Self-adaptive AI interview method and device, electronic equipment and storage medium

Abstract

The invention provides a self-adaptive AI interview method and device, electronic equipment and a storage medium, and belongs to the technical field of artificial intelligence, in the method, the next operation (namely rule instructions, or strategy optimization behaviors, specifically next interview questions, or interview termination) can be dynamically adjusted according to interview answers, the dynamic adaptability is good, the method is more flexible, and in addition, the method can be applied to the field of AI interview. In the comprehensive scoring process, a mixed scoring strategy is adopted, the method is more scientific, when the current state does not meet the trigger condition in the preset rule, the offline reinforcement learning agent is adopted to make an intelligent decision on the current state, then the strategy optimization behavior is obtained, the accuracy is better, in addition, the preset rule is pre-configured, the controllability is good, and the user experience is improved. And the offline reinforcement learning agent is a conservative Q learning algorithm, and during training, Q value overestimation of behaviors outside distribution of a training data set can be inhibited through a conservative regularization item, so that non-compliant behaviors can be inhibited, namely, the behaviors are more controllable and compliant.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a self-adaptive AI interview method and device, electronic equipment and storage medium. BACKGROUND

[0002] Traditional manual interview process has inherent defects such as low efficiency, high cost, strong subjectivity of evaluation, and inconsistent standards.

[0003] Traditional automated interview tools aim to solve some problems, but still have the following shortcomings: (1) The system based on fixed scripts lacks flexibility and cannot conduct personalized interaction and in-depth evaluation according to the performance of the candidate; (2) Simple question and answer or FAQ systems are usually based on keyword or basic intent recognition and are difficult to understand the deep meaning of complex answers, with single evaluation dimension; (3) Traditional AI dialogue systems (such as customer service robots) usually do not have the multi-dimensional in-depth evaluation capability required for interview scenarios, and the online reinforcement learning (Online RL) strategy adopted by them has a high risk of online exploration problem if applied to interviews. Inappropriate exploration actions may result in a poor candidate experience, inaccurate evaluation results, and even cause enterprise reputation risks; (4) Pure end-to-end learning systems face the "black box" problem, and their decision-making process is difficult to explain and control, which cannot guarantee that the interview process strictly follows specific business rules and compliance requirements (such as question time limit and mandatory questions), which is unacceptable in serious interview scenarios.

[0004] In summary, traditional interview methods lack dynamic adaptability, are not flexible, have poor accuracy, and are uncontrollable. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a self-adaptive AI interview method, device, electronic equipment and storage medium to alleviate the technical problems of traditional interview methods lacking dynamic adaptability, being not flexible, having poor accuracy, and being uncontrollable.

[0006] In a first aspect, the embodiments of the present application provide a self-adaptive AI interview method, comprising: obtaining an interview answer input by an interviewee for a current interview question, and comprehensively scoring the interview answer by using a hybrid scoring strategy to obtain a comprehensive score vector, wherein the hybrid scoring strategy includes a semantic similarity scoring strategy, a fine-tuning model scoring strategy, and a keyword matching scoring strategy; updating a current state of a current interview session based on the comprehensive score vector, wherein the current interview session corresponds to a current entire interview process, and the current state comprises a current interview stage, a historical interview question feature, a comprehensive score sequence feature, an interview time consumption, an interviewer dynamic evaluation label, and an interview warning label; sequentially checking whether the current state meets a trigger condition in a preset rule according to a priority of the trigger condition in the preset rule; if the current state meets the trigger condition, executing a rule instruction corresponding to the trigger condition; if the current state does not meet the trigger condition, performing intelligent decision-making on the current state by an offline reinforcement learning agent to obtain a policy optimization behavior, and executing the policy optimization behavior to generate a next interview question or terminate the interview according to the rule instruction or the policy optimization behavior, if the next interview question is generated, taking the next interview question as the current interview question, returning to the step of obtaining the interview answer input by the interviewer for the current interview question, until all interview questions of all interview stages are completed or the interview is terminated, and then generating an interview report of the current interview session, wherein the offline reinforcement learning agent comprises a conservative Q-learning algorithm, and during training, the conservative Q-learning algorithm suppresses overestimation of Q values of behaviors outside a distribution of a training data set by a conservative regularization term.

