Interview tutoring method and device based on large language model, equipment and medium

Abstract

The invention relates to the technical field of deep learning, in particular to an interview tutoring method and device based on a large language model, equipment and a medium, and the method comprises the steps: obtaining resume data of a job seeker and target post demand data corresponding to the resume data; inputting the resume data of the job seeker and the target post demand data into the large language model, and outputting a first interview question and a corresponding first interview answer; inputting the first interview question and the first interview answer into a relevancy scoring model and outputting a relevancy score; inputting the resume data of the job seeker and the first interview answer into a credibility scoring model, and outputting a credibility score; if the comprehensive score is smaller than a preset threshold value, inputting the relevancy score and the credibility score into the large language model again until the comprehensive score is larger than the preset threshold value, and outputting the first interview question and the corresponding interview answer. According to the technical scheme, the quality of the finally output answer can be effectively improved, and job seekers are helped to improve interview performance.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to an interview coaching method, device, electronic device and storage medium based on a large language model. Background Art

[0002] In the modern job search process, interviews have become a key link in determining whether job seekers can obtain their target positions. Since interview performance directly affects the results of recruitment, many job seekers hope to improve their competitiveness through interview coaching services. Such services can help job seekers better demonstrate their personal advantages and highlight their highlights in accordance with job requirements, and thus are gradually becoming widely welcomed by job seekers.

[0003] At present, interview coaching services mainly rely on manual services, providing one-on-one customized coaching for job seekers through a team of experts such as senior HR, headhunters and career planners. This type of service can deeply tap the personal ability advantages of job seekers and provide detailed interview skills guidance. However, due to the high labor costs of such services, the prices are usually expensive, resulting in low penetration in the market. In addition, the limited expert resources also restrict the large-scale popularization of such services, making it difficult to effectively meet the needs of a large number of job seekers.

[0004] On the other hand, the rapid development of artificial intelligence technology has spawned a class of AI-based interview simulation systems. Such systems usually have the function of automatically generating interview questions and evaluating interview performance, which can help improve interview efficiency to a certain extent. However, the main purpose of such systems is to improve the efficiency of recruitment screening and evaluation. They do not fully explore the advantages of job seekers' experiences, and it is difficult to effectively provide personalized coaching suggestions for job seekers. Summary of the invention

[0005] In view of the above problems, the present application proposes an interview coaching method based on a large language model, comprising: obtaining resume data of a job seeker and target position requirement data corresponding to the resume data; inputting the resume data of the job seeker and the target position requirement data into the large language model and outputting a first interview question and a corresponding first interview answer; inputting the first interview question and the first interview answer into a pre-trained relevance scoring model and outputting a relevance score, wherein the relevance scoring model is trained using a plurality of first sample data, wherein the first sample data includes interview questions and interview answers, and according to the relevance of the interview questions and the interview answers, the first sample data is divided into positive samples and negative samples; the resume data of the job seeker and the target position requirement data are input into the large language model and outputting a first interview question and a corresponding first interview answer; inputting the first interview question and the first interview answer into a pre-trained relevance scoring model and outputting a relevance score. The first interview answer is input into a pre-trained credibility scoring model and a credibility score is output. The credibility scoring model is trained using multiple second sample data. The second sample data includes resume data and interview answers. According to the logical relationship between the resume data and the interview answers, the types of the second sample data are divided into: implication, contradiction and neutrality. A comprehensive score is determined according to the relevance score and the credibility score. If the comprehensive score is less than a preset threshold, the relevance score and the credibility score are input into the large language model again, the second interview answer is output, and the comprehensive score of the first interview question and the second interview answer is calculated until the comprehensive score is greater than the preset threshold, and the first interview question and the corresponding interview answer are output.

[0006] In the interview coaching method as described above, the relevance scoring model is trained by using the relevance scores of the same interview question and the relevant interview answers and the irrelevant interview answers respectively.

[0007] In the interview coaching method as described above, the training process of the relevance scoring model includes: constructing multiple first sample data, wherein positive samples include interview questions and interview answers related to the interview questions, and negative samples include interview questions and interview answers unrelated to the interview questions; indexing the interview questions and interview answers in the first sample data to obtain a first sample sequence code, wherein the first sample sequence code includes a classification identifier for indicating whether the interview questions and interview answers are related; inputting the first sample sequence code into the relevance scoring model to output a relevance score; and updating the model parameters in the relevance scoring model by minimizing a first loss function value until the first loss function value is less than a first loss threshold, wherein the first loss function includes the relevance score of the positive sample and the relevance score of the negative sample.

[0008] In the interview coaching method described above, the first loss function is expressed by the following formula:

[0009]

[0010] Among them, lossA is the value of the first loss function, s θ (q,a+ ) is the correlation score of the positive sample, is the correlation score of the negative sample, and k is the number of negative samples.

[0011] In the interview coaching method as described above, the training process of the credibility scoring model includes: constructing multiple second sample data, the second sample data including resume data and interview answers; indexing the resume data and interview answers in the second sample data to obtain a second sample sequence code, the second sample sequence code including a classification identifier for representing the logical relationship between the resume data and the interview answer; inputting the second sample sequence code into the credibility scoring model, outputting the unnormalized prediction value on each logical relationship and converting it into a probability distribution value on each logical relationship; updating the model parameters in the credibility scoring model by minimizing the second loss function value until the second loss function value is less than the second loss threshold, the second loss function including the true type of the second sample data and the probability distribution value on each logical relationship.

