodiment of the present invention;

[0052] Figure 6 A flowchart illustrating another interview method based on an AI master-slave expert model provided in an embodiment of the present invention;

[0053] Figure 7A structure schematic diagram of an interview device based on an AI primary and backup expert model is provided for an embodiment of the present application.

[0054] Figure 8 A structure schematic diagram of an electronic device is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0055] The technical solutions of the present application will be described clearly and completely in combination with embodiments. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0056] The MoE model strategy decomposes complex tasks or data spaces into multiple specialized "expert" networks, which are responsible for processing different parts of the input. Unlike traditional dense models, which activate all parameters when processing each input, the MoE model dynamically selects and activates a small subset of experts that are most relevant to the current input through a gating network or router. As shown in the following figure, the MoE model typically includes the following core components: Figure 4

[0057] Experts: Each expert is usually an independent neural network, and there are many of them. In the Transformer architecture, they replace the feed-forward network layer (FFN). Experts focus on learning unique and specialized knowledge and are responsible for processing inputs in specific domains.

[0058] Shared Experts: Independent of the experts, these shared experts are always activated for all inputs and are responsible for processing all inputs. Shared experts focus on cross-domain general knowledge and effectively reduce the overlap of learning content between routing experts and alleviate the problem of knowledge redundancy.

[0059] Gating Network: Receives input and decides which experts to route the input to and how to combine the outputs of these experts, usually generating weights for each expert.

[0060] Routing Mechanism: Selects experts based on the output of the gating network. The most common is Top-K routing, which selects the top K experts with the highest scores.

[0061] ​Although the current mainstream MoE hybrid expert model is more efficient than the traditional dense architecture, and mostly adopts a fixed Top-K expert routing strategy (as shown in Figure 4 Although this method is simple and effective, it ignores the fact that different tokens may have different complexities or diversities. For simple tokens, activating K experts may waste computing resources; for tokens that require complex reasoning or knowledge, the fixed K experts may not be able to provide sufficient computing power or model capacity, thereby limiting the upper limit of the model's performance. This "one-size-fits-all" approach obviously lacks flexibility and is difficult to achieve optimal allocation of computing resources.

[0062] To overcome the above limitations of the traditional Top-K expert routing strategy, the AI primary and backup expert model-based interview method and device provided by the embodiments of the present application adopts a dynamic adaptive routing strategy that combines gated confidence and Sinkhorn double random matrix, which can improve the large model reasoning ability of the AI interviewer. The double random matrix is a square matrix, where the sum of the elements in each row and each column is equal to 1.

[0063] The core idea of the embodiments of the present application is to drive dynamic expert activation through confidence, i.e., set a cumulative confidence threshold (such as 0.8), and add it in order according to the score ranking of the experts, and stop selecting experts when the cumulative confidence threshold is met, and when the number of activated experts is insufficient, use the Sinkhorn algorithm to convert the scoring matrix of the remaining experts into an approximate double random distribution, thereby selecting the most globally balanced experts as backups, ensuring fair load distribution and avoiding idle experts. This ensures that enough experts are activated and provides more uniform gradient signals for backpropagation.

[0064] Unlike Top-K selection of a fixed number of experts, the dynamic adaptive routing strategy based on the combination of gated confidence and double random matrix is more refined and more intelligent in adaptively allocating computing resources compared to the static expert selection strategy, improving computing efficiency and avoiding unnecessary computational overhead, while taking into account sparsity and balance to improve model performance.

[0065] To facilitate understanding of the present embodiment, first, a AI primary and backup expert model-based interview method is described in detail.

[0066] The AI primary and backup expert model-based interview method provided by the embodiments of the present application can be executed by an electronic device with data processing capability. Referring to the flowchart of the AI primary and backup expert model-based interview method shown in Figure 5 The method mainly includes the following steps S510 to S540:

[0067] Step S510, obtaining at least one interview feature information of the candidate.

[0068] The candidate can be a candidate for a certain post.

[0069] In some possible embodiments, the step S510 can include obtaining an interview video of the candidate, the interview video including question answering voice and interview image data; and performing feature extraction on the interview video to obtain at least one interview feature information, the interview feature information can include one or more of text features, voice features, expression features and behavior features. The feature extraction can be but not limited to realized through a neural network (such as a Transformer).

[0070] In specific implementation, the question answering voice can be converted into answering text, and the answering text can be analyzed by natural language processing technology to obtain text features. By analyzing the vocabulary, syntax structure and semantics used by the candidate when answering the question, the communication ability, professional knowledge and logical thinking of the candidate can be evaluated.

[0071] The question answering voice can be subjected to voice feature recognition such as tone, rhythm, fluency, etc. to obtain voice features, so as to evaluate one or more of the emotional state, self-confidence and nervousness of the candidate.

[0072] The interview image data can be subjected to facial expression recognition and behavior analysis to obtain expression features and behavior features. Computer vision technology can be used to recognize micro-expressions on the face, and the emotional reactions of the candidate such as excitement, nervousness, honesty, etc. can be inferred therefrom; non-verbal behaviors of the candidate such as gesture use, eye contact, etc. can be analyzed to obtain information on the personality traits of the candidate.

[0073] Step S520, calculating an affinity score between each interview feature information and a plurality of preset candidate experts.

