An interview guidance method and system based on big model and artificial intelligence

Abstract

The present invention proposes an interview guidance method and system based on a large model and artificial intelligence, the method comprising: collecting interview guidance data; performing preprocessing, using a large model to convert the data into a high-dimensional vector, using a machine learning model to extract semantic features and time series features according to semantic information, and extracting emotional features; constructing a dynamic knowledge graph database according to semantic features, time series features and emotional features; using natural language processing technology to extract key features and construct an interview guidance model; constructing a deep learning interview score prediction model to predict interview scores according to a recommended user data set; designing a user privacy mechanism; designing a simulated interviewer model according to a recommended user data set, and obtaining a result to form a report output to the user. The present invention integrates a large language model, natural language processing, machine learning technology and privacy protection technology, and uses a number of advanced technologies and design concepts to improve the success rate of user interviews.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and in particular relates to an interview guidance method and system based on a large model and artificial intelligence. Background Art

[0002] Although interview coaching methods and systems based on large language models and artificial intelligence have shown significant potential and advantages in the field of interview preparation, the technology currently has some defects and challenges that limit its full functionality and optimization of user experience. Although large language models can provide reasonable responses in many situations, they may be biased or inaccurate in understanding complex, field-specific professional questions or situations with large cultural background differences. Existing AI systems have difficulty accurately identifying and responding to emotionally charged communications involved in interviews, such as reading the interviewer's non-verbal cues and emotional reactions. The interview preparation suggestions and simulations provided by the system may be too generalized and lack in-depth consideration of each user's unique background, skills, and career goals, making it difficult to achieve full personalization. In the face of user feedback and progress, existing systems may not be able to flexibly adjust strategies and suggestions in real time to achieve the best coaching results.

[0003] The training of AI systems relies on a large amount of data. If the data set is biased, the system may also give biased recommendations, such as unfair treatment of certain groups. Uploading personal resumes and sensitive information to participate in interview simulations may cause users to worry about data security and privacy leaks.

[0004] Therefore, it is urgent to study an intelligent and safe interview guidance method to solve the above problems. Summary of the invention

[0005] The purpose of this invention is to propose an interview guidance method based on big models and artificial intelligence, which integrates big language models, natural language processing, machine learning technology and privacy protection technology, and comprehensively applies a number of advanced technologies and design concepts to provide users with comprehensive and personalized interview preparation support and improve the success rate of interviews.

[0006] In order to achieve the above object, the present invention provides an interview guidance method based on a large model and artificial intelligence, the method comprising:

[0007] S1. Collecting interview guidance data; wherein the interview guidance data includes at least two of the following: interview questions, job seeker's answers, interviewer's evaluation, interview scenario, job information and job seeker's background;

[0008] S2. Preprocess the interview guidance data, use a large model to convert the data into a high-dimensional vector, capture semantic information, use a machine learning model to extract semantic features and time series features based on the semantic information, perform sentiment annotation on the semantic information, and extract sentiment features;

[0009] S3. Based on semantic features, time series features and sentiment features, a dynamic knowledge graph database is constructed in combination with historical interview guidance data;

[0010] S4. Extract key features using natural language processing technology based on the resume and target position uploaded by the user in real time, use the graph database query language to query the dynamic knowledge graph database, and build an interview guidance model. Use the interview guidance model to match the user profile with the position requirements to obtain a recommended user data set.

[0011] S5. Build an interview score prediction model based on deep learning to predict interview scores based on the recommended user dataset;

[0012] S6. Design user privacy mechanisms to ensure secure storage and transmission of user information;

[0013] S7. Design a simulated interviewer model based on the recommended user data set, interact with the user, evaluate the user's answer and calculate the actual interview score, dynamically adjust the recommended user data set based on the predicted interview score and the actual interview score, and output the recommended user data set to the user in the form of a report;

[0014] Wherein, the step S3 specifically includes:

[0015] S301. Define the knowledge graph as G = {S, K, E}, where S represents a feature triple, K represents an entity, and E represents a relationship matrix;

[0016] S302, the feature triple is defined as S={<c,t,p> |c, t∈K, p∈E}, where c represents the head entity, t represents the tail entity, and p represents the relationship;

[0017] S303, mapping the entity K embedding representation matrix to a d-dimensional embedding space; wherein, both the head entity and the tail entity are mapped to the d-dimensional embedding space, and a d-dimensional mapping head entity c is obtained. c 、c t and d-dimensional mapping tail entity t c ,t t , and c c ,t c ∈R d , R represents an integer;

[0018] S304, map the relationship matrix E to the d*d dimensional embedding space, convert the relationship vector into a relationship matrix, and obtain the relationship matrix p-1 ∈R d×d ;

[0019] S305. Design of bilinear score S core (c, t, p), expressed as follows:

[0020] S core (c,t,p)=α·f(c c ,p,t t )+β·f(t c , p -1 , c t )

[0021] Where f() represents a bilinear score function based on a single component, α and β represent weight factors respectively;

[0022] S306, construct a loss function to train the scores of feature triples to minimize the loss function while avoiding overfitting; the loss function L γ It is expressed as follows:

[0023] L γ =∑log(1-exp(-l*clamp(S core (c,t,p),δ)))

[0024] Among them, clamp represents the clamp function, which is used to control the range within (S core (c, t, p), δ), δ represents the cropping factor, and l represents the label.