[0007] Further, a hybrid scoring strategy is used to comprehensively score the interview answer, comprising: calculating semantic similarity between the interview answer and a preset reference answer by using a pre-trained language model to obtain a semantic score; scoring the interview answer by using a fine-tuned scoring model to obtain a quality score, wherein the fine-tuned scoring model is obtained by fine-tuning a BERT-based classification model on an interview data set, and the interview data set comprises questions, answers, and star-level qualities corresponding to each answer; matching the interview answer with a preset keyword list to obtain a keyword matching score; weighting and calculating the semantic score, the quality score, and the keyword matching score to obtain a comprehensive score of the interview answer, and determining the comprehensive score vector according to the comprehensive score of the interview answer and a comprehensive score corresponding to a historical interview question.

[0008] Further, the training process of the conservative Q-learning algorithm comprises: obtaining a historical interview interaction log, wherein the historical interview interaction log comprises historical semantic scores, historical quality scores, historical keyword matching scores, historical states, rule instructions corresponding to the historical states, or policy optimization behaviors; filtering the historical interview interaction logs based on historical states in the historical interview interaction logs to obtain filtered historical interview interaction logs; calculating a reward value of each of the filtered historical interview interaction logs according to historical semantic scores, historical quality scores, and historical keyword matching scores in the filtered historical interview interaction logs; periodically training the hybrid scoring strategy and the conservative Q-learning algorithm offline according to the filtered historical interview interaction logs and corresponding reward values.

[0009] Further, the method further comprises: obtaining a semantic score, a quality score, and a keyword matching score corresponding to a current historical interview question in the filtered historical interview interaction logs, and a semantic score, a quality score, and a keyword matching score corresponding to a previous historical interview question; According to the reward value calculation formula: calculating the reward value of the filtered historical interview interaction log corresponding to the current historical interview question, wherein, represents the reward value of the filtered historical interview interaction log corresponding to the current historical interview question, represents a weight corresponding to a score of a dimension, , represents a score of a dimension corresponding to the current historical interview question, represents a score of a dimension corresponding to the previous historical interview question, represents a score of a dimension corresponding to the previous historical interview question, represents a penalty threshold value, represents a penalty intensity coefficient, represents an indicator function.

[0010] Further, the preset rules include a preset mandatory rule, and the preset mandatory rule includes: if an interview time consumption of the current interview stage is greater than a preset time, or if a comprehensive score of a continuous preset number of times is less than a preset threshold value, the interview is terminated. When the strategy optimization behavior is a follow-up question, if the number of follow-up questions reaches a maximum number threshold value, the strategy optimization behavior is not performed.

[0011] Further, the method further comprises: ​If a next interview question is generated, a target interview question is obtained from the question library, and the target interview question is taken as the next interview question, wherein the target interview question is an interview question that meets the rule instruction or the strategy optimization behavior, and the target interview question is an interview question that has not been answered by the interviewee, and the target interview question has a correlation with the current interview question that is less than a preset correlation degree.

[0012] Further, the interview report includes information of the interviewee, comprehensive scores and interview answers corresponding to respective interview questions, an interview total score corresponding to the current entire interview process, and an interview conclusion.