[0074] The candidate expert is a pre-trained expert network (such as a neural network), different candidate experts have different processing methods for input, and thus the interview feature information they adapt to can be different, so it is necessary to match a suitable expert for each interview feature information. One interview feature information can be processed by one or more experts. In order to determine the expert suitable for the interview feature information, it is necessary to calculate an affinity score between each interview feature information and a plurality of preset candidate experts, the affinity score being used to represent the matching degree between the interview feature information and the corresponding candidate expert, so as to filter out the expert set corresponding to each interview feature information based on the affinity score.

[0075] In some possible embodiments, each expert has a gating vector, which is used to calculate an affinity score with the input token. The expert gating vector is a parameter weight matrix trained and used to represent the characteristics of the corresponding expert. Based on this, step S520 above may include: for each interview feature information, calculating the original routing score of the interview feature information for each candidate expert based on the interview feature information and the expert gating vector of each candidate expert; and obtaining the affinity score of the interview feature information for each candidate expert by normalizing it using the Sigmoid function based on the original routing score of the interview feature information for each candidate expert.

[0076] A linear computation (e.g., dot product plus bias) can be performed on the input token and the expert gating vector, and then the affinity score of each token to each candidate expert can be obtained through the sigmoid function. Based on this, in one possible implementation, the formula for calculating the original routing score can be:

[0077] ;

[0078] in, Indicates the first t The interview feature information for the first i The raw routing scores of each candidate expert. Indicates the first t Interview characteristics information, Indicates the first i The expert gating vector of each candidate expert. Indicates the first i Trainable bias terms for each candidate expert. Right now and The dot product of represents the first . t The first interview characteristic information and the first i The degree of similarity or matching between candidate experts. A non-assisted loss strategy is adopted (it does not directly rely on an externally defined loss function to adjust the bias, but instead optimizes indirectly based on the results of route selection). It only affects the route decision and is used to adjust the activation probability of experts to avoid route collapse (always select several experts with higher weights), thereby dynamically achieving load balancing (if the utilization rate of an expert is low, the bias is increased, and vice versa).

[0079] The original routing score can be normalized to a score in the interval [0,1] using the Sigmoid function. In one possible implementation, the affinity score can be calculated as follows:

[0080] ;

[0081] in, Indicates the first t The interview feature information for the first i The affinity score of each candidate expert. This indicates that the Sigmoid function is applied to z.

[0082] In another possible implementation, a temperature hyperparameter can be introduced when calculating the affinity score. Lower temperatures result in a sharper output, making the selected expert more certain; conversely, higher temperatures result in a smoother output, making the selected expert more random. That is, to control the sharpness or smoothness of the output, a temperature hyperparameter T can be introduced into the sigmoid function. Adjusting this hyperparameter can affect the certainty of the expert's selection. Based on this, the formula for calculating the affinity score can be:

[0083] ;

[0084] When T is low: the output of the sigmoid function is sharper, meaning that some experts have much higher affinity scores than others. This makes the model more inclined to select the few experts with the highest affinity scores, thus increasing the certainty of the selection;

[0085] When T is high: the output of the sigmoid function is smoother, and the differences in affinity scores among all experts decrease. In this case, the probability of choosing any one expert increases, making the expert selection more random.

[0086] Step S530: Determine the expert set corresponding to each interview feature based on the affinity score of each candidate expert for each interview feature; wherein, the expert set includes the main experts determined based on the ranking of affinity scores and a preset cumulative confidence threshold; when the number of main experts is less than the preset minimum number of experts, the expert set also includes backup experts determined by the double random matrix transformation of the Sinkhorn algorithm.

[0087] When determining the expert set corresponding to each interview feature, a primary expert can be selected using Confidence-Based Routing, and backup experts can be selected using Sinkhorn Routing. When selecting the primary expert, the affinity scores of each candidate expert can be sorted in descending order, and then the scores are accumulated starting from the highest score until the cumulative sum is greater than or equal to a preset threshold (cumulative confidence threshold). The candidate experts selected during the accumulation process are marked as the initially selected primary experts, and their original affinity scores and weights are recorded. The weights can be obtained by normalizing the affinity scores of each primary expert. When selecting backup experts, it can be checked whether the number of initially selected primary experts is less than the preset minimum number of experts (min).k If the number of selected experts is not less than the minimum number of experts, the backup expert is not needed to be selected; if the number of selected experts is less than the minimum number of experts, all the candidate experts not initially selected are used, and the selected primary experts are shielded (e.g., the original routing score logits of the primary experts are set to negative infinity), and then the Sinkhorn normalization is applied to the logits of the candidate experts not selected to obtain the backup experts (i.e., the backup experts).

[0088] In some possible embodiments, a maximum allowed number of experts is preset, the maximum allowed number of experts being a maximum number of experts allowed to be activated, to prevent too many experts from being selected. The number of the backup experts and the number of the primary experts should be not less than the minimum number of experts and not more than the maximum allowed number of experts. Based on this, the step S530 can include: for each interview feature information, sorting the affinity scores of the candidate experts for the interview feature information in descending order to obtain a sorting result; accumulating the affinity scores according to the sorting result, and determining, as the primary expert corresponding to the interview feature information, the candidate expert corresponding to the cumulative confidence threshold of the accumulated sum of the affinity scores or the candidate expert corresponding to the maximum allowed number of experts of the number of candidate experts participating in the accumulation; determining, as simple feature information, the interview feature information whose number of primary experts is less than the minimum number of experts; obtaining, by double random matrix transformation of the Sinkhorn algorithm, a normalized score matrix corresponding to the simple feature information in the case of shielding the primary experts, the normalized score matrix including the normalized scores of each candidate expert for the simple feature information; and determining, as the backup experts corresponding to the simple feature information, the candidate experts with the largest normalized scores in the normalized score matrix.