[0025] Furthermore, the large model is used to convert the data into a high-dimensional vector to capture semantic information; wherein the large model is a pre-trained Transformer model; the specific steps are as follows:

[0026] Input data sequence X = {x1, x2, ..x i ., x n}, where x i is a word or subword in the data;

[0027] The data sequence is input into the encoding layer and represented as follows:

[0028] E 0 =W scr X+P i

[0029] A t =Layer(E t-1 +Head(E t-1 , E t-1 , E t-1 ))

[0030] E t =Layer(A t +FEN(A t ))

[0031] Among them, E 0 represents the output of the 0th layer encoder, W scr Represents the word embedding matrix of the source data, P i represents the position encoding corresponding to the source data, Head represents the multi-head self-attention function, E t-1 represents the output of the t-1th layer encoder, A t represents the output of multi-head self-attention, E t represents the output of the t-th layer encoder, FEN represents the feedforward neural function, and Layer represents the normalization function;

[0032] The output E of the t-th layer encoder t The input is passed through the encoder to extract the semantic vector, which is expressed as follows:

[0033]

[0034]

[0035]

[0036]

[0037] Among them, D t represents the semantic feature vector, Represents the word embedding matrix of the target data, D 0 represents the output vector of the encoder at layer 0, X hot Represents the one-hot matrix corresponding to the target data, represents the output vector of the multi-head self-attention of the t-th layer decryptor, D t-1 represents the output vector of the encoder at layer t-1, MASK represents the multi-head attention function with occlusion mechanism in the decoder, represents the output vector of the multi-head attention with occlusion mechanism in the t-th layer decryptor-decoder, E n Represents the output of the nth layer encoder.

[0038] Furthermore, the machine learning model is used to extract semantic features and time series features according to the semantic information, sentiment annotation is performed on the semantic information, and sentiment features are extracted, as follows:

[0039] Introducing time series features X′ t , according to the encoder output D t , train a machine learning model to classify the sentiment of semantic features, which is expressed as follows:

[0040] Classifier(D t )→y

[0041] Among them, y represents the sentiment label;

[0042] The time series feature X′ t Combined with the semantic sentiment features of the corresponding moment, a fusion feature vector F is formed c , which is expressed as follows:

[0043] F c =Concatenate(X′ t , SentimentScore t )

[0044] Among them, Concatenate represents the feature fusion function, SentimentScore t represents the sentiment score;

[0045] Define a loss function to measure the gap between the model prediction and the actual label, and use gradient descent to minimize the loss function to update the model parameters; wherein the loss function is expressed as follows:

[0046]

[0047] in, Represents the actual sentiment score.

[0048] Furthermore, the dynamic knowledge graph database is constructed as follows:

[0049] Define the entities, relationships and attributes of the knowledge graph; wherein the entities include questions, answers, sentiment states, and time series features;

[0050] Choose Neo4j to build a knowledge graph database, import entities and as nodes, relationships as edges, and attributes into the knowledge graph database;

[0051] Design an update mechanism for the dynamic knowledge graph to update the knowledge graph in real time or regularly based on new interview data.

[0052] Furthermore, the update mechanism of the dynamic knowledge graph is expressed as follows:

[0053] Construct a dynamic knowledge graph G = (V, E), where V is the set of entity nodes and E is the set of relationship edges;

[0054] Let D be the newly arrived interview dataset, each data item d∈D can be an interview record, including semantic features S(d), time series features T(d), and emotional features Emo(d);

[0055] For each d, identify the corresponding entity e∈E, let f_match(e, d) be the matching function, and return the identifier of the existing entity or the newly created entity related to d;

[0056] For entity e and its attribute set A, let update_attribute(A, feature) be the update function, which updates the entity attributes according to the features of the new data d;

[0057] If each data item d introduces a new relational attribute, or if a new question is answered and a new answer is obtained, an edge r∈R needs to be created. Let create_relation(e1, e2, r) be the edge creation function, where e1, e2 are entities and r is the relation type.

[0058] Furthermore, the interview guidance model is constructed as follows:

[0059] Assume that user resume is represented as a document set R = {r1, r2…, r n}, where each resume r n is a string consisting of a series of words or phrases. The target job description is represented by a document set J = {j1, j2, ..., j m}, each job descriptionj m Also made up of words or phrases;

[0060] Construct an interview guidance model for matching, as shown below:

[0061]

[0062] Among them, M represents the interview guidance model, S nm Resume n With position j m The matching score of A feature vector representing a user's resume, The feature vector representing the job description, N nm According to the resume n and position j m The relevant node subsets and their relationships in the graph obtained by feature query;

[0063] According to the matching score S nm , sort all the combinations of resumes and positions, and select the combinations with the highest matching degree as the recommended user data set:

[0064]

[0065] Among them, θ is the set matching threshold.

[0066] Further, the S5, constructing an interview score prediction model based on deep learning to predict the interview score according to the recommended user data set, specifically includes:

[0067] The interview score prediction model is constructed as follows:

[0068] Let the input feature vector be s∈R n , represents the characteristics of each applicant; the actual interview score is y∈R;

[0069] Input layer to hidden layer weight W1∈R H×n , the bias term b1∈R H ; Hidden layer to output layer weight W2∈1×H, bias term b2∈R; where H represents the number of neurons in the hidden layer;

[0070] Construct forward propagation; wherein the forward propagation includes the input layer to the hidden layer and the hidden layer to the output layer, and the input layer to the hidden layer is expressed as follows:

[0071] Z1=σ(sW1+b1)

[0072] h=σ(Z)

[0073] Wherein, Z1 represents the weighted sum of neurons from the input layer to the hidden layer, σ represents the activation function, and h represents the output of the hidden layer; the hidden layer to the output layer is represented as follows:

[0074] Z2=hW2+b2

[0075]

[0076] Among them, Z2 represents the weighted sum from the hidden layer to the output layer, Represents the predicted output of the output layer;

[0077] Using the cross entropy loss function, it is expressed as follows:

[0078]

[0079] Among them, N represents the total number of samples, C represents the total number of categories, and y i represents the real sample;

[0080] Use back propagation to calculate the gradient and update the weights;

[0081] Taking multiple initial interview score prediction models as individuals of the population, using the accuracy of interview score prediction as the objective function, and using the sparrow search algorithm to perform iterative search to optimize and adjust the parameters of the interview score prediction model; wherein the sparrow search algorithm is expressed as follows:

[0082]

[0083] Among them, ε n+1 represents the updated interview score prediction model variable, ε n represents the interview score prediction model variable, μ represents the control parameter; wherein, the interview score prediction model variables are weights and bias terms.