[0012] In the interview coaching method described above, the second loss function is expressed by the following formula:

[0013]

[0014] Among them, lossB is the value of the second loss function, p c is the probability that the sample data belongs to category C, C = 3, the categories are implication, contradiction and neutral, y c is the true label of the sample in category c. If the true category of the sample is c, then y c is 1, otherwise y c is 0.

[0015] The interview coaching method as described above further includes: taking the sum of the probability distribution values ​​of the resume data and the interview answers in the implication and neutral categories respectively as the credibility score.

[0016] According to another aspect of the present application, an interview coaching device based on a large language model is proposed, comprising: an acquisition module, used to acquire resume data of a job seeker and target position requirement data corresponding to the resume data; an interview question and answer generation module, used to input the resume data of the job seeker and the target position requirement data into the large language model and output a first interview question and a corresponding first interview answer; a relevance calculation module, used to input the first interview question and the first interview answer into a pre-trained relevance scoring model and output a relevance score, wherein the relevance scoring model is trained using a plurality of first sample data, wherein the first sample data includes interview questions and interview answers, and the first sample data is divided into positive samples and negative samples according to the relevance of the interview questions and the interview answers; a credibility calculation module, used to input the resume data of the job seeker and the target position requirement data into the large language model and output a first interview question and a corresponding first interview answer; The first interview answer is input into a pre-trained credibility scoring model and a credibility score is output. The credibility scoring model is trained using multiple second sample data. The second sample data includes resume data and interview answers. According to the logical relationship between the resume data and the interview answers, the types of the second sample data are divided into: implication, contradiction and neutrality; a comprehensive score calculation module is used to determine the comprehensive score according to the relevance score and the credibility score; a question and answer optimization module is used to input the relevance score, the credibility score and the first interview question into the large language model again if the comprehensive score is less than a preset threshold, optimize the first interview question, output the second interview answer, and calculate the comprehensive score of the first interview question and the second interview answer until the comprehensive score is greater than the preset threshold, and output the first interview question and the corresponding interview answer.

[0017] According to another aspect of the present application, an electronic device is proposed, which includes a processor and a memory, wherein a computer program instruction set is stored in the memory, and when the processor executes the computer program instruction set in the memory, the interview coaching method based on the large language model as described above is implemented.

[0018] According to another aspect of the present application, a computer-readable storage medium is provided, wherein a computer program instruction set is stored on the computer-readable storage medium, and when the computer program instruction set is executed by a processor, the interview coaching method based on a large language model as described above is implemented.

[0019] This application can effectively improve the quality of the final output answers by evaluating the relevance and credibility of the generated answers. It ensures that the answers provided by the system are not only semantically consistent, but also meet actual needs in terms of details and expression, helping job seekers improve their interview performance. Through the optimization of relevance and credibility evaluation, it can generate tailored interview questions and answer suggestions for each job seeker, providing a highly personalized interview coaching experience. This customized coaching method helps job seekers improve their interview performance in a targeted manner and enhance their competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The preferred embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, wherein:

[0021] Figure 1 This is a flow chart of an interview coaching method based on a large language model according to an embodiment of the present application.

[0022] Figure 2 It is a flow chart of a training method of a relevance scoring model according to an embodiment of the present application.

[0023] Figure 3 It is a flow chart of a training method of a relevance scoring model according to an embodiment of the present application.

[0024] Figure 4 It is a schematic diagram of the structure of an interview coaching device based on a large language model according to an embodiment of the present application.

[0025] Figure 5 It is a schematic diagram of the hardware structure principle of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] In the following detailed description, reference may be made to the various specification drawings that are part of the present application and are used to illustrate specific embodiments of the present application. In the accompanying drawings, similar reference numerals describe substantially similar components in different figures. The various specific embodiments of the present application are described below in sufficient detail so that a person of ordinary skill in the art with relevant knowledge and skills in the art can implement the technical solutions of the present application. It should be understood that other embodiments may also be used or structural, logical or electrical changes may be made to the embodiments of the present application.

[0028] Large Language Model (LLM) is a natural language processing (NLP) technology based on artificial intelligence, which is usually trained by deep learning methods and can understand, generate and process human language. Large language models are usually trained using large amounts of text data to master the structure, grammar, meaning and context of the language. The large language model in this application can learn complex language patterns, which can generate relevant interview questions based on the resume text of the job seeker, and combine the information of the resume text to personalize the interview answers corresponding to the interview questions.