[0089] In a possible implementation, the primary expert can be determined by the following process: the affinity scores are sequentially taken out according to the sorting result; it is judged whether the accumulated sum of the currently taken-out affinity score reaches the cumulative confidence threshold; if the cumulative confidence threshold is not reached, it is judged whether the number of candidate experts corresponding to the currently taken-out affinity score reaches the maximum allowed number of experts; if the maximum allowed number of experts is not reached, the step of sequentially taking out the affinity scores according to the sorting result is re-executed; if the cumulative confidence threshold is reached or the maximum allowed number of experts is reached, the candidate expert corresponding to the currently taken-out affinity score is determined as the primary expert.

[0090] In a possible implementation, the normalized score matrix corresponding to the simple feature information can be obtained through the following process: obtaining the original routing scores of the simple feature information on each candidate expert, and resetting the original routing score of the master expert to negative infinity to shield the master expert; through exponential transformation, the original routing scores of the simple feature information on each candidate expert are converted into an exponential transformation score matrix to ensure that all elements in the exponential transformation score matrix (i.e., the exponential transformation score corresponding to each candidate expert) are non-negative; through the double random matrix transformation of the Sinkhorn algorithm, the exponential transformation score matrix is converted into a normalized score matrix.

[0091] The above exponential transformation can be expressed as:

[0092] ;

[0093] If the original routing score z i of the i i th candidate expert is -∞ (negative infinity), then its exponential transformation score Q i = 0.

[0094] The Sinkhorn algorithm is a method for converting any non-negative matrix into a doubly stochastic matrix (Doubly Stochastic Matrix) through an iterative process, that is, a matrix whose row sum and column sum are both 1. It converts the original unbalanced allocation matrix into a balanced, derivable, and approximately optimal allocation structure by alternately normalizing the rows and columns of the matrix. The Sinkhorn algorithm can effectively achieve load balancing and global optimal allocation. The data in the normalized score matrix obtained through the Sinkhorn algorithm represents the probability or weight of being routed to different experts. Through this matrix, the input data (i.e., each interview feature information) can be allocated to the corresponding expert network for processing, that is, to select a backup expert for the interview feature information.

[0095] In step S540, at least one interview feature information is calculated by activating the expert set corresponding to each interview feature information to obtain the interview result of the candidate object.

[0096] For each interview feature information, the maximum weight of each expert in the expert set can be obtained through normalization first, and then the result is output based on the maximum weight of each expert in the expert set, and the optimization is performed through back propagation. Specifically, the weights (such as affinity scores of the main expert and affinity scores / normalized scores of the backup expert) of each expert in the expert set corresponding to each Token (that is, interview feature information) can be normalized so that the sum is 1, and the gating weight of each expert in the expert set corresponding to each Token is obtained. For each Token, the gating weight of the main expert obtained by confidence routing and the gating weight of the backup expert obtained by Sinkhorn routing are combined with the expert output result to calculate the final output result. Based on the final output result corresponding to each Token, the interview result of the candidate object is generated. According to the interview result of the candidate object and the preset true label, the loss signal is calculated through the selected loss function. According to the loss signal, the network parameters and the routing gating parameters of the corresponding expert are updated through back propagation, and the routing gating parameters are used to determine the expert gating vector.

[0097] In some possible embodiments, the above step S540 can include: for each interview feature information, according to the affinity score of each activated expert in the corresponding expert set with respect to the interview feature information, the gating weight of the interview feature information on each activated expert is obtained through normalization processing; the interview feature information is input into the corresponding expert set for processing to obtain the expert result output by each activated expert; the expert results output by each activated expert are weighted and aggregated according to the gating weight of the interview feature information on each activated expert to obtain the target output result of the interview feature information; and the interview result of the candidate object is generated according to the target output result of each interview feature information.

[0098] The interview method based on the AI main-backup expert model provided by the embodiment of the application realizes adaptive allocation of computing resources by driving dynamic expert activation through the cumulative confidence threshold, improves the computing efficiency under the condition of ensuring the accuracy of the interview result output by the model, and avoids unnecessary computing overhead; and when the number of activated experts is insufficient, the Sinkhorn algorithm is used for double-random matrix transformation to select the expert with the most global balance as the backup expert, so that the load distribution is balanced and the experts are not idle.

[0099] The interview method based on the AI main-backup expert model provided by the embodiment of the application mainly includes the following steps:

[0100] 1. Calculate the affinity score of the input token and the expert gating network.

[0101] 2. Sort the affinity scores.

[0102] 3. Determine the initially activated main expert based on the cumulative affinity score and the cumulative confidence threshold.

[0103] 4. Check if the number of activated experts meets the minimum requirement min k .

[0104] 5. If not min k , compute Sinkhorn normalized probabilities for non-activated experts and select backup experts until min k is met.

[0105] 6. Weight normalize the experts selected by confidence and Sinkhorn.