[0084] Furthermore, the user privacy mechanism is designed to specifically include:

[0085] Let the plaintext be S m , belongs to the plaintext space m, and the key is S k , belongs to the key space k;

[0086] The encryption function is Φ, which maps the elements of the plaintext space m and the key space k to the ciphertext space Θ;

[0087] The encryption process is expressed as:

[0088]

[0089] Among them, Ψ represents the ciphertext, belongs to the ciphertext space Θ, and represents the plaintext S m Through the encryption function Φ and the key S k Converted from

[0090] The decrypted representation is:

[0091]

[0092] in, Used to ensure that the encrypted ciphertext Ψ can be encrypted with the same key S k Restore to the original plaintext S k .

[0093] The second invention of the present invention provides an interview guidance system based on a large model and artificial intelligence, the system comprising:

[0094] A data collection module, used to collect interview guidance data; wherein the interview guidance data includes at least two of the following: interview questions, job seeker's answers, interviewer's evaluation, interview scenario, job information and job seeker's background;

[0095] Feature extraction module, used to pre-process the interview guidance data, use a large model to convert the data into a high-dimensional vector, capture semantic information, use a machine learning model to extract semantic features and time series features based on the semantic information, perform sentiment annotation on the semantic information, and extract sentiment features;

[0096] The knowledge graph construction module is used to build a dynamic knowledge graph database based on semantic features, time series features, and sentiment features combined with historical interview guidance data;

[0097] The data matching module is used to extract key features based on the resume and target position uploaded by the user in real time using natural language processing technology, and to query the dynamic knowledge graph database using the graph database query language. At the same time, an interview guidance model is built and used to match the user profile with the position requirements to obtain a recommended user data set.

[0098] Score prediction module, which is used to build an interview score prediction model based on deep learning to predict interview scores according to the recommended user dataset;

[0099] Privacy design module, used to design user privacy mechanisms to ensure the secure storage and transmission of user information;

[0100] The report output module is used to design a simulated interviewer model based on the recommended user data set, interact with the user, evaluate the user's answer and calculate the actual interview score, dynamically adjust the recommended user data set according to the predicted interview score and the actual interview score, and output the recommended user data set to the user in the form of a report;

[0101] The steps performed by the knowledge graph construction module specifically include:

[0102] S301. Define the knowledge graph as G = {S, K, E}, where S represents a feature triple, K represents an entity, and E represents a relationship matrix;

[0103] S302, the feature triple is defined as S={<c,t,p> |c, t∈K, p∈E}, where c represents the head entity, t represents the tail entity, and p represents the relationship;

[0104] S303, mapping the entity K embedding representation matrix to a d-dimensional embedding space; wherein, both the head entity and the tail entity are mapped to the d-dimensional embedding space, and a d-dimensional mapping head entity c is obtained. c 、c t and d-dimensional mapping tail entity t c ,t t , and c c ,t c ∈R d , R represents an integer;

[0105] S304, map the relationship matrix E to the d*d dimensional embedding space, convert the relationship vector into a relationship matrix, and obtain the relationship matrix p -1 ∈R d×d ;

[0106] S305. Design of bilinear score S core (c, t, p), expressed as follows:

[0107] S core(c, t, p) = α·f(c c ,p,t t )+β·f(t c , p -1 , c t )

[0108] Where f() represents a bilinear score function based on a single component, α and β represent weight factors respectively;

[0109] S306, construct a loss function to train the scores of feature triples to minimize the loss function while avoiding overfitting; the loss function L γ It is expressed as follows:

[0110] L γ =∑log(1-exp(-l*clamp(S core (c, t, p), δ)))

[0111] Among them, clamp represents the clamp function, which is used to control the range within (S core (c, t, p), δ), δ represents the cropping factor, and l represents the label.

[0112] The beneficial technical effects of the present invention are at least as follows:

[0113] (1) The present invention analyzes the user's resume, target position requirements, and historical interview data to provide each user with a customized interview preparation strategy. This personalized guidance helps users focus on improving the abilities and knowledge they need most in a specific position interview, thereby improving preparation efficiency.

[0114] (2) The present invention uses a large model to generate a variety of interview questions, covering multiple dimensions such as technology, behavior, and case analysis, ensuring that users can easily handle different types of interview questions and enhance their ability to deal with unknown problems.

[0115] (3) As the user practices more, the present invention can dynamically adjust the learning plan and difficulty according to the user's progress, ensuring continuous challenges and promoting growth. At the same time, the model itself will be continuously optimized according to user feedback and usage to improve service quality. Compared with traditional face-to-face training or purchasing expensive training materials, AI-based interview guidance is more flexible and convenient. Users can practice according to their own schedule, which greatly saves time and money.

[0116] (4) When processing user data, the present invention adopts advanced encryption technology and anonymous processing means to ensure the security of user information and increase user trust. BRIEF DESCRIPTION OF THE DRAWINGS

[0117] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0118] Figure 1 The present invention is a flow chart of an interview guidance method based on a large model and artificial intelligence. DETAILED DESCRIPTION

[0119] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0120] like Figure 1 As shown, an interview guidance method based on a large model and artificial intelligence is provided in an embodiment of the present invention, comprising:

[0121] S1. Collecting interview guidance data; wherein the interview guidance data includes at least two of the following: interview questions, job seeker's answers, interviewer's evaluation, interview scenario, job information and job seeker's background;

[0122] S2. Preprocess the interview guidance data, use a large model to convert the data into a high-dimensional vector, capture semantic information, use a machine learning model to extract semantic features and time series features based on the semantic information, perform sentiment annotation on the semantic information, and extract sentiment features;

[0123] S3. Based on semantic features, time series features and sentiment features, a dynamic knowledge graph database is constructed in combination with historical interview guidance data;

[0124] S4. Extract key features using natural language processing technology based on the resume and target position uploaded by the user in real time, use the graph database query language to query the dynamic knowledge graph database, and build an interview guidance model. Use the interview guidance model to match the user profile with the position requirements to obtain a recommended user data set.