[0029] Figure 1 FIG. 1 is a flow chart of an interview coaching method based on a large language model according to an embodiment of the present application. Figure 1 As shown, the method includes:

[0030] S11, obtaining the resume data of the job seeker and the target position requirement data corresponding to the resume data;

[0031] S12, inputting the resume data of the job seeker and the target position requirement data into the large language model and outputting the first interview question and the corresponding first interview answer;

[0032] S13, inputting the first interview question and the first interview answer into a pre-trained relevance scoring model and outputting a relevance score, wherein the relevance scoring model is trained using a plurality of first sample data, wherein the first sample data includes interview questions and interview answers, and the first sample data is divided into positive samples and negative samples according to the relevance between the interview questions and the interview answers;

[0033] S14, inputting the resume data and the first interview answer of the job seeker into a pre-trained credibility scoring model and outputting a credibility score, wherein the credibility scoring model is trained using a plurality of second sample data, wherein the second sample data includes the resume data and the interview answer, and according to the logical relationship between the resume data and the interview answer, the types of the second sample data are classified into: implication, contradiction and neutrality;

[0034] S15, determining a comprehensive score according to the relevance score and the credibility score;

[0035] S16: If the comprehensive score is less than the preset threshold, the relevance score and the credibility score are input into the large language model again, the second interview answer is output, and the comprehensive score of the first interview question and the second interview answer is calculated until the comprehensive score is greater than the preset threshold, and the first interview question and the corresponding interview answer are output.

[0036] The resume data of job seekers includes text information such as the basic information, educational background, work experience, project experience, and personal skills of the job seekers. The target position requirement data is the position that the job seeker wants to obtain in the resume. The job fair will retrieve the general job requirement text for this position according to the position name. Before inputting the resume data and the target position requirement data into the large language model, it is necessary to preprocess this text information. The preprocessing steps include:

[0037] Text cleaning: Remove irrelevant characters, special symbols, HTML tags, spaces, extra line breaks, etc. from the resume and position requirement text to ensure the cleanliness and standardization of the text. This step can improve the accuracy of the model's understanding of the text content.

[0038] Word segmentation and lemmatization: Segment the text to break the sentence into words or phrases. At the same time, perform lemmatization (such as restoring different tenses of verbs to the base form and plural forms of nouns to the singular form) to reduce the interference of lexical redundancy and variants on the model analysis.

[0039] Stop word filtering: Remove common words with no practical meaning (such as "of", "and", "is", etc.) to retain the meaningful key content and ensure that the model focuses on important information such as educational background and work experience.

[0040] Entity recognition and extraction: Through the named entity recognition (NER) technology in natural language processing, identify and extract important entity information in the resume, such as name, position, company name, time, location, etc. Similarly, perform entity recognition on the position information and extract the position name, job responsibilities, skill requirements, etc.

[0041] Data standardization: Standardize the formats of dates and numerical information (such as work years, educational background, etc.) in the resume and position data to make them unified into a consistent format, so that the model can more accurately understand and compare different data points.

[0042] Vectorization processing: Perform vectorization processing on the resume and position information through an encoder, convert the text into a numerical form suitable for model processing, and generate vector indexes for the resume and position. This step embeds the text information into a high-dimensional space, enabling the large language model to better understand the semantic relationships between texts.

[0043] Duplicate removal and content verification: Check and remove duplicate content in the resume or position information to ensure that the data input into the model is concise and without redundancy. At the same time, verify the integrity and accuracy of the resume and position information to avoid missing key data.

[0044] The processed resume data and target job requirement data are input into the big language model. The big language model will generate multiple interview questions according to different question types based on the parsed resume information and job information, including basic background questions, behavioral interview questions, scenario simulation questions, etc. Then, for the generated interview questions, the most relevant content is queried from the resume and job vector indexes respectively to determine the best answer material. The queried content is used as context and submitted to the big model together with the interview questions to generate a preliminary interview answer, i.e. the first interview answer.

[0045] The big language big model of the present application includes but is not limited to any one of Tongyi Qianwen, Wenxin Yiyan, and Zhipu GLM. Alternatively, the big language big model can also be fine-tuned in vertical fields on the open source versions provided by the suppliers of big models such as Tongyi Qianwen, Wenxin Yiyan, and Zhipu GLM.

[0046] The relevance of this application solves the problem of judging whether the interview answers generated by the large model are "irrelevant" to the interview questions. For example, if the interview question is "What school did you graduate from?" and the interview answer is "I worked at xx company", the correlation between the two is irrelevant. This pair of sample data is a negative sample, indicating that the interview answer is an inappropriate answer to the interview question (it can be a completely irrelevant or partially relevant answer). If the answer is "I graduated from xx University", the correlation between the two is relevant, then this pair of sample data is a positive sample, indicating that the interview answer is an appropriate answer to the interview question.

[0047] The credibility of this application solves the problem of judging whether the interview answers generated by the large model are inconsistent with the information in the job applicant's resume data. Inspect the implicit relationship between the interview answers and the relevant content of the resume, and examine the consistency of the information. For example, part of the answer is "worked for 5 years", but the resume information says "worked for 3 years", which results in a contradictory relationship. The generated answer may be inconsistent with the resume content and has low credibility. If the resume information says "worked for 3 years", the two are in an implicit relationship, indicating that the generated answer is semantically consistent with the resume content and has a high credibility; if the resume information says "graduated 5 years ago", the relationship between the two is neutral, indicating that the generated answer has no direct contradiction or implicit relationship with the resume content.

[0048] The scoring module of this application can effectively improve the quality of the final output answers by evaluating the relevance and credibility of the generated answers. It ensures that the answers provided by the system are not only semantically consistent, but also meet actual needs in terms of details and expression, helping job seekers improve their interview performance. Through the optimization of relevance and credibility evaluation, it can generate tailored interview questions and answer suggestions for each job seeker, providing a highly personalized interview coaching experience. This customized coaching method helps job seekers improve their interview performance in a targeted manner and enhance their competitiveness.