[0106] 7. Return the final routing weights and the set of activated experts.

[0107] 8. Experts output results and compute final results combined with weights, backpropagate to update network parameters.

[0108] For ease of understanding, the following refers to Figure 6 the process of determining the expert set in the above AI primary-backup expert model-based interview method. As Figure 6 shown, the AI primary-backup expert model-based interview method includes the following steps:

[0109] Step S610, compute the affinity score between the token and the expert gate vector.

[0110] Step S620, sort the expert affinity scores in descending order.

[0111] Step S630, determine if there is an affinity score that reaches the cumulative confidence threshold. If not, execute step S640; if yes, execute step S650.

[0112] Step S640, accumulate the affinity scores and select the primary expert. Then re-execute step S630.

[0113] Step S650, determine the primary expert. Then execute step S660.

[0114] Step S660, check if the number of primary experts meets the minimum number of experts. If not, execute step S670; if yes, execute step S690.

[0115] Step S670, apply the Sinkhorn routing strategy to assign backup experts.

[0116] Step S680, determine the backup experts.

[0117] Step S690, determine the routing expert set.

[0118] For ease of understanding, the above interview method based on the AI primary and backup expert model is exemplarily introduced as follows.

[0119] The implementation scheme is set as follows:

[0120] Input token number: 3 (i.e. 3 interview feature information), denoted as , each vector feature dimension D=4 (actually there is also a batch size , B The token will be calculated with the gating network and the expert network.

[0121] Gating network: N=4, including expert gating vector , Each expert has a gating vector for vector dot product with the token to calculate the affinity score.

[0122] Expert network: selected according to the gating network.

[0123] Cumulative confidence threshold threshold: τ =0.8, cumulative score threshold for confidence routing. Stop selecting experts when this threshold is reached or exceeded. A higher threshold generally means that more activated experts need to be selected to meet the confidence requirement.

[0124] Minimum number of activated experts (i.e. minimum number of experts): min k=2, minimum number of activated experts. If the number of experts selected by confidence routing is less than this number, the Sinkhorn backup mechanism will be started. Configurable.

[0125] Maximum number of activated experts (i.e. maximum number of allowed experts): max k=3, maximum number of activated experts. If the number of experts selected by confidence routing is greater than this number, the selection will be stopped to prevent too many experts from being activated. Configurable. Entropy can also be introduced as a dynamic loss, with higher entropy indicating that the candidate experts are more dispersed, with similar scores for each expert, and lower expert certainty; on the contrary, the expert scores are more concentrated, i.e. one or a few experts are significantly higher than the others, and the expert certainty is higher. Entropy is used here as a measure to evaluate the distribution of candidate expert scores; the entropy of the selected experts is calculated and added as an additional term to the total loss function, so that the model optimizes not only the main task loss but also the certainty of expert selection during the training process.

[0126] Selected activated expert set: the selected expert set when the cumulative probability reaches the cumulative confidence threshold and the expert set selected by the fallback (i.e. backup mechanism), which together constitute the final activated expert set for the token.

[0127] Step 1: Calculate the gating network original routing score Logits.

[0128] For each token , calculate the dot product score with all experts:

[0129] ;

[0130] where, is the t th token, represents the semantic features of the current input (i.e., interview feature information), such as the output vector from the previous Transformer layer. is the i th expert gating vector, is the parameter weight matrix learned by training, represents the characteristics of the expert. represents the similarity or matching degree (dot product) between the token and the expert. is the i th expert's trainable bias term. is the linear combination score of the token and the expert gating network.

[0131] For convenience of demonstration, simplify the calculation process, assume:

[0132] Expert gating vector (d=4): , , , .

[0133] Input token vector: , , .

[0134] Expert gating bias (initial value, trainable): .

[0135] Take x1 as an example, calculate the dot product + bias with each expert gate:

[0136] .

[0137] Step 2: Expert gating affinity Sigmoid calculation.

[0138] Calculate the affinity score (i.e., affinity score) of input token with each expert gating network , and normalize it to a score in the interval [0, 1] through Sigmoid. The formula is as follows:

[0139] ;

[0140] where, , the feature vector (dimension t ) of the d th input token; , the i th expert gate vector; , the bias (learnable) of the i th expert gate; , , the linear combination score of the token and the expert gate network; , apply the Sigmoid function to the original score z, map the score to the (0, 1) interval, and convert the score to probability; , the expert score of the Sigmoid output, indicating the affinity of the token t with the expert i , that is, the degree of fit or probability tendency of the token to the expert.

[0141] ;

[0142] z=0.37 represents the matching degree of the token to the expert. s 1,1 =0.591 is the affinity score of the token x 1 assigned to the expert e 1. All , close to 1 indicates "more confident" to be handled by the expert.

[0143] Calculate other affinities, and the final affinity calculation results are shown in Table 1 as follows:

[0144] Table 1

[0145]

[0146] Step 3: Confidence selects experts (cumulative affinity score).

[0147] Take x 1 as an example:

[0148] s i,t The affinity scores are sorted from high to low: [0.591, 0.552, 0.527, 0.512].

[0149] Detect s i,t the maximum value s max in s max whether it is greater than or equal to the cumulative confidence threshold τ=0.8, if s maxThe corresponding expert. Conversely, continue to accumulate until greater than or equal to τ, or reach the preset maximum allowed number of experts max k . The formula is:

[0150] ;

[0151] wherein, is the accumulated expert affinity score; denotes the accumulated next expert affinity score until the accumulated confidence threshold is met or the maximum allowed number of experts is exceeded.