[0125] S5. Build an interview score prediction model based on deep learning to predict interview scores based on the recommended user dataset;

[0126] S6. Design user privacy mechanisms to ensure secure storage and transmission of user information;

[0127] S7. Design a simulated interviewer model based on the recommended user data set, interact with the user, evaluate the user's answer and calculate the actual interview score, dynamically adjust the recommended user data set based on the predicted interview score and the actual interview score, and output the recommended user data set to the user in the form of a report;

[0128] Wherein, the step S3 specifically includes:

[0129] S301. Define the knowledge graph as G = {S, K, E}, where S represents a feature triple, K represents an entity, and E represents a relationship matrix;

[0130] S302, the feature triple is defined as S={<c,t,p> |c, t∈K, p∈E}, where c represents the head entity, t represents the tail entity, and p represents the relationship;

[0131] S303, mapping the entity K embedding representation matrix to a d-dimensional embedding space; wherein, both the head entity and the tail entity are mapped to the d-dimensional embedding space, and a d-dimensional mapping head entity c is obtained. c 、c t and d-dimensional mapping tail entity t c ,t t , and c c ,t c ∈R d , R represents an integer;

[0132] S304, map the relationship matrix E to the d*d dimensional embedding space, convert the relationship vector into a relationship matrix, and obtain the relationship matrix p -1 ∈R d×d ;

[0133] S305. Design of bilinear score S core (c, t, p), expressed as follows:

[0134] S core (c,t,p)=α·f(c c ,p,t t )+β·f(t c , p -1 , c t )

[0135] Where f() represents a bilinear score function based on a single component, α and β represent weight factors respectively;

[0136] S306, construct a loss function to train the scores of feature triples to minimize the loss function while avoiding overfitting; the loss function L γ It is expressed as follows:

[0137]

[0138] Among them, clamp represents the clamp function, which is used to control the range within (S core (c, t, p), δ), δ represents the cropping factor, and l represents the label.

[0139] Furthermore, the large model is used to convert the data into a high-dimensional vector to capture semantic information; wherein the large model is a pre-trained Transformer model; the specific steps are as follows:

[0140] Input data sequence X = {x1, x2, ..x i .,x n}, where x i is a word or subword in the data;

[0141] The data sequence is input into the encoding layer and represented as follows:

[0142] E 0 =W scr X+P i

[0143] A t =Layer(E t-1 +Head(E t-1 , E t-1 , E t-1 ))

[0144] E t =Layer(A t +FEN(A t ))

[0145] Among them, E 0 represents the output of the 0th layer encoder, W scr Represents the word embedding matrix of the source data, P i represents the position encoding corresponding to the source data, Head represents the multi-head self-attention function, E t-1 represents the output of the t-1th layer encoder, A t represents the output of multi-head self-attention, E t represents the output of the t-th layer encoder, FEN represents the feedforward neural function, and Layer represents the normalization function;

[0146] The output E of the t-th layer encoder t The input is passed through the encoder to extract the semantic vector, which is expressed as follows:

[0147]

[0148]

[0149]

[0150]

[0151] Among them, D t represents the semantic feature vector, Represents the word embedding matrix of the target data, D 0 represents the output vector of the encoder at layer 0, X hot Represents the one-hot matrix corresponding to the target data, represents the output vector of the multi-head self-attention of the t-th layer decryptor, D t-1 represents the output vector of the encoder at layer t-1, MASK represents the multi-head attention function with occlusion mechanism in the decoder, represents the output vector of the multi-head attention with occlusion mechanism in the t-th layer decryptor-decoder, E n Represents the output of the nth layer encoder.

[0152] Furthermore, the machine learning model is used to extract semantic features and time series features according to the semantic information, sentiment annotation is performed on the semantic information, and sentiment features are extracted, as follows:

[0153] Introducing time series features X′ t , according to the encoder output D t , train a machine learning model to classify the sentiment of semantic features, which is expressed as follows:

[0154] Classifier(D t )→y

[0155] Among them, y represents the sentiment label;

[0156] The time series feature X′ t Combined with the semantic sentiment features of the corresponding moment, a fusion feature vector F is formed c , which is expressed as follows:

[0157] F c =Concatenate(X′ t , SentimentScore t )

[0158] Among them, Concatenate represents the feature fusion function, SentimentScore t represents the sentiment score;

[0159] Define a loss function to measure the gap between the model prediction and the actual label, and use gradient descent to minimize the loss function to update the model parameters; wherein the loss function is expressed as follows:

[0160]

[0161] in, Represents the actual sentiment score.

[0162] Furthermore, the dynamic knowledge graph database is constructed as follows:

[0163] Define the entities, relationships and attributes of the knowledge graph; wherein the entities include questions, answers, sentiment states, and time series features;

[0164] Choose Neo4j to build a knowledge graph database, import entities and as nodes, relationships as edges, and attributes into the knowledge graph database;

[0165] Design an update mechanism for the dynamic knowledge graph to update the knowledge graph in real time or regularly based on new interview data.

[0166] Furthermore, the update mechanism of the dynamic knowledge graph is expressed as follows:

[0167] Construct a dynamic knowledge graph G = (V, E), where V is the set of entity nodes and E is the set of relationship edges;

[0168] Let D be the newly arrived interview dataset, each data item d∈D can be an interview record, including semantic features S(d), time series features T(d), and emotional features Emo(d);

[0169] For each d, identify the corresponding entity e∈E, let f match(e, d) be the matching function, and return the identifier of the existing entity or the newly created entity related to d;

[0170] For entity e and its attribute set A, let update_attribute(A, feature) be the update function, which updates the entity attributes according to the features of the new data d;

[0171] If each data item d introduces a new relational attribute, or if a new question is answered and a new answer is obtained, an edge r∈R needs to be created. Let create_relation(e1, e2, r) be the edge creation function, where e1, e2 are entities and r is the relation type.