[0049] Furthermore, compared with traditional manual interview coaching services, this technical solution greatly reduces the cost of interview coaching by automatically generating questions and answers. It can provide services to a large number of job seekers at the same time, is not restricted by the limited resources of experts, and has stronger scalability and popularity.

[0050] According to one embodiment of the present application, the relevance scoring model is trained using the relevance scores of the same interview question with relevant interview answers and irrelevant interview answers. The first sample data of the present application is a triple, and the form of a triple is: (interview question, positive sample answer, negative sample answer list). The positive sample answer is a suitable answer to the question, and the negative sample answer is an inappropriate answer to the same question (which can be a completely irrelevant or partially relevant answer). The positive sample answer and the negative sample answer of the same interview question are input into the relevance scoring model together. The relevance scoring model optimizes its parameters by contrastive learning, which effectively improves the training efficiency of the relevance scoring model. Thereby providing job seekers with more accurate answer suggestions that meet the actual job requirements in interview coaching, improving the quality and effect of interview coaching.

[0051] Figure 2 FIG. 1 is a flow chart of a training method for a relevance scoring model according to an embodiment of the present application. Figure 2 As shown, the method includes:

[0052] S131, constructing a plurality of first sample data, wherein a positive sample includes an interview question and an interview answer related to the interview question, and a negative sample includes an interview question and an interview answer unrelated to the interview question;

[0053] S132, indexing the interview questions and the interview answers in the first sample data to obtain a first sample sequence code, wherein the first sample sequence code includes a classification identifier for indicating whether the interview questions and the interview answers are related;

[0054] S133, inputting the first sample sequence code into a relevance scoring model to output a relevance score;

[0055] S134, updating the model parameters in the relevance scoring model by minimizing the first loss function value until the first loss function value is less than a first loss threshold, wherein the first loss function includes the relevance score of the positive sample and the relevance score of the negative sample.

[0056] This application takes the interview question and the interview answer in the triple as a pair of input samples, and inputs them into the relevance scoring model to be trained. The relevance scoring model can be a cross encoder model (such as bge-reranker), which encodes the text. The cross encoder considers the entire sequence of inputs, encodes the semantic context of the question and answer together, and inputs the encoding into the model. The model outputs a relevance score (usually between 0 and 1) to measure the match between the answer and the interview question. The score is used to evaluate whether the generated answer semantically answers the question.

[0057] The input of the cross encoder is the text pair (interview question q, answer a). The text pair needs to be preprocessed so that the cross encoder can accept these texts and generate corresponding vector representations. Specifically, first, the two sentences are concatenated by the separator [SEP], and a special symbol [CLS] is added at the beginning of the sentence as a marker for the classification task, i.e., a classification identifier.

[0058] Then, the concatenated text is segmented (sentence -> word sequence) and indexed (word sequence -> id sequence), and finally the following sequence [CLS], q0, q1, …, qn, [SEP], a0, a1, …, am is obtained as the input of the encoder.

[0059] Finally, the concatenated text sequence is input into the relevance scoring model. Since the relevance scoring model processes both interview questions and answers at the same time, the model can capture the interdependencies and interactive information between interview questions and answers. In this process, the model generates a context vector for each word in the input sequence. The output vector corresponding to the "[CLS]" tag of the last layer is extracted as the comprehensive representation vector of the entire concatenated text. This vector is input into a linear layer and outputs a scalar score, which can be used as the semantic relevance score between the interview question and the answer. This calculation process is expressed as s θ express:

[0060] score1=s θ (q,a);

[0061] For a piece of training data, the loss function can be expressed using the following formula:

[0062]

[0063] Among them, lossA is the value of the first loss function, s θ (q,a + ) is the correlation score of the positive sample, is the correlation score of the negative sample, and k is the number of negative samples.

[0064] During the contrastive learning training process, the model parameters of the cross encoder are updated by minimizing the contrastive learning loss value of the cross encoder, so that the text similarity between the interview questions and the positive sample answers can be as high as possible, and the text similarity between the interview questions and the negative sample answers can be as low as possible.

[0065] This solution constructs a dataset of positive and negative samples and uses a relevance scoring model to evaluate the semantic relevance between interview questions and answers. The model is based on a deep learning encoder that converts interview questions and answers into numerical vector representations and performs supervised learning on positive and negative samples through classification identifiers. The model uses a loss function optimization strategy to maximize the scores of relevant answers and minimize the scores of irrelevant answers, thereby gradually improving the ability to judge the relevance of answers during the training process.

[0066] Figure 3 FIG. 1 is a flow chart of a training method for a relevance scoring model according to an embodiment of the present application. Figure 3 As shown, the method includes:

[0067] S141, constructing a plurality of second sample data, the second sample data including resume data and interview answers;

[0068] S142, indexing the resume data and the interview answers in the second sample data to obtain a second sample sequence code, wherein the second sample sequence code includes a classification identifier for indicating a logical relationship between the resume data and the interview answers;

[0069] S143, inputting the second sample sequence code into the credibility scoring model, outputting the unnormalized prediction value on each logical relationship and converting it into a probability distribution value on each logical relationship;

[0070] S144, updating the model parameters in the credibility scoring model by minimizing the second loss function value until the second loss function value is less than the second loss threshold, wherein the second loss function includes the true type of the second sample data and the probability distribution value on each logical relationship.