[0152] The scores are sequentially accumulated until the accumulation exceeds the threshold τ = 0.8:

[0153] 1st expert (0.591) → accumulation = 0.591;

[0154] 2nd expert (0.552) → accumulation = 0.591 + 0.552 = 1.143 ≥ 0.8;

[0155] The accumulated expert is the selected activated master expert, so the Top-2 selected {e1, e2} experts.

[0156] Similarly, select experts for x2 and x3. The final expert selection results are shown in Table 2:

[0157] Table 2

[0158]

[0159] Here, each token selects an expert according to the confidence score, which satisfies ≥ 2 experts (min k = 2) and does not exceed max k = 3, without fallback. This reflects the sparsity of conditional calculation (i.e., only expand the calculation when needed, thereby reducing the total calculation amount).

[0160] Step 4: Normalize expert gate weight.

[0161] The selected expert gate score is normalized for weighting the expert output. The formula is as follows:

[0162] ;

[0163] wherein, denotes the token t gate value (weight) on the activated expert i , used to weight control the expert output; denotes the original affinity, used for weight calculation; Score of all selected K experts activated by current token, for normalization; normalized result satisfies , and .

[0164] Normalized gating weight of token x1:

[0165] ;

[0166] Normalized gating weight of token x2:

[0167] ;

[0168] Normalized gating weight of token x3:

[0169] .

[0170] Gating weight determines how much each activated expert contributes to the output.

[0171] Step 5: If the number of selected experts is less than min k , enter Fallback mechanism.

[0172] Backup experts will be assigned for each token (according to the balance principle).

[0173] On the basis of the previous, use the following settings:

[0174] Input token x4 is [0.9, -0.8, 0.2, 0.5];

[0175] The confidence threshold becomes τ = 0.7;

[0176] The minimum number of experts is still min k = 2.

[0177] After the above formula calculation, the affinity score of token x4 is shown in Table 3:

[0178] Table 3

[0179]

[0180] The scores are sorted from high to low: [0.774, 0.670, 0.625, 0.283].

[0181] The confidence threshold τ = 0.7, the fourth expert e4 directly exceeds τ, so only one main expert is selected, but it does not meet min k = 2, triggering Sinkhorn fallback mechanism.

[0182] Step 6: Mask the exponent scores.

[0183] The main expert e4 is selected by the confidence, then masked, and its logits (original dot product scores) are set to -∞ to prevent repeated selection. Then, one backup expert is selected from the remaining experts e1, e2, and e3.

[0184] Through the above calculation, before selecting the expert, the unactivated logits are [0.71, -0.93, 0.51, 1.23], and after masking, the masked_logits are [0.71, -0.93, 0.51, -inf], i.e., setting e4 to negative infinity (-∞). Then, apply Sinkhorn to the logits of the remaining experts and normalize.

[0185] Exponential transformation to ensure all elements are non-negative: if not masked, Q i = exp(z i ); if z i is -∞, Q i = 0.

[0186] The calculation result is:

[0187] ;

[0188] .

[0189] Q The exponential transformation vector score (i.e., the exponential transformation score matrix).

[0190] Step 7: Sinkhorn iteration normalization.

[0191] Through multiple iterations of Sinkhorn regularization, it is approximately satisfied that the sum of its rows and columns is 1:

[0192] , .

[0193] The normalization formula is:

[0194] .

[0195] The calculation result is:

[0196] ;

[0197] .

[0198] Here, there is only one token, i.e. , row normalization can be performed; if there are other tokens that do not meet the minimum number of experts, such as e5, then , it is necessary to perform row and column normalization simultaneously to ensure that the total weight of each token is allocated to 1 (each row) and the proportion of each expert load is consistent (each column). This process makes the matrix Q approach a doubly stochastic matrix (1 for both rows and columns) to achieve the goal of a doubly stochastic matrix, which balances the load of experts among multiple fallback tokens. is the normalized score (i.e., normalized score).

[0199] Step 8: Select backup experts and normalize expert weights.

[0200] The original primary expert is e4, with a score of 0.774; the fallback calculation results are shown in step 7, and the top-1 is selected after sorting, i.e., selecting backup expert e1, and finally selecting experts {e4, e1}. Compared to Top-K which can only select a fixed K, the combination of the above two methods (i.e., confidence routing strategy and Sinkhorn routing selection strategy) provides more flexibility and fine-grained control granularity.

[0201] The scores of the selected two experts are: e4 = 0.774, e1 = 0.497.

[0202] Calculate the gating weight:

[0203] .

[0204] Step 9: Route expert output and perform backpropagation gradient update.

[0205] After selecting the above experts, the input token x t will be sent to the network of these experts for processing to obtain their respective outputs y i,t . The experts not selected will not be calculated. Then, the gating weight calculated above is used to weight and aggregate the output y i,t to obtain the final output y t of the MoE layer. The final output is updated through the loss function backpropagation to update the expert gating network parameters and expert network parameters. The above process forms a direct feedback mechanism that affects the weights of the final output, and the gradient signal will be received during backpropagation, which enables the gating network to learn how to make better routing and weight allocation decisions, thereby optimizing the performance of the entire MoE layer.