[0172] Furthermore, the interview guidance model is constructed as follows:

[0173] Assume that user resume is represented as a document set R = {r1, r2…, r n}, where each resume r n is a string consisting of a series of words or phrases. The target job description is represented by a document set J = {j1, j2, ..., j m}, each job descriptionj mAlso made up of words or phrases;

[0174] Construct an interview guidance model for matching, as shown below:

[0175]

[0176] Among them, M represents the interview guidance model, S nm Resume n With position j m The matching score of A feature vector representing a user's resume, The feature vector representing the job description, N nm According to the resume n and position j m The relevant node subsets and their relationships in the graph obtained by feature query;

[0177] According to the matching score S nm , sort all the combinations of resumes and positions, and select the combinations with the highest matching degree as the recommended user data set:

[0178]

[0179] Among them, θ is the set matching threshold.

[0180] Further, the S5, constructing an interview score prediction model based on deep learning to predict the interview score according to the recommended user data set, specifically includes:

[0181] The interview score prediction model is constructed as follows:

[0182] Let the input feature vector be s∈R n , represents the characteristics of each applicant; the actual interview score is y∈R;

[0183] Input layer to hidden layer weight W1∈R H×n , the bias term b1∈R H ; Hidden layer to output layer weight W2∈1×H, bias term b2∈R; where H represents the number of neurons in the hidden layer;

[0184] Construct forward propagation; wherein the forward propagation includes the input layer to the hidden layer and the hidden layer to the output layer, and the input layer to the hidden layer is expressed as follows:

[0185] Z1=σ(sW1+b1)

[0186]

[0187] Among them, Z1 represents the weighted sum of neurons from the input layer to the hidden layer, σ represents the activation function, and h represents the hidden layer output;

[0188] The hidden layer to the output layer is represented as follows:

[0189] Z2=hW2+b2

[0190]

[0191] Among them, Z2 represents the weighted sum from the hidden layer to the output layer, Represents the predicted output of the output layer;

[0192] Using the cross entropy loss function, it is expressed as follows:

[0193]

[0194] Among them, N represents the total number of samples, C represents the total number of categories, and y i represents the real sample;

[0195] Back propagation is used to calculate the gradient and update the weights, which is expressed as follows:

[0196]

[0197]

[0198]

[0199]

[0200] b1=sum(δ 1 ,axis=0)

[0201] Among them, axis operates by column, sum represents the accumulation operation, and δ 1 Represents the partial derivative from the input layer to the hidden layer;

[0202] Among them, the partial derivative can be calculated as follows:

[0203]

[0204] Among them, f′(Z l ) represents the derivative of the activation function at point Z1;

[0205] Taking multiple initial interview score prediction models as individuals of the population, using the accuracy of interview score prediction as the objective function, and using the sparrow search algorithm to perform iterative search to optimize and adjust the parameters of the interview score prediction model; wherein the sparrow search algorithm is expressed as follows:

[0206]

[0207] Among them, ε n+1represents the updated interview score prediction model variable, ε n represents the interview score prediction model variable, μ represents the control parameter; wherein, the interview score prediction model variables are weights and bias terms.

[0208] Furthermore, the user privacy mechanism is designed to specifically include:

[0209] Let the plaintext be S m , belongs to the plaintext space m, and the key is S k , belongs to the key space k;

[0210] The encryption function is Φ, which maps the elements of the plaintext space m and the key space k to the ciphertext space Θ;

[0211] The encryption process is expressed as:

[0212]

[0213] Among them, Ψ represents the ciphertext, belongs to the ciphertext space Θ, and represents the plaintext S m Through the encryption function Φ and the key S k Converted from

[0214] The decrypted representation is:

[0215]

[0216] in, Used to ensure that the encrypted ciphertext Ψ can be encrypted with the same key S k Restore to the original plaintext S k .

[0217] Further, S7, designing a simulated interviewer model based on the recommended user data set, interacting with the user, evaluating the user's answer to calculate the actual interview score, dynamically adjusting the recommended user data set based on the predicted interview score and the actual interview score, and outputting the recommended user data set to the user in the form of a report, including the following steps:

[0218] Collect rich interview guidance materials, including historical interview records (such as questions, answers, scoring criteria, and final scores), user background information (such as education background and work experience), and interviewer evaluations and feedback. Remove invalid or incomplete records, handle missing values, and standardize text data to ensure data quality. Extract useful features, such as text sentiment analysis, keyword extraction, and user skill tags, to prepare for model training.

[0219] Develop mock interviewers that include the following:

[0220] Use NLP techniques to understand user responses, including semantic understanding, sentiment analysis, and response relevance assessment. Build models to assess response quality, which may include classification models (to determine whether responses are relevant and of high quality), regression models (to predict interview scores), or sequence-to-sequence models (to generate feedback suggestions). Design a conversation flow so that the mock interviewer can ask appropriate questions based on the progress of the interview, simulating a real interview environment.

[0221] Specifically, a user-friendly interface is developed to allow users to participate in mock interviews, receive and submit responses. The model instantly analyzes the user's responses, calculates scores based on pre-set scoring criteria, and provides instant feedback.

[0222] Optionally, an emotion recognition function is integrated to analyze the tone and speed of the user's answer and give emotion management suggestions.

[0223] Based on the difference between the user's actual interview performance and the simulated interview score, we collect feedback and analyze the model prediction deviation. Based on the analysis results, we dynamically adjust the practice questions and learning resources recommended to the user to strengthen the user's weak links. We continue to iterate the model and optimize the algorithm to improve the evaluation accuracy and efficiency of personalized recommendations.

[0224] Regularly or after the mock interview, automatically generate personalized reports summarizing user performance, including strengths, areas for improvement, and recommended learning paths and resources. At the same time, establish a mechanism to collect user feedback on the mock interview experience to further improve the model and services.