[0071] The format of the second sample data of this application is: [resume text, generated interview answer] sample pair, and the logical relationship between them is marked (implied, contradictory or neutral). At the same time, open source datasets such as OCNLI, CSNLI, STS-B-Chinese can be used to expand the dataset.

[0072] First, the sample pairs of [resume text, generated interview answers] need to be preprocessed. The preprocessing steps are similar to the above-mentioned process of calculating the relevance, and both of them are to segment and index the sample pairs to obtain the second sample sequence code.

[0073] Then, the second sample sequence encoding is input into the credibility scoring model. The credibility scoring model can be a natural language inference (NLI) model trained by supervised learning using a pre-trained model as a base (for example, based on BERT and its derivative models). The representation vector corresponding to the [CLS] tag output by the model is input into the linear classification layer to obtain the unnormalized prediction value (logits) on each logical relationship category (implication, contradiction and neutrality), and the logits are converted into a probability distribution (p entailment ,p contradiction ,p neutral ). The unnormalized predicted value is converted into a probability distribution value using the following formula:

[0074]

[0075] Among them, p c is the probability that the second sample data belongs to category C, C = 3, the categories are implication, contradiction and neutral, Z c is the unnormalized predicted value of the sample data in the analogy C.

[0076] Finally, the cross-entropy loss function is used to train the NLI model. For a single sample, the second loss function is expressed as follows:

[0077]

[0078] Among them, lossB is the value of the second loss function, p c is the probability that the sample data belongs to category C, C = 3, the categories are implication, contradiction and neutral, y c is the true label of the sample in category c. If the true category of the sample is c, then y c is 1, otherwise y c is 0.

[0079] If the second loss function value is greater than or equal to the second loss threshold, after updating the parameters in the model, continue to calculate the second loss function value until the second loss function value is less than the second loss threshold, indicating that the model has met the target requirements, stop optimization training, and fix the parameter values ​​in the model.

[0080] After the NLI model inference is completed, the sum of the probability (i.e., non-contradiction rate) distribution values ​​of the resume data and the interview answers in the implied and neutral categories is used as the credibility score. It is expressed as follows:

[0081] score2=p entailment +p neutral ;

[0082] Among them, score2 is the credibility score, p entailment is the probability distribution value under the logical relationship of "implied" category, p neutral is the probability distribution value under the “neutral” category of the logical relationship.

[0083] This application can ensure that the interview answers are semantically consistent with the resume and avoid self-contradictions by determining the logical relationship between the interview answers and the resume, i.e. calculating the relevance score. Probability distributions (such as implication, contradiction, and neutrality) provide a quantitative evaluation method to quickly identify which answers need to be optimized and which are already a good match. Based on the credibility score, the current interview answers can be optimized and modified in a targeted manner. Better demonstrate your strengths in the interview answers and improve your performance in the interview.

[0084] According to an embodiment of the present application, the relevance score and the credibility score are weighted and summed to obtain a comprehensive score. The comprehensive score is expressed by the following formula:

[0085] S=score1×W A +score2×W B ;

[0086] Among them, S is the comprehensive score, score1 is the relevance score, and W A is the weight of the relevance score, score2 is the credibility score, and W B is the weight of the credibility score.

[0087] Among them, W A The value range is 0.5-0.6, W B The value range is 0.4-0.5.

[0088] The above describes the implementation methods and advantages of the embodiments of the present application through multiple embodiments. This application describes the specific processing process of the embodiments of the present application in detail with reference to specific examples.

[0089] S1: Assume that a job seeker submits a resume and a recruitment requirement for the target position of "Data Analyst".

[0090] S2: The system uses NLP technology to parse resumes and job requirements, extract key information, and use pre-trained encoders to vectorize the corresponding text segments and build vector indexes for subsequent matching and scoring.

[0091] S3: Interview question generation

[0092] Based on the parsed resume information and job requirements, the system generates multiple interview questions. Take the following question as an example:

[0093] Q1.Please briefly describe your work experience in the field of data analysis.

[0094] S4: Preliminary answer generation

[0095] The following information (summary) was retrieved from the resume index and job requirements index using Q1:

[0096] - Resume context: "Company A (2019-Present): Senior Data Analyst, Job Description: ..., Company B (2016-2018): Junior Data Analyst, Job Description: ...";

[0097] -Job requirement context: "Experience in data analysis and machine learning is required."

[0098] Generate preliminary interview answers based on interview questions, resume context, and job requirements context:

[0099] A1: "I worked as a junior data analyst at Company A, mainly responsible for data cleaning and analysis. After joining Company B, I was promoted to a senior data analyst, participated in a certain project, and used machine learning technology to support the decision-making process."

[0100] S5: Scoring module

[0101] S5.1 Relevance score

[0102] -Q1&A1 relevance score: 0.65 (Although relevant, the answer is not substantial and the match between the answer and the question is moderate)

[0103] S5.2 Credibility Rating

[0104] - Resume context & A1 credibility score: 0.3 (the answer is obviously inconsistent with the resume)

[0105] S5.3 Calculation of comprehensive score

[0106] -A1: 0.65 (correlation) × 0.6 (correlation weight) + 0.3 (credibility) × 0.4 (credibility weight) = 0.51

[0107] -Threshold judgment: Assume that the preset threshold is 0.85

[0108] -A1: 0.51<0.85 → Failed

[0109] S6: Result Feedback and Optimization

[0110] Since A1's comprehensive score is lower than the threshold, the interview questions, resume and position-related content, previously generated answers, and previous scoring results of each dimension are fed back to the big model, and it is required to optimize the interview answers again.