[0013] In a second aspect, an apparatus for adaptive AI interview is also provided, and the apparatus includes: The obtaining and comprehensive scoring unit is configured to obtain an interview answer input by the interviewee for a current interview question, and perform comprehensive scoring on the interview answer by using a mixed scoring strategy to obtain a comprehensive score vector, wherein the mixed scoring strategy includes a semantic similarity scoring strategy, a fine-tuning model scoring strategy, and a keyword matching scoring strategy. The updating unit is configured to update a current state of a current interview session based on the comprehensive score vector, wherein the current interview session corresponds to the current entire interview process, and the current state includes a current interview stage, a historical interview question feature, a comprehensive score sequence feature, an interview time consumption, an interviewee dynamic evaluation label, and an interview warning label. The sequential checking unit is configured to sequentially check whether the current state meets a trigger condition in a preset rule according to a priority of the trigger condition in the preset rule. The first execution unit is configured to execute a rule instruction corresponding to the trigger condition if the current state meets the trigger condition. The second execution unit is configured to perform intelligent decision-making on the current state by using an offline reinforcement learning agent to obtain a strategy optimization behavior and execute the strategy optimization behavior, so as to generate a next interview question or terminate the interview according to the rule instruction or the strategy optimization behavior, wherein if the next interview question is generated, the next interview question is taken as the current interview question, and the step of obtaining the interview answer input by the interviewee for the current interview question is executed until interview questions in all interview stages are completed or the interview is terminated, and further, an interview report of the current interview session is generated, wherein the offline reinforcement learning agent includes a conservative Q-learning algorithm, and during training, the conservative Q-learning algorithm suppresses overestimation of Q values of behaviors outside a distribution of a training data set by using a conservative regularization term.

[0014] In a third aspect, an electronic device is provided, which comprises a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and the processor implements the steps of the method of any one of the first aspect when running the computer program.

[0015] In a fourth aspect, a computer readable storage medium is provided, which stores machine executable instructions, and the machine executable instructions, when invoked and run by a processor, cause the processor to run the method of any one of the first aspect.

[0016] In the embodiment of the present application, a method of adaptive AI interview is provided, comprising: obtaining an interview answer input by an interviewee for a current interview question, and comprehensively scoring the interview answer by using a mixed scoring strategy to obtain a comprehensive scoring vector, wherein the mixed scoring strategy comprises: a semantic similarity scoring strategy, a fine-tuning model scoring strategy, and a keyword matching scoring strategy; updating a current state of a current interview session based on the comprehensive scoring vector, wherein the current interview session corresponds to a current entire interview process, and the current state comprises: a current interview stage, a historical interview question feature, a comprehensive scoring sequence feature, an interview time consumption, an interviewee dynamic evaluation label, and an interview warning label; sequentially checking whether the current state meets a trigger condition in a preset rule according to a priority of each trigger condition in the preset rule; if yes, executing a rule instruction corresponding to the trigger condition; if not, intelligently deciding the current state by using an offline reinforcement learning agent to obtain a strategy optimization behavior, and executing the strategy optimization behavior to generate a next interview question according to the rule instruction or the strategy optimization behavior, or terminate the interview, if the next interview question is generated, taking the next interview question as the current interview question, returning to the step of obtaining the interview answer input by the interviewee for the current interview question, until all interview questions of all interview stages are completed or the interview is terminated, thereby generating an interview report of the current interview session, wherein the offline reinforcement learning agent comprises a conservative Q-learning algorithm, and during training, the conservative Q-learning algorithm suppresses overestimation of Q values of behaviors outside a distribution of a training data set by using a conservative regularization term. As can be seen from the above description, in the method of adaptive AI interview of the present application, the current state of the current interview session is automatically updated based on the comprehensive scoring vector of the interview answer of the interviewee, and then the rule instruction or the strategy optimization behavior corresponding to the current state is determined, that is, the method of the present application can dynamically adjust the next operation (i.e. the rule instruction or the strategy optimization behavior, specifically the next interview question or the termination of the interview) according to the interview answer, has good dynamic adaptability, is more flexible, in addition, the mixed scoring strategy is used during comprehensive scoring, is more scientific, when the current state does not meet the trigger condition in the preset rule, the offline reinforcement learning agent is also used to intelligently decide the current state to obtain the strategy optimization behavior, has better accuracy, furthermore, the preset rule is pre-configured, has good controllability, and the offline reinforcement learning agent is the conservative Q-learning algorithm, during training, overestimation of Q values of behaviors outside the distribution of the training data set is suppressed by using the conservative regularization term, so that the behaviors that do not meet the rules are suppressed, that is, are more controllable and meet the rules, thereby solving the technical problems of the traditional interview method, such as lack of dynamic adaptability, poor flexibility, poor accuracy, and uncontrollability. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the specific embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the specific embodiments or prior art description. Obviously, the drawings described below are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0018] Figure 1 A flow chart of a method for adaptive AI interview provided by an embodiment of the present application; Figure 2 A schematic diagram of a device for adaptive AI interview provided by an embodiment of the present application; Figure 3 A schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0019] The technical solutions of the present application will be described clearly and completely below in conjunction with embodiments. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0020] The traditional interview method lacks dynamic adaptability, is not good in flexibility, is poor in accuracy, and is uncontrollable.