[0225] In addition, an embodiment of the present application also provides an interview guidance system based on a large model and artificial intelligence, the system comprising:

[0226] The data collection module 101 is used to collect interview guidance data; wherein the interview guidance data includes at least two of the following: interview questions, job seeker's answers, interviewer's evaluation, interview scenario, job information and job seeker's background;

[0227] The feature extraction module 102 is used to pre-process the interview guidance data, convert the data into a high-dimensional vector using a large model, capture semantic information, extract semantic features and time series features based on the semantic information using a machine learning model, perform sentiment annotation on the semantic information, and extract sentiment features;

[0228] The knowledge graph construction module 103 is used to construct a dynamic knowledge graph database based on semantic features, time series features and sentiment features in combination with historical interview guidance data;

[0229] The data matching module 104 is used to extract key features using natural language processing technology based on the resume and target position uploaded by the user in real time, query the dynamic knowledge graph database using the graph database query language, and build an interview guidance model. The interview guidance model is used to match the user profile with the position requirements to obtain a recommended user data set;

[0230] The score prediction module 105 is used to construct an interview score prediction model based on deep learning to predict the interview score according to the recommended user data set;

[0231] The privacy design module 106 is used to design a user privacy mechanism to ensure the secure storage and transmission of user information;

[0232] A report output module 107 is used to design a simulated interviewer model based on the recommended user data set, interact with the user, evaluate the user's answer and calculate the actual interview score, dynamically adjust the recommended user data set according to the predicted interview score and the actual interview score, and output the recommended user data set to the user in the form of a report;

[0233] The steps performed by the knowledge graph construction module 103 specifically include:

[0234] S301. Define the knowledge graph as G = {S, K, E}, where S represents a feature triple, K represents an entity, and E represents a relationship matrix;

[0235] S302, the feature triple is defined as S={<c,t,p> |c, t∈K, p∈E}, where c represents the head entity, t represents the tail entity, and p represents the relationship;

[0236] S303, mapping the entity K embedding representation matrix to a d-dimensional embedding space; wherein, both the head entity and the tail entity are mapped to the d-dimensional embedding space, and a d-dimensional mapping head entity c is obtained. c 、c t and d-dimensional mapping tail entity t c ,t t , and c c ,t c ∈R d , R represents an integer;

[0237] S304, map the relationship matrix E to the d*d dimensional embedding space, convert the relationship vector into a relationship matrix, and obtain the relationship matrix p -1 ∈R d×d ;

[0238] S305. Design of bilinear score S core (c, t, p), expressed as follows:

[0239] S core(c,t,p)=α·f(c c ,p,t t )+β·f(t c , p -1 , c t )

[0240] Where f() represents a bilinear score function based on a single component, α and β represent weight factors respectively;

[0241] S306, construct a loss function to train the scores of feature triples to minimize the loss function while avoiding overfitting; the loss function L γ It is expressed as follows:

[0242] L γ =∑log(1-exp(-l*clamp(S core (c,t,p),δ)))

[0243] Among them, clamp represents the clamp function, which is used to control the range within (S core (c, t, p), δ), δ represents the cropping factor, and l represents the label.

[0244] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0245] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a division of logical functions. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the device or unit can be electrical, mechanical or other forms.

[0246] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0247] Although embodiments of the present invention have been shown and described, those skilled in the art will appreciate that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. An interview guidance method based on a large model and artificial intelligence, characterized in that: The method comprises: S1. Collecting interview guidance data; wherein the interview guidance data includes at least two of the following: interview questions, job seeker's answers, interviewer's evaluation, interview scenario, job information and job seeker's background; S2. Preprocess the interview guidance data, use a large model to convert the data into a high-dimensional vector, capture semantic information, use a machine learning model to extract semantic features and time series features based on the semantic information, perform sentiment annotation on the semantic information, and extract sentiment features; S3. Based on semantic features, time series features and sentiment features, a dynamic knowledge graph database is constructed in combination with historical interview guidance data; S4. Extract key features using natural language processing technology based on the resume and target position uploaded by the user in real time, use the graph database query language to query the dynamic knowledge graph database, and build an interview guidance model. Use the interview guidance model to match the user profile with the position requirements to obtain a recommended user data set. S5. Build an interview score prediction model based on deep learning to predict interview scores based on the recommended user dataset; S6. Design user privacy mechanisms to ensure secure storage and transmission of user information; S7. Design a simulated interviewer model based on the recommended user data set, interact with the user, evaluate the user's answer and calculate the actual interview score, dynamically adjust the recommended user data set based on the predicted interview score and the actual interview score, and output the recommended user data set to the user in the form of a report; Wherein, the step S3 specifically includes: S301. Define the knowledge graph as G = {S, K, E}, where S represents a feature triple, K represents an entity, and E represents a relationship matrix; S302, the feature triple is defined as S={<c,t,p> |c, t∈K, p∈E}, where c represents the head entity, t represents the tail entity, and p represents the relationship; S303, mapping the entity K embedding representation matrix to a d-dimensional embedding space; wherein, both the head entity and the tail entity are mapped to the d-dimensional embedding space, and a d-dimensional mapping head entity c is obtained. c 、c t and d-dimensional mapping tail entity t c ,t t , and c c ,t c ∈R d , R represents an integer; S304, map the relationship matrix E to the d*d dimensional embedding space, convert the relationship vector into a relationship matrix, and obtain the relationship matrix p -1 ∈R d×d ; S305. Design of bilinear score S core (c, t, p), expressed as follows: S core (c,t,p)=α·f(c c ,p,t t )+β·f(t c ,p -1 ,c t ) Where f() represents a bilinear score function based on a single component, α and β represent weight factors respectively; S306, construct a loss function to train the scores of feature triples to minimize the loss function while avoiding overfitting; the loss function L γ It is expressed as follows: L γ =∑log(1-exp(-l*clamp(S core (c,t,p),δ))) Among them, clamp represents the clamp function, which is used to control the range within (S core (c, t, p), δ); δ represents the cropping factor, l represents the label; The interview guidance model is constructed as follows: Assume that user resume is represented as document set R = {r1,r2...,r n }, where each resume r n is a string consisting of a series of words or phrases. The target job description is represented by a document set J = {j1, j2, ..., j m }, each job descriptionj m Also made up of words or phrases; Construct an interview guidance model for matching, as shown below: Among them, M represents the interview guidance model, S nm Resume n With position j m The matching score of A feature vector representing a user's resume, The feature vector representing the job description, N nm According to the resume n and position j m The relevant node subsets and their relationships in the graph obtained by feature query; According to the matching score S nm , sort all the combinations of resumes and positions, and select the combinations with the highest matching degree as the recommended user data set: Among them, θ is the set matching threshold; S5, constructing an interview score prediction model based on deep learning to predict interview scores according to the recommended user data set, specifically includes: The interview score prediction model is constructed as follows: Let the input feature vector be s∈R n , represents the characteristics of each applicant; the actual interview score is y∈R; Input layer to hidden layer weight W1∈R H×n , the bias term b1∈R H ; Hidden layer to output layer weight W2∈1×H, bias term b2∈R; where H represents the number of neurons in the hidden layer; Construct forward propagation; wherein the forward propagation includes the input layer to the hidden layer and the hidden layer to the output layer, and the input layer to the hidden layer is expressed as follows: Z1=σ(sW1+b1) h=σ(Z) Among them, Z1 represents the weighted sum of neurons from the input layer to the hidden layer, σ represents the activation function, and h represents the hidden layer output; The hidden layer to the output layer is represented as follows: Z2=hW2+b2 Among them, Z2 represents the weighted sum from the hidden layer to the output layer, Represents the predicted output of the output layer; Using the cross entropy loss function, it is expressed as follows: Among them, N represents the total number of samples, C represents the total number of categories, and y i represents the real sample; Use back propagation to calculate the gradient and update the weights; Taking multiple initial interview score prediction models as individuals of the population, using the accuracy of interview score prediction as the objective function, and using the sparrow search algorithm to perform iterative search to optimize and adjust the parameters of the interview score prediction model; wherein the sparrow search algorithm is expressed as follows: Among them, ε n+1 represents the updated interview score prediction model variable, ε n represents the interview score prediction model variable, and μ represents the control parameter.