[0111] Generate answer A2 again:

[0112] "From 2016 to 2018, I worked as a junior data analyst at Company B, mainly responsible for data cleaning and analysis using SQL and Python, optimizing data processing processes and improving efficiency. After joining Company A in 2019, I was promoted to senior data analyst, participated in a certain project, and used machine learning technology to support the decision-making process."

[0113] The scoring module evaluates A2: relevance score 0.75 (the answer is still not convincing), credibility score 0.9 (very reliable), comprehensive score = 0.75 × 0.6 + 0.9 × 0.4 = 0.81 < 0.85 → Fail

[0114] Feedback again to generate answer A3:

[0115] “From 2016 to 2018, I worked as a junior data analyst at Company B. I was mainly responsible for collecting and organizing data from different departments, and using SQL and Python for preliminary data cleaning and analysis. In addition, I wrote scripts to automate report generation, reducing manual operation time by 30 hours per month and improving team work efficiency.

[0116] After joining Company A in 2019, I was promoted to Senior Data Analyst, and I began to perform more complex statistical analysis on large data sets. I participated in multiple cross-departmental projects, using machine learning techniques such as linear regression and random forest algorithms to analyze business data and provide data-driven decision support for the team. ”

[0117] The scoring module evaluates A2: relevance score 0.88 (the answer is relevant and the content is detailed), credibility score 0.9 (very reliable), comprehensive score = 0.88×0.6+0.9×0.4=0.89>0.85→passed, and the system returns the optimized answer A3.

[0118] As can be seen from the above, this application uses a scoring module to score interview answers in terms of relevance and credibility. When the score is unqualified, based on the relevance and credibility scores, the current interview answer is guided to be optimized, so that the optimized interview answer is relevant and detailed, and has high credibility and reliable content.

[0119] The optimized interview answers can effectively guide job seekers in dealing with interviews, at least in the following aspects:

[0120] (1) Improve the success rate of interviews. Through correlation analysis with resumes and job descriptions, job seekers can more accurately convey their experience and skills. This accuracy is not only reflected in the consistency of the content, but also in the depth and details of the answers, so that interviewers can clearly understand the professional capabilities of job seekers, thereby increasing the possibility of passing the interview;

[0121] (2) Interview answers are highly targeted. By analyzing the weaknesses and strengths of job seekers, the optimization process can help job seekers focus on the content that is most relevant to the job requirements. This targeting allows job seekers to more effectively demonstrate their match with job requirements during the interview, enhancing the purposefulness and persuasiveness of their answers.

[0122] (3) Improve the matching degree of the interview. In the interview, job seekers can show their understanding and adaptability to the position through precise answers, which not only reflects their serious attitude towards the position, but also shows how they can combine their background with the company's needs. This adaptability makes job seekers stand out from many candidates and leave a deep impression on the interviewer.

[0123] Experimental comparison before and after using the scoring module

[0124] Control group: The first interview question and the first interview answer generated by the large language model are directly pushed to job seekers and interviewers;

[0125] Experimental group: Generate the same first interview question and first interview answer as the control group from the large language model, use the scoring module to optimize the first interview answer in terms of relevance and credibility, output the second interview answer, and push the first interview question and the second interview answer to the job seeker and the interviewer.

[0126] 1. Recruit 50 job seekers and randomly assign them to the control group and the experimental group (25 in each group).

[0127] 2. Recruit 50 interviewers and randomly assign them to the control group and the experimental group (25 in each group).

[0128] 3. The two groups of job seekers and interviewers make evaluations based on the interview questions and answers in their respective groups.

[0129] 4. Collect information on whether job seekers and interviewers are satisfied with the interview questions and answers (satisfied is 1, dissatisfied is 0).

[0130] 5. Compare the evaluations of the control group and the experimental group, and analyze whether the optimized answers significantly improve the job seekers’ satisfaction.

[0131] The experimental results are shown in Table 1 below:

[0132] Table 1

[0133] Group Interviewer satisfaction rating (0 / 1) Job seeker satisfaction rating (0 / 1) Control group 15 / 25 18 / 25 Experimental Group 20 / 25 22 / 25

[0134] Experimental conclusion: According to the above data, the job seeker satisfaction in the experimental group is significantly higher than that in the control group, indicating that the optimized answers generated by the large language model have a positive impact on job seeker satisfaction. The interviewer satisfaction in the experimental group is also significantly higher than that in the control group, indicating that the optimized interview answers are also more popular in the interviewer's evaluation. Therefore, using the optimized interview questions and interview answers to push to job seekers is conducive to improving the job search success rate of job seekers and improving user satisfaction.