[0021] Therefore, in the method for adaptive AI interview of the present application, the current state of the current interview session is automatically updated based on the comprehensive score vector of the interview answers of the interviewee, and then the rule instruction corresponding to the current state or the strategy optimization behavior is determined. That is, the method of the present application can dynamically adjust the next operation (i.e. the rule instruction or the strategy optimization behavior, specifically the next interview question or the termination of the interview) according to the interview answers, has good dynamic adaptability, is more flexible, and in addition, a mixed scoring strategy is used in the comprehensive scoring, which is more scientific. When the current state does not satisfy the trigger condition in the preset rule, an offline reinforcement learning agent is also used to make intelligent decisions on the current state, and then the strategy optimization behavior is obtained, which is better in accuracy. In addition, the preset rule is pre-configured, has good controllability, and the offline reinforcement learning agent is a conservative Q learning algorithm. During training, the conservative regularization term is used to suppress the overestimation of the Q value of the behavior outside the distribution of the training data set. In this way, the behavior that does not comply with the rules will be suppressed, that is, it is more controllable and compliant.

[0022] In order to facilitate the understanding of the present embodiment, first, a method for adaptive AI interview disclosed by the present embodiment will be described in detail.

[0023] Embodiment one: According to an embodiment of the present application, an embodiment of a method for adaptive AI interview is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown.

[0024] Figure 1 is a flowchart of a method for adaptive AI interview according to an embodiment of the present application, as shown, the method comprises the following steps: Figure 1 Step S102, obtaining the interview answer input by the interviewee for the current interview question, and adopting a mixed scoring strategy to comprehensively score the interview answer to obtain a comprehensive score vector, wherein the mixed scoring strategy comprises: a semantic similarity scoring strategy, a fine-tuning model scoring strategy, and a keyword matching scoring strategy. In an embodiment of the present application, the method for adaptive AI interview can be applied to intelligent automatic interview of human resource management or education training.

[0025] For example, when conducting online interview for the position of algorithm engineer, the interviewee X accesses the AI interviewer strategy management platform (i.e. adaptive AI interview system) through the web browser to participate in the online interview for the position of algorithm engineer. The whole interview process is cooperated by multiple modules, including: natural language understanding module (i.e. NLU), dialogue state tracker module (i.e. DST), rule engine module and offline reinforcement learning intelligent agent module (i.e. RL), the above system is deployed on the cloud server cluster, and the interviewee accesses the interview interface through the standard web browser for interaction. The system modules adopt micro-service architecture and communicate through API gateway.

[0026] The interview starts, the interviewee X logs in the platform and clicks the "start interview" button. System (i.e. platform) response: the backend API gateway receives the request of starting the interview, starts a new interview session, generates a unique session_id, the dialogue state tracker module initializes the current state S_t of the current interview session as: the current interview stage (current_stage) is opening, and stores it in Redis.

[0027] The system randomly selects an opening question (i.e. current interview question, such as "please briefly introduce yourself") from the question bank (MongoDB) and displays it to the interviewee X through the front end.

[0028] The interviewee answers the above current interview question as "I majored in computer science during my university period, focusing on algorithm design and optimization, and participated in XXX project".

[0029] ​The system obtains the interview answer, and a natural language understanding module adopts a mixed scoring strategy to comprehensively score the interview answer to obtain a comprehensive score vector.