[0206] The formula is:

[0207] .

[0208] In summary, the key technical points of the embodiments of the present application include:

[0209] Confidence routing strategy: abandon fixed top-k, use threshold type confidence cumulative score to dynamically select experts, that is, prefer to select experts with high affinity to input tokens, accumulate their affinity scores, until a preset confidence threshold is reached, thereby determining a part of "high confidence" master experts, and improving the adaptability of decision-making.

[0210] Sinkhorn routing (Fallback) strategy: ensure that sparse experts can still be load balanced by global normalization distribution, that is, when the number of confidence routing experts is insufficient, use Sinkhorn normalization to provide a global balanced probability distribution for the remaining experts as a basis for selecting complementary experts, and obtain backup experts.

[0211] Expert deduplication and masking mechanism: ensure that fallback experts and confidence experts are complementary and do not conflict; ensure that experts selected by Sinkhorn do not repeat experts selected by confidence, avoiding redundant calculations and weight conflicts.

[0212] Weight balancing processing to ensure reverse trainability: normalize the final routing weights to ensure they are valid probability distributions and enable stable backpropagation of gradients.

[0213] The scheme combines confidence routing and Sinkhorn routing strategies, which can further improve routing quality and expert utilization rate, improve model performance, and thus improve the accuracy and reliability of the final output interview results while maintaining the lightweight characteristics of the DeepSeek-MoE architecture.

[0214] Corresponding to the above AI master-backup expert model-based interview method, the embodiments of the present application also provide an AI master-backup expert model-based interview device. Referring to Figure 7 the structural schematic diagram of an AI master-backup expert model-based interview device, the device comprises:

[0215] The feature acquisition module 701 is configured to acquire at least one interview feature information of a candidate object.

[0216] The score calculation module 702 is configured to calculate the affinity score between each interview feature information and a plurality of preset candidate experts.

[0217] The expert set determination module 703 is used to determine the expert set corresponding to each interview feature based on the affinity score of each candidate expert for each interview feature; wherein, the expert set includes the main experts determined based on the ranking of affinity scores and a preset cumulative confidence threshold; when the number of main experts is less than the preset minimum number of experts, the expert set also includes the backup experts determined by the double random matrix transformation of the Sinkhorn algorithm.

[0218] The result determination module 704 is used to calculate at least one interview feature by activating the expert set corresponding to each interview feature information to obtain the interview result of the candidate.

[0219] The interview device based on the AI ​​master-slave expert model provided in this embodiment of the invention achieves adaptive allocation of computing resources by driving dynamic expert activation through the cumulative confidence threshold. While ensuring the accuracy of the interview results output by the model, it improves computing efficiency and avoids unnecessary computing overhead. Furthermore, when the number of activated experts is insufficient, the Sinkhorn algorithm is used to transform the double random matrix to select the expert with the most global balance as the backup expert, ensuring balanced load distribution and avoiding expert idleness.

[0220] Further, the score calculation module 702 is specifically used for: for each interview feature information, calculating the original routing score of the interview feature information for each candidate expert based on the interview feature information and the expert gating vector of each candidate expert; and obtaining the affinity score of the interview feature information for each candidate expert by normalizing it with the Sigmoid function based on the original routing score of the interview feature information for each candidate expert.

[0221] Furthermore, the formula for calculating the original routing score is as follows:

[0222] ;

[0223] in, Indicates the first t The interview feature information for the first i The raw routing scores of each candidate expert. Indicates the first t Interview characteristics information, Indicates the first i The expert gating vector of each candidate expert. Indicates the first i Trainable biases for each candidate expert;

[0224] The formula for calculating the affinity score is:

[0225] ;

[0226] wherein, denotes the affinity score of the i-th interview feature information to the j-th candidate expert, t denotes the application of a Sigmoid function to z. i Further, the expert set determination module 703 is specifically configured to: for each of the interview feature information, sort the affinity scores of the interview feature information to each of the candidate experts in descending order to obtain a sorting result; accumulate the affinity scores according to the sorting result, and determine the candidate expert corresponding to the cumulative confidence threshold when the cumulative sum of the affinity scores reaches the cumulative confidence threshold, or the candidate expert corresponding to the maximum allowed number of experts when the number of candidate experts participating in the accumulation reaches the maximum allowed number of experts as the primary expert corresponding to the interview feature information; determine the interview feature information with a number of primary experts less than the minimum number of experts as simple feature information; obtain a normalized score matrix corresponding to the simple feature information by double random matrix transformation of the Sinkhorn algorithm in the case of shielding the primary expert, wherein the normalized score matrix comprises a normalized score of each of the candidate experts for the simple feature information; and determine a number of candidate experts with the maximum normalized scores in the normalized score matrix as backup experts corresponding to the simple feature information.

[0227] Further, the expert set determination module 703 is further configured to: sequentially take out the affinity scores according to the sorting result; determine whether the cumulative sum of the currently taken out affinity scores reaches the cumulative confidence threshold; if the cumulative confidence threshold is not reached, determine whether the number of candidate experts corresponding to the currently taken out affinity scores reaches the maximum allowed number of experts; if the maximum allowed number of experts is not reached, re-perform the step of sequentially taking out the affinity scores according to the sorting result; and if the cumulative confidence threshold is reached or the maximum allowed number of experts is reached, determine the candidate expert corresponding to the currently taken out affinity scores as the primary expert.