[0135] Corresponding to the method embodiment of the present application, the present application also provides an interview coaching device based on a large language model, such as Figure 4 As shown, the interview coaching device 100 based on the large language model includes:

[0136] An acquisition module 110 is used to acquire the resume data of a job seeker and target job requirement data corresponding to the resume data from a resume database and a job requirement database respectively;

[0137] An interview question and answer generating module 120 is used to input the resume data of the job seeker and the target position requirement data into the large language model and output a first interview question and a corresponding first interview answer;

[0138] A relevance calculation module 130, configured to input the first interview question and the first interview answer into a pre-trained relevance scoring model and output a relevance score, wherein the relevance scoring model is trained using a plurality of first sample data, wherein the first sample data includes interview questions and interview answers, and the first sample data is divided into positive samples and negative samples according to the relevance between the interview questions and the interview answers;

[0139] The credibility calculation module 140 is used to input the resume data and the first interview answer of the job seeker into a pre-trained credibility scoring model and output a credibility score, wherein the credibility scoring model is trained using a plurality of second sample data, wherein the second sample data includes the resume data and the interview answer, and according to the logical relationship between the resume data and the interview answer, the type of the second sample data is classified into: implication, contradiction and neutrality;

[0140] Scoring module 150, used to determine a comprehensive score based on the relevance score and the credibility score;

[0141] The question and answer optimization module 160 is used to input the relevance score, credibility score and the first interview question into the large language model again if the comprehensive score is less than a preset threshold, optimize the first interview question, output the second interview answer, and calculate the comprehensive score of the first interview question and the second interview answer until the comprehensive score is greater than the preset threshold, and output the first interview question and the corresponding interview answer.

[0142] Figure 5 1 is a schematic diagram of the hardware structure principle of an electronic device according to an embodiment of the present invention. The electronic device can be implemented as a server or other various terminal devices, such as a desktop personal computer, a tablet computer, a laptop computer, a mobile phone, etc., including a processor 601 and a memory 602. The memory 602 stores a program instruction set. When the processor 601 executes the program instruction set on the memory 602, any of the aforementioned recommendation information generation methods is implemented.

[0143] Specifically, the processor 601 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiment of the present invention.

[0144] The memory 602 may include a large capacity memory for data or instructions. By way of example and not limitation, the memory 602 may include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive or a combination of two or more of these. In appropriate cases, the memory 602 may include a removable or non-removable (or fixed) medium. In appropriate cases, the memory 602 may be inside or outside the integrated gateway disaster recovery device. In a specific embodiment, the memory 602 is a non-volatile solid-state memory.

[0145] The memory may include a read-only memory (ROM), a random access memory (RAM), a magnetic disk storage medium device, an optical storage medium device, a flash memory device, an electrical, optical or other physical / tangible memory storage device. Therefore, generally, the memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., a memory device) encoded with software including computer executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the recommendation information generation method provided by the present invention.

[0146] In one example, the electronic device may further include a communication interface 603 and a bus 604. The processor 601, the memory 602, and the communication interface 603 are connected via the bus 604 and communicate with each other.

[0147] The communication interface 603 is mainly used to implement communication between various modules, devices, units and / or equipment in the embodiment of the present invention.

[0148] Bus 604 includes hardware, software or both, and the components of online data flow billing equipment are coupled to each other. For example, but not limitation, bus may include accelerated graphics port (AGP) or other graphics bus, enhanced industrial standard architecture (EISA) bus, front side bus (FSB), hypertransport (HT) interconnection, industrial standard architecture (ISA) bus, infinite bandwidth interconnection, low pin count (LPC) bus, memory bus, micro channel architecture (MCA) bus, peripheral component interconnection (PCI) bus, PCI-Express (PCI-X) bus, serial advanced technology attachment (SATA) bus, video electronics standard association local (VLB) bus or other suitable bus or two or more of these combinations. In appropriate cases, bus 604 may include one or more buses. Although the embodiment of the present invention describes and shows a specific bus, the present invention considers any suitable bus or interconnection.

[0149] The present invention also provides a computer-readable storage medium having computer program instructions stored thereon, which can be executed by a processor to implement any of the interview coaching methods based on a large language model in the aforementioned embodiments. The computer-readable storage medium can be any medium that can tangibly contain or store computer-executable instructions for use by or in combination with an instruction execution system, device, and apparatus. The storage medium can be a transient computer-readable storage medium or a non-transient computer-readable storage medium. Non-transient computer-readable storage media may include, but are not limited to, magnetic storage devices, optical storage devices, and / or semiconductor storage devices. Examples of such storage devices include, for example, magnetic disks, optical disks based on CD, DVD, or Blu-ray technology, and persistent solid-state memories such as flash memory, solid-state drives, and the like.

[0150] The present invention also provides a computer program product, which includes a set of computer program instructions, and when the set of computer program instructions is executed by a processor, any one of the interview coaching methods based on a large language model in the aforementioned embodiments is implemented. The computer program product includes but is not limited to application installation packages published on websites and application stores, application plug-ins, and small programs that can be run in certain applications.

[0151] It should be clear that the present invention is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present invention.

[0152] The above embodiments are only used to illustrate the present invention, but not to limit the present invention. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the scope of the present invention. Therefore, all equivalent technical solutions should also fall within the scope of the present invention.