[0030] The mixed scoring strategy includes a semantic similarity scoring strategy, a fine-tuning model scoring strategy, and a keyword matching scoring strategy. The semantic similarity scoring strategy uses a pre-trained language model ERNIE to calculate the semantic similarity between the interview answer and a preset reference answer. A BERT classification model that is fine-tuned based on an internally labeled interview dataset scores the interview answer to obtain a quality score. Meanwhile, the interview answer is matched with a corresponding preset keyword list to check whether the interview answer contains the corresponding keywords (such as "algorithm design" and "optimization"), and then a keyword matching score is obtained. Finally, the three scores are weighted and calculated to obtain the comprehensive score of the interview answer.

[0031] In step S104, a current state of a current interview session is updated based on the comprehensive score vector. The current interview session corresponds to the entire current interview process, and the current state includes a current interview stage, a historical interview question feature, a comprehensive score sequence feature, an interview time consumption, an interviewee dynamic evaluation label, and an interview warning label. Specifically, as exemplified above, the dialogue state tracker module updates the current state of the current interview session based on the comprehensive score vector, records the comprehensive score and timestamp of the current interview question in the current interview stage, and further obtains the current state of the current interview session, which includes: the current interview stage, the historical interview question feature, the comprehensive score sequence feature, the interview time consumption, the interviewer dynamic evaluation label, and the interview warning label. The historical interview question feature can be specifically TF-IDF feature (i.e., TF-IDF statistics are performed on the interviewed questions (i.e., historical interview questions) to obtain a high-dimensional sparse vector), average embedding feature (each historical interview question is encoded into a semantic vector by a pre-trained model (such as Sentence-BERT), and the arithmetic average of the semantic vectors of all historical interview questions is taken); the comprehensive score sequence feature can be specifically the comprehensive score of the current interview question, the average score (such as the average score of the current interview stage, and / or the average score of all interview answers), and the trend; the interview time consumption includes the total time consumption, the time consumption of the current interview stage, and the time consumption of the current interview question; the interviewer dynamic evaluation label includes the number of consecutive low scores, the average score trend, the suspected cheating label, the emotional fluctuation label, the follow-up acceptance label, the stage timeout label, and the number of repeated questions; and the interview warning label includes consecutive low scores (i.e., lower than a threshold), high repetition of answers (detection of copy-paste or templated answers), low keyword hit rate, timeout unanswer, frequent interruption request (such as multiple clicks on the “repeat question” or “pause interview” button), violation content label (such as containing discriminatory remarks or sensitive information in the answer), and identity verification failure (such as abnormal camera / microphone or low face matching degree).

[0032] In a specific implementation, a Redis Hash structure is used to store the current state S_t of each active interview session. The key is session_id. The storage content includes all fields in the aforementioned current state definition, and is updated atomically by the API gateway or a dedicated state management service after each interaction.

[0033] In step S106, the current state is sequentially checked according to the priority of each trigger condition in the preset rule. Specifically, the preset rule includes: trigger condition, rule instruction corresponding to each trigger condition, and priority corresponding to each trigger condition.

[0034] As exemplified above, the rule engine module checks the current state S_t and finds that the current_stage is “opening speech” and the answer_score (comprehensive score) is high (≥4.0). The following preset rule is satisfied.

[0035] Preset rule: if: - stage: "Opening Remarks" - answer_score>= 4.0 do: - transition_to_stage: "Basic algorithm" priority: 100 In this embodiment of the present invention, the rule engine module stores predefined deterministic rules (i.e., preset rules, such as IF-THEN structures, which include triggering conditions, rule instructions, and priorities). During each decision-making round, the rule engine module is first invoked. It reads the current state S_t in the DST and checks, in order of priority, whether any of the triggering conditions for the preset rules are met. If a matching rule is found, the deterministic action A_t (i.e., rule instruction) defined by that rule is output.