2. The interview guidance method based on a large model and artificial intelligence according to claim 1, characterized in that: The large model is used to convert data into high-dimensional vectors to capture semantic information; wherein the large model is a pre-trained Transformer model; the specific steps are as follows: Input data sequence X = {x1, x2, .., x i .,x n }, where x i is a word or subword in the data; The data sequence is input into the encoding layer and represented as follows: E 0 =W scr X+P i A t =Layer(E t-1 +Head(E t-1 ,E t-1 ,E t-1 )) E t =Layer(A t +FEN(A t )) Among them, E 0 represents the output of the 0th layer encoder, W scr Represents the word embedding matrix of the source data, P i represents the position encoding corresponding to the source data, Head represents the multi-head self-attention function, E t-1 represents the output of the t-1th layer encoder, A t represents the output of multi-head self-attention, E t represents the output of the t-th layer encoder, FEN represents the feedforward neural function, and Layer represents the normalization function; The output E of the t-th layer encoder t The input is passed through the encoder to extract the semantic vector, which is expressed as follows: Among them, D t represents the semantic feature vector, Represents the word embedding matrix of the target data, D 0 represents the output vector of the encoder at layer 0, X hot Represents the one-hot matrix corresponding to the target data, represents the output vector of the multi-head self-attention of the t-th layer decryptor, D t-1 represents the output vector of the encoder at layer t-1, MASK represents the multi-head attention function with occlusion mechanism in the decoder, represents the output vector of the multi-head attention with occlusion mechanism in the t-th layer decryptor-decoder, E n Represents the output of the nth layer encoder.

3. The interview guidance method based on a large model and artificial intelligence according to claim 2, characterized in that: The machine learning model is used to extract semantic features and time series features according to semantic information, sentimentally annotate the semantic information, and extract sentiment features, as follows: Introducing time series features X′ t , according to the encoder output D t , train a machine learning model to classify the sentiment of semantic features, which is expressed as follows: Classifier(D t )→y Among them, y represents the sentiment label; The time series feature X′ t Combined with the semantic sentiment features of the corresponding moment, a fusion feature vector F is formed c , which is expressed as follows: F c =Concatenate(X′ t ,SentimentScore t ) Among them, Concatenate represents the feature fusion function, SentimentScore t represents the sentiment score; Define a loss function to measure the gap between the model prediction and the actual label, and use gradient descent to minimize the loss function to update the model parameters; wherein the loss function is expressed as follows: in, Represents the actual sentiment score.

4. The interview guidance method based on a large model and artificial intelligence according to claim 1, characterized in that: The dynamic knowledge graph database construction is expressed as follows: Define the entities, relationships and attributes of the knowledge graph; wherein the entities include questions, answers, sentiment states, and time series features; Choose Neo4j to build a knowledge graph database, import entities and as nodes, relationships as edges, and attributes into the knowledge graph database; Design an update mechanism for the dynamic knowledge graph to update the knowledge graph in real time or regularly based on new interview data.

5. The interview guidance method based on a large model and artificial intelligence according to claim 4, characterized in that: The update mechanism of the dynamic knowledge graph is expressed as follows: Construct a dynamic knowledge graph G = (V, E), where V is the set of entity nodes and E is the set of relationship edges; Let D be the newly arrived interview dataset, each data item d∈D is an interview record, containing semantic features S(d), time series features T(d), and emotional features Emo(d); For each d, identify the corresponding entity e∈E, let f_match(e,d) be the matching function, and return the identifier of the existing entity or the newly created entity related to d; For entity e and its attribute set A, let update_attribute(A,feature) be the update function, which updates the entity attributes according to the features of the new data d; If each data item d introduces a new relational attribute, or if a new question is answered and a new answer is obtained, an edge r∈R needs to be created. Let create_relation(e1,e2,r) be the edge creation function, where e1, e2 are entities and r is the relation type.