[0228] Further, the expert set determination module 703 is further configured to: obtain the original routing scores of each of the candidate experts for the simple feature information, and reset the original routing score corresponding to the primary expert to negative infinity; convert the original routing scores of each of the candidate experts for the simple feature information into an exponential transformation score matrix through exponential transformation; and convert the exponential transformation score matrix into the normalized score matrix through double random matrix transformation of the Sinkhorn algorithm.

[0229] Further, the expert set determination module 703 is further configured to: obtain the original routing scores of each of the candidate experts for the simple feature information, and reset the original routing score corresponding to the primary expert to negative infinity; convert the original routing scores of each of the candidate experts for the simple feature information into an exponential transformation score matrix through exponential transformation; and convert the exponential transformation score matrix into the normalized score matrix through double random matrix transformation of the Sinkhorn algorithm.

[0230] ​Further, the result determination module 704 is specifically used for: for each of the interview feature information, obtaining the gating weight of the interview feature information on each of the activated experts in the corresponding expert set through normalization processing based on the affinity score of the interview feature information on each activated expert; inputting the interview feature information into the corresponding expert set for processing to obtain the expert results output by each activated expert; performing weighted aggregation on the expert results output by each activated expert based on the gating weight of the interview feature information on each activated expert to obtain the target output result of the interview feature information; and generating the interview results of the candidate based on the target output results of each of the interview feature information.

[0231] The interview device based on the AI ​​master-slave expert model provided in this embodiment has the same implementation principle and technical effect as the aforementioned interview method based on the AI ​​master-slave expert model. For the sake of brevity, any parts not mentioned in the embodiment of the interview device based on the AI ​​master-slave expert model can be referred to the corresponding content in the aforementioned interview method embodiment based on the AI ​​master-slave expert model.

[0232] like Figure 8 As shown, an electronic device 800 provided in this embodiment of the invention includes: a processor 801, a memory 802 and a bus. The memory 802 stores a computer program that can run on the processor 801. When the electronic device 800 is running, the processor 801 and the memory 802 communicate through the bus, and the processor 801 executes the computer program to implement the above-mentioned interview method based on an AI master-slave expert model.

[0233] Specifically, the memory 802 and processor 801 mentioned above can be general-purpose memory and processor, without any specific limitations.

[0234] This invention also provides a computer-readable storage medium storing a computer program. When a processor runs this computer program, it executes the interview method based on an AI master-slave expert model described in the preceding method embodiments. The computer-readable storage medium includes various media capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), RAM, magnetic disk, or optical disk.

[0235] The term "and / or", merely describes association between associated objects, indicates that there can be three types of relationships, for example, A and / or B can indicate that A exists alone, A and B exist together, and B exists alone. In addition, the term "at least one" herein indicates any one of a plurality or at least any combination of two or more of a plurality, for example, at least one of A, B, and C includes any one or more elements selected from the set consisting of A, B, and C.

[0236] In all examples shown and described herein, any specific values should be interpreted as merely exemplary and not as a limitation, and thus other examples of the example embodiments can have different values.

[0237] The flowcharts and block diagrams in the drawings show the architectural, functional and operational aspects of possible implementations of apparatuses, methods and computer program products according to embodiments of the present application. In this regard, each block in the flowcharts or block diagrams can represent a module, a segment or a portion of code which includes one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession can in fact be executed substantially concurrently or the blocks can sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowcharts, and combinations thereof, can be implemented by dedicated hardware-based systems which perform the specified functions or acts, or can be implemented by a combination of dedicated hardware and computer instructions.

[0238] In several embodiments provided in the present application, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, and for example, a plurality of modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed ones can be indirect coupling or communication connection between the devices or modules through some communication interfaces, and can be electrical, mechanical or other forms.

[0239] The modules described as separate components can or can not be physically separated, and the components shown as modules can or can not be physical modules, i.e. they can be located in one place or distributed to a plurality of network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment.

[0240] In addition, each functional module in each embodiment of the present application can be integrated in one processing module, or each module can exist physically separately, or two or more modules can be integrated in one module.

[0241] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of each embodiment of the present application.

Claims

1. An interview method based on an AI master-slave expert model, characterized in that, include: Obtain at least one interview characteristic information of the candidate; The affinity score between each of the interview feature information and multiple preset candidate experts is calculated; Based on the affinity score of each of the interview feature information for each of the candidate experts, an expert set corresponding to each of the interview feature information is determined; wherein, the expert set includes the main experts determined based on the ranking of affinity scores and a preset cumulative confidence threshold; when the number of the main experts is less than the preset minimum number of experts, the expert set also includes the backup experts determined by the double random matrix transformation of the Sinkhorn algorithm. By activating the expert set corresponding to each of the interview feature information, the interview result of the candidate is obtained by calculating the at least one interview feature information; The step of determining the expert set corresponding to each interview feature based on the affinity score of each candidate expert for each interview feature includes: For each interview feature, the affinity scores of each candidate expert are sorted in descending order to obtain a sorting result; the affinity scores are accumulated according to the sorting result, and the candidate experts whose accumulated affinity scores reach the accumulated confidence threshold, or whose number of candidate experts participating in the accumulation reaches the preset maximum allowed number of experts, are determined as the main experts corresponding to the interview feature. The interview feature information where the number of lead experts is less than the minimum number of experts is identified as simple feature information. With lead experts masked, a normalized score matrix corresponding to the simple feature information is obtained through a double random matrix transformation using the Sinkhorn algorithm. The normalized score matrix includes the normalized score of the simple feature information for each candidate expert. The candidate experts with the largest normalized scores in the normalized score matrix are identified as backup experts corresponding to the simple feature information.