[0036] You can use an open-source rule engine library (such as the Drools Python interface or a self-developed dictionary-based matching engine). This allows administrators to define rules in the background using a graphical user interface or by writing rule files in YAML or JSON format. The rule library is bundled with the interview process template.

[0037] Example of a preset rule: "IF state.current_stage == 'Algorithm Basics' AND state.avg_score_stage<2.5 AND state.low_score_count>= 2 THEN action='ask_question(difficulty=easy, type=Basic)' =90".

[0038] Step S108: If the trigger condition is satisfied, the rule instruction corresponding to the trigger condition is executed; Specifically, as in the above example, the rule instructions corresponding to the trigger conditions are executed, that is, jumping to the next stage "basic algorithm".

[0039] Entering the basic algorithm stage, the system selects interview questions suitable for the "basic algorithm" stage from the question bank as the current interview question (such as "Please explain the time complexity of the binary search algorithm") and displays it to interviewee X.

[0040] Interviewee X enters interview answer: "The binary search algorithm has a time complexity of O(log n) because it cuts the search range in half each time." The natural language understanding module again comprehensively scores the interview answers and updates the current status in real time. Assume that the comprehensive score is 3.8 (slightly lower than expected). This means that the current status does not meet the trigger conditions in the preset rules.

[0041] Step S110, if it is not satisfied, an offline reinforcement learning agent is used to make an intelligent decision on the current state, obtain a strategy optimization behavior, and execute the strategy optimization behavior to generate the next interview question according to the rule instructions, or the strategy optimization behavior, or terminate the interview. If the next interview question is generated, the next interview question is used as the current interview question, and the step of obtaining the interview answer input by the interviewee for the current interview question is returned to execute until the interview questions of all interview stages are completed, or the interview is terminated, and then an interview report for the current interview session is generated. The offline reinforcement learning agent includes: a conservative Q learning algorithm. During training, the conservative Q learning algorithm suppresses the overestimation of the Q value of the behavior outside the distribution of the training data set through a conservative regularization term.

[0042] Specifically, as in the example above, an offline reinforcement learning agent is used to make intelligent decisions about the current state and determine the optimal strategy. In an interview scenario, if RL (reinforcement learning) online exploration mistakenly selects inappropriate or inefficient behaviors, it can lead to a poor interview experience for the interviewee and even risk the company's reputation. Offline RL eliminates the need for real-time trial and error; strategies can be trained using existing historical data (historical interview interaction logs). Offline datasets have already proven the viability of certain "good" or "compliant" behaviors, and the algorithm uses "conservative Q-learning" to minimize overestimation of unseen or high-risk behaviors. This significantly improves the safety of the interview process.

[0043] For example, because the overall score is low and does not reach the threshold (<4.0), the offline reinforcement learning agent chooses an action that reduces the difficulty (i.e., a policy optimization behavior, such as ask_question(difficulty=easy, type=basic)).

[0044] Then, the question bank is queried based on the strategy optimization behavior to obtain a list of questions that meet the strategy optimization behavior (such as "Please explain the basic principles of bubble sort"), and an interview question is randomly selected (i.e., the next interview question) and returned to interviewee X.

[0045] It should be noted that the offline reinforcement learning agent (i.e., the offline reinforcement learning agent module) contains a policy model π(a|s) trained using a reinforcement learning algorithm (such as conservative Q-learning). This offline reinforcement learning agent module is activated only when the rule engine module does not trigger any pre-set rules. It receives the current state S_t provided by the DST and, based on its policy model π(a|s), selects and outputs an adaptive action A_t with optimal or near-optimal probability (i.e., a policy-optimizing behavior).

[0046] The training data set can specifically be the filtered historical interview interaction log and its corresponding reward value. CQL can effectively reduce the probability of selecting uncertain behavior by penalizing the Q value of behavior lacking support in the training data set through conservative regularization terms, so that the system does not make rash or abnormal decisions in the interview process, and the safety and reliability of the strategy are improved. In the training, the interviewer dynamic evaluation label, interview warning label and reward value in the historical state are combined, so that the strategy inherently avoids bad behavior, which will be described in detail below. At the same time, the selectable behavior of RL can be dynamically constrained according to the current state S_t, further ensuring safety.