Claims

1. An interview coaching method based on a large language model, characterized in that: include: Obtaining the resume data of job seekers and target job demand data corresponding to the resume data; Input the applicant's resume data and target position requirement data into the large language model and output the first interview question and the corresponding first interview answer; Inputting the first interview question and the first interview answer into a pre-trained relevance scoring model and outputting a relevance score, wherein the relevance scoring model is trained using a plurality of first sample data, wherein the first sample data includes interview questions and interview answers, and the first sample data is divided into positive samples and negative samples according to the relevance between the interview questions and the interview answers; Inputting the resume data and the first interview answer of the job seeker into a pre-trained credibility scoring model and outputting a credibility score, wherein the credibility scoring model is trained using a plurality of second sample data, wherein the second sample data includes the resume data and the interview answer, and according to the logical relationship between the resume data and the interview answer, the types of the second sample data are classified into: implication, contradiction and neutrality; Determine a comprehensive score based on the relevance score and the credibility score; If the comprehensive score is less than the preset threshold, the relevance score and credibility score are input into the large language model again, the second interview answer is output, and the comprehensive score of the first interview question and the second interview answer is calculated until the comprehensive score is greater than the preset threshold, and the first interview question and the corresponding interview answer are output.

2. The interview coaching method according to claim 1, characterized in that: The relevance scoring model is trained by using the relevance scores of the same interview question and the relevant interview answers and the irrelevant interview answers.

3. The interview coaching method according to claim 1, characterized in that: The training process of the relevance scoring model includes: Constructing a plurality of first sample data, wherein a positive sample includes an interview question and an interview answer related to the interview question, and a negative sample includes an interview question and an interview answer unrelated to the interview question; Indexing the interview questions and the interview answers in the first sample data to obtain a first sample sequence code, wherein the first sample sequence code includes a classification identifier for indicating whether the interview questions and the interview answers are related; Input the first sample sequence code into the relevance scoring model to output the relevance score; The model parameters in the relevance scoring model are updated by minimizing a first loss function value until the first loss function value is less than a first loss threshold, wherein the first loss function includes a relevance score of a positive sample and a relevance score of a negative sample.

4. The interview coaching method according to claim 3, characterized in that: The first loss function is expressed by the following formula: Among them, lossA is the value of the first loss function, s θ (q,a + ) is the correlation score of the positive sample, is the correlation score of the negative sample, and k is the number of negative samples.

5. The interview coaching method according to claim 1, characterized in that: The training process of the credibility scoring model includes: Constructing a plurality of second sample data, wherein the second sample data includes resume data and interview answers; Indexing the resume data and the interview answers in the second sample data to obtain a second sample sequence code, wherein the second sample sequence code includes a classification identifier for indicating a logical relationship between the resume data and the interview answers; Input the second sample sequence encoding into the credibility scoring model, output the unnormalized prediction value on each logical relationship and convert it into a probability distribution value on each logical relationship; The model parameters in the credibility scoring model are updated by minimizing the second loss function value until the second loss function value is less than a second loss threshold, wherein the second loss function includes the true type of the second sample data and the probability distribution value on each logical relationship.

6. The interview coaching method according to claim 5, characterized in that: The second loss function is expressed by the following formula: Among them, lossB is the value of the second loss function, p c is the probability that the sample data belongs to category C, C = 3, the categories are implication, contradiction and neutral, y c is the true label of the sample in category c. If the true category of the sample is c, then y c is 1, otherwise y c is 0.

7. The interview coaching method according to claim 5, characterized in that: Further including: The sum of the probability distribution values ​​of the resume data and interview answers in the implication and neutral categories respectively is taken as the credibility score.

8. An interview coaching device based on a large language model, characterized in that: include: An acquisition module is used to acquire the resume data of job seekers and target job demand data corresponding to the resume data; An interview question and answer generation module, used to input the resume data of the job seeker and the target position requirement data into the large language model and output the first interview question and the corresponding first interview answer; a relevance calculation module, configured to input the first interview question and the first interview answer into a pre-trained relevance scoring model and output a relevance score, wherein the relevance scoring model is trained using a plurality of first sample data, wherein the first sample data includes interview questions and interview answers, and the first sample data is divided into positive samples and negative samples according to the relevance between the interview questions and the interview answers; a credibility calculation module, configured to input the resume data and the first interview answer of the job seeker into a pre-trained credibility scoring model and output a credibility score, wherein the credibility scoring model is trained using a plurality of second sample data, wherein the second sample data includes the resume data and the interview answer, and according to the logical relationship between the resume data and the interview answer, the types of the second sample data are classified into: implication, contradiction and neutrality; A comprehensive score calculation module, used to determine the comprehensive score according to the relevance score and the credibility score; The question and answer optimization module is used to input the relevance score, credibility score and the first interview question into the large language model again if the comprehensive score is less than the preset threshold, optimize the first interview question, output the second interview answer, and calculate the comprehensive score of the first interview question and the second interview answer until the comprehensive score is greater than the preset threshold, and output the first interview question and the corresponding interview answer.

9. An electronic device, characterized in that: It includes a processor and a memory, wherein a computer program instruction set is stored in the memory, and when the processor executes the computer program instruction set in the memory, the interview coaching method based on a large language model as described in any one of claims 1-7 is implemented.

10. A computer-readable storage medium, characterized in that: in, The computer-readable storage medium stores a computer program instruction set, and when the computer program instruction set is executed by the processor, the interview coaching method based on a large language model described in any one of claims 1-7 is implemented.

Images