6. The interview guidance method based on a large model and artificial intelligence according to claim 1, characterized in that: in, The interview score prediction model variables are weights and bias terms.

7. The interview guidance method based on a large model and artificial intelligence according to claim 1, characterized in that: The user privacy mechanism is designed to specifically include: Let the plaintext be S m , belongs to the plaintext space m, and the key is S k , belongs to the key space k; The encryption function is Φ, which maps the elements of the plaintext space m and the key space k to the ciphertext space Θ; The encryption process is expressed as: Among them, Ψ represents the ciphertext, belongs to the ciphertext space Θ, and represents the plaintext S m Through the encryption function Φ and the key S k Converted from The decrypted representation is: in, Used to ensure that the encrypted ciphertext Ψ can be encrypted with the same key S k Restore to the original plaintext S k .

8. An interview guidance system based on a large model and artificial intelligence, characterized in that: The system comprises: A data collection module, used to collect interview guidance data; wherein the interview guidance data includes at least two of the following: interview questions, job seeker's answers, interviewer's evaluation, interview scenario, job information and job seeker's background; Feature extraction module, used to pre-process the interview guidance data, use a large model to convert the data into a high-dimensional vector, capture semantic information, use a machine learning model to extract semantic features and time series features based on the semantic information, perform sentiment annotation on the semantic information, and extract sentiment features; The knowledge graph construction module is used to build a dynamic knowledge graph database based on semantic features, time series features, and sentiment features combined with historical interview guidance data; The data matching module is used to extract key features based on the resume and target position uploaded by the user in real time using natural language processing technology, and to query the dynamic knowledge graph database using the graph database query language. At the same time, an interview guidance model is built and used to match the user portrait with the position requirements to obtain a recommended user data set. Score prediction module, which is used to build a deep learning interview score prediction model to predict interview scores based on the recommended user dataset; Privacy design module, used to design user privacy mechanisms to ensure the secure storage and transmission of user information; The report output module is used to design a simulated interviewer model based on the recommended user dataset, interact with the user, evaluate the user's answer and calculate the actual interview score, dynamically adjust the recommended user dataset based on the predicted interview score and the actual interview score, and output the recommended user dataset to the user in the form of a report; The steps performed by the knowledge graph construction module specifically include: S301. Define the knowledge graph as G = {S, K, E}, where S represents a feature triple, R represents an entity, and E represents a relationship; S302, the feature triple is defined as S={<c,t,p> |c, t∈K, p∈E}, where c represents the head entity, t represents the tail entity, and p represents the relationship; S303, mapping the entity K embedding representation matrix to a d-dimensional embedding space; wherein, both the head entity and the tail entity are mapped to the d-dimensional embedding space, and a d-dimensional mapping head entity c is obtained. c 、c t and d-dimensional mapping tail entity t c ,t t , and c c ,t c ∈R d , R represents an integer; S304, map the relationship matrix E to the d*d dimensional embedding space, convert the relationship vector into a relationship matrix, and obtain the relationship matrix p -1 ∈R d×d ; S305. Design of bilinear score S core (c, t, p), expressed as follows: S core (c,t,p)=α·f(c c ,p,t t )+β·f(t c ,p -1 ,c t ) Where f() represents a double-line scoring function based on a single component, α and β represent weight factors respectively; S306, construct a loss function to train the scores of feature triples to minimize the loss function while avoiding overfitting; the loss function L γ It is expressed as follows: L γ =∑log(1-exp(-l*clamp(S core (c,t,p),δ))) Among them, clamp represents the clamp function, which is used to control the range within (S core (c, t, p), δ); δ represents the cropping factor, l represents the label; The interview guidance model is constructed as follows: Assume that user resume is represented as document set R = {r1,r2...,r n }, where each resume r n is a string consisting of a series of words or phrases. The target job description is represented by a document set J = {j1, j2, ..., j m }, each job descriptionj m Also made up of words or phrases; Construct an interview guidance model for matching, as shown below: Among them, M represents the interview guidance model, S nm Resume n With position j m The matching score of A feature vector representing a user's resume, The feature vector representing the job description, N nm According to the resume n and position j m The relevant node subsets and their relationships in the graph obtained by feature query; According to the matching score S nm , sort all the combinations of resumes and positions, and select the combinations with the highest matching degree as the recommended user data set: Among them, θ is the set matching threshold; S5, constructing a deep learning interview score prediction model to predict the interview score according to the recommended user data set, specifically includes: The interview score prediction model is constructed as follows: Let the input feature vector be s∈R n , represents the characteristics of each applicant; the actual interview score is y∈R; Input layer to hidden layer weight W1∈R H×n , the bias term b1∈R H ; Hidden layer to output layer weight W2∈1×H, bias term b2∈R; where H represents the number of neurons in the hidden layer; Construct forward propagation; wherein the forward propagation includes the input layer to the hidden layer and the hidden layer to the output layer, and the input layer to the hidden layer is expressed as follows: Z1=σ(sW1+b1) h=σ(Z) Among them, Z1 represents the weighted sum of neurons from the input layer to the hidden layer, σ represents the activation function, and h represents the hidden layer output; The hidden layer to the output layer is represented as follows: Z2=hW2+b2 Among them, Z2 represents the weighted sum from the hidden layer to the output layer, Represents the predicted output of the output layer; Using the cross entropy loss function, it is expressed as follows: Among them, N represents the total number of samples, C represents the total number of categories, and y i represents the real sample; Use back propagation to calculate the gradient and update the weights; Taking multiple initial interview score prediction models as individuals of the population, using the accuracy of interview score prediction as the objective function, and using the sparrow search algorithm to perform iterative search to optimize and adjust the parameters of the interview score prediction model; wherein the sparrow search algorithm is expressed as follows: Among them, ε n+1 represents the updated interview score prediction model variable, ε n represents the interview score prediction model variable, and μ represents the control parameter.

Images