[0047] Continuing the above illustration, the above interaction and feedback are continuously performed, such as the interviewer X continues to answer the interview questions, and the system repeats the above process to dynamically adjust the question difficulty or jump stage according to the comprehensive score and the current state. After each answer, the system will update the current state, and select the next interview question according to the decision of the rule engine or offline reinforcement learning agent.

[0048] If the score is below the threshold (such as 2.5) for two consecutive stages, the rule engine triggers the rule instruction to terminate the interview, and the system generates an interview report and notifies the interviewer X.

[0049] In the embodiment of the present application, a method of adaptive AI interview is provided, comprising: obtaining an interview answer input by an interviewee for a current interview question, and comprehensively scoring the interview answer by using a mixed scoring strategy to obtain a comprehensive score vector, wherein the mixed scoring strategy comprises: a semantic similarity scoring strategy, a fine-tuning model scoring strategy, and a keyword matching scoring strategy; updating a current state of a current interview session based on the comprehensive score vector, wherein the current interview session corresponds to a current entire interview process, and the current state comprises: a current interview stage, a historical interview question feature, a comprehensive score sequence feature, an interview time consumption, an interviewee dynamic evaluation label, and an interview warning label; sequentially checking whether the current state meets a trigger condition in a preset rule according to a priority of each trigger condition in the preset rule; if yes, executing a rule instruction corresponding to the trigger condition; if not, intelligently deciding the current state by using an offline reinforcement learning agent to obtain a strategy optimization behavior, and executing the strategy optimization behavior to generate a next interview question according to the rule instruction or the strategy optimization behavior, or terminate the interview, if the next interview question is generated, taking the next interview question as the current interview question, returning to the step of obtaining the interview answer input by the interviewee for the current interview question, until the interview question of all interview stages is completed or the interview is terminated, thereby generating an interview report of the current interview session, wherein the offline reinforcement learning agent comprises a conservative Q-learning algorithm, and during training, the conservative Q-learning algorithm suppresses overestimation of Q values of behaviors outside a training data set distribution by using a conservative regularization term. As can be seen from the above description, in the method of adaptive AI interview, the current state of the current interview session is automatically updated based on the comprehensive score vector of the interview answer of the interviewee, and then the rule instruction or the strategy optimization behavior corresponding to the current state is determined, that is, the method of the present application can dynamically adjust the next operation (i.e. the rule instruction or the strategy optimization behavior, specifically the next interview question or the termination of the interview) according to the interview answer, has good dynamic adaptability, is more flexible, and in addition, during comprehensive scoring, the mixed scoring strategy is used, which is more scientific, when the current state does not meet the trigger condition in the preset rule, the offline reinforcement learning agent is also used to intelligently decide the current state to obtain the strategy optimization behavior, which is more accurate, in addition, the preset rule is pre-configured, has good controllability, and the offline reinforcement learning agent is the conservative Q-learning algorithm, which suppresses overestimation of Q values of behaviors outside a training data set distribution by using a conservative regularization term during training, so that irregular behaviors are suppressed, that is, more controllable and compliant, thereby solving the technical problems of lack of dynamic adaptability, poor flexibility, poor accuracy, and uncontrollability of traditional interview methods.

[0050] The above describes the method of adaptive AI interview of the present application briefly, and the specific contents involved therein are described in detail below.

[0051] In an optional embodiment of the present application, a mixed scoring strategy is adopted to comprehensively score the interview answers, specifically including the following steps: (1) The semantic similarity between the interview answers and the preset reference answers is calculated by using a pre-trained language model, and then a semantic score is obtained; Specifically, the pre-trained language model can be Sentence-BERT, and the pre-trained model (such as ERNIE or MacBERT) is used as the basis of Sentence-BERT to calculate the semantic similarity between the interview answers and the preset reference answers (which can include reference answers of multiple angles). The embedding vectors of the preset reference answers are pre-calculated and stored in a vector database (such as Milvus).