2. The interview method based on an AI master-slave expert model according to claim 1, characterized in that, The calculation yields an affinity score between each interview feature and multiple preset candidate experts, including: For each of the interview feature information, the original routing score of the interview feature information for each of the candidate experts is calculated based on the interview feature information and the expert gating vector of each candidate expert; Based on the original routing score of each candidate expert according to the interview feature information, the affinity score of the interview feature information for each candidate expert is obtained by normalization using the Sigmoid function.

3. The interview method based on an AI master-slave expert model according to claim 2, characterized in that, The formula for calculating the original routing score is as follows: ； in, Indicates the first t The interview feature information for the first i The raw routing scores of each candidate expert. Indicates the first t Interview characteristics information, Indicates the first i The expert gating vector of each candidate expert. Indicates the first i Trainable biases for each candidate expert; The formula for calculating the affinity score is: ； in, Indicates the first t The interview feature information for the first i The affinity score of each candidate expert. This indicates that the Sigmoid function is applied to z.

4. The interview method based on an AI master-slave expert model according to claim 1, characterized in that, The step of accumulating affinity scores based on the ranking results, and determining the candidate experts whose cumulative affinity scores reach the cumulative confidence threshold, or whose number of participating candidate experts reaches the preset maximum allowed number of experts, as the main experts, includes: Extract the affinity scores sequentially according to the sorting results; Determine whether the sum of the currently extracted affinity scores reaches the cumulative confidence threshold; If the cumulative confidence threshold is not reached, determine whether the number of candidate experts corresponding to the currently extracted affinity score has reached the maximum allowed number of experts; If the maximum allowed number of experts is not reached, repeat the step of retrieving affinity scores sequentially according to the sorting results; If the cumulative confidence threshold or the maximum number of allowed experts is reached, the candidate expert corresponding to the currently extracted affinity score will be determined as the master expert.

5. The interview method based on an AI master-slave expert model according to claim 1, characterized in that, The step of obtaining the normalized score matrix corresponding to the simple feature information through the double random matrix transformation of the Sinkhorn algorithm, under the condition of masking the master expert, includes: Obtain the original routing scores of each candidate expert based on the simple feature information, and reset the original routing score corresponding to the master expert to negative infinity; Through exponential transformation, the original routing scores of each candidate expert based on the simple feature information are transformed into an exponential transformation score matrix. The exponentially transformed fractional matrix is ​​transformed into the normalized fractional matrix through the double random matrix transformation of the Sinkhorn algorithm.

6. The interview method based on an AI master-slave expert model according to claim 1, characterized in that, The step of activating the expert set corresponding to each of the interview feature information, calculating the interview result of the candidate by activating the expert set corresponding to each of the interview feature information, includes: For each of the interview feature information, the gating weight of the interview feature information on each of the activated experts in the corresponding expert set is obtained by normalization based on the affinity score of the interview feature information on each activated expert. The interview feature information is input into the corresponding expert set for processing to obtain the expert results output by each activated expert; The expert results output by each of the activation experts are weighted and aggregated according to the gating weights of the interview feature information on each of the activation experts to obtain the target output result of the interview feature information; Based on the target output results of each of the interview feature information, the interview results of the candidate are generated.

7. An interview device based on an AI master-slave expert model, characterized in that, include: The feature acquisition module is used to acquire at least one interview feature information of the candidate; The score calculation module is used to calculate the affinity score between each of the interview feature information and multiple preset candidate experts; The expert set determination module is used to determine the expert set corresponding to each of the interview feature information based on the affinity score of each of the candidate experts; wherein, the expert set includes the main experts determined based on the ranking of affinity scores and a preset cumulative confidence threshold; when the number of the main experts is less than the preset minimum number of experts, the expert set also includes the backup experts determined by the double random matrix transformation of the Sinkhorn algorithm. The result determination module is used to calculate the interview result of the candidate by activating the expert set corresponding to each of the interview feature information; The expert set determination module is specifically used for: for each interview feature information, sorting the affinity scores of each candidate expert in descending order of the interview feature information to obtain a sorting result; accumulating the affinity scores according to the sorting result, and determining the candidate experts whose accumulated affinity scores reach the accumulated confidence threshold, or whose number of participating candidate experts reaches the preset maximum allowed number of experts, as the main experts corresponding to the interview feature information; and determining the interview feature information with a number of main experts less than the minimum number of experts as simple feature information. With the primary expert masked, a normalized score matrix corresponding to the simple feature information is obtained through the double random matrix transformation of the Sinkhorn algorithm. The normalized score matrix includes the normalized score of the simple feature information for each candidate expert. The candidate experts with the largest normalized scores in the normalized score matrix are determined as the backup experts corresponding to the simple feature information.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the interview method based on the AI ​​master-slave expert model as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when executed by the processor, performs the interview method based on the AI ​​master-slave expert model as described in any one of claims 1-6.

Images