[0052] The semantic score can be specifically understood as a correctness score.

[0053] (2) A fine-tuned scoring model is used to score the interview answers to obtain a quality score, wherein the fine-tuned scoring model is obtained by fine-tuning a BERT-based classification model on an interview data set, and the interview data set includes: questions, answers, and star-level quality corresponding to each answer; Specifically, a BERT-based classification model is constructed, which is fine-tuned on an internally labeled interview data set (containing questions, answers, and 1-5 star quality ratings given by senior interviewers) to predict the quality level of the interview answers, and then the corresponding quality score is obtained.

[0054] The quality score can be specifically understood as a depth score.

[0055] (3) The interview answers are matched with the corresponding preset keyword list to obtain a keyword matching score; Specifically, the interview answers are matched with the corresponding preset keyword list to obtain a keyword hit rate, and then the keyword matching score is obtained.

[0056] The keyword matching score can be understood as a completeness score.

[0057] As can be seen, the present application quantitatively evaluates the interview answers from the aspects of correctness, depth and completeness of the interview answers.

[0058] (4) The semantic score, quality score and keyword matching score are weighted and calculated to obtain a comprehensive score of the interview answers, and a comprehensive score vector is determined according to the comprehensive score of the interview answers and the comprehensive score corresponding to the historical interview questions.

[0059] Specifically, the weight corresponding to the semantic score is set to 0.4, the weight corresponding to the quality score is set to 0.4, and the weight corresponding to the keyword matching score is set to 0.2. The weights can be configured by an administrator in the background according to different interview processes or question types.

[0060] In an optional embodiment of the present application, the training process of the conservative Q learning algorithm specifically includes the following steps: (1) Obtain a historical interview interaction log, wherein the historical interview interaction log includes a historical semantic score, a historical quality score, a historical keyword matching score, a historical state, a rule instruction corresponding to the historical state, or a strategy optimization behavior; Specifically, the API gateway or each service asynchronously sends detailed historical interview interaction logs to a Kafka message queue, which is consumed by a log processing service and stored in a database (such as AWS S3).

[0061] Use Apache Spark or Flink for offline data processing (state feature extraction, reward calculation, and data cleaning).

[0062] Implement the CQL algorithm (i.e., the conservative Q learning algorithm) using TensorFlow or PyTorch. The Q network uses a multi-layer perceptron (MLP) structure, with the input being the encoded historical state vector S_t in the historical interview interaction log and the output being the Q value of each discrete behavior (i.e., the strategy optimization behavior).

[0063] Historical state encoding: categorical features (such as interview stages) are encoded using One-hot encoding, numerical features (such as time and scores) are normalized, historical sequence features are encoded or statistical features are extracted using RNN or Transformer.

[0064] (2) Filter the historical interview interaction logs based on the historical states in each historical interview interaction log to obtain filtered historical interview interaction logs; Specifically, if the historical state includes a specific interviewer dynamic evaluation label or a specific interview warning label (such as high risk), the corresponding historical interview interaction log is deleted and filtered to obtain the filtered historical interview interaction log, i.e., before calling argmax Q(s, a), the behavior that is not allowed to be executed is filtered out.

[0065] (3) Calculate the reward value of each filtered historical interview interaction log according to the historical semantic score, the historical quality score, and the historical keyword matching score in each filtered historical interview interaction log; Specifically, according to the reward value design concept, the reward value R is calculated for each filtered historical interview interaction log record (s, a, s') in the offline data processing stage. A larger negative reward is given to the record marked as "interviewer forced intervention" or "candidate quits halfway" (included in the interview warning label); the "high-risk behavior label" (such as refusing the interview or extremely reducing the difficulty) is introduced, and a "significant negative reward" is set in the historical data, so as to automatically learn to avoid such behaviors in offline training; the punishment for "abnormal end" is strengthened, and the strategy is encouraged to learn safe and controllable behavior in offline training.

[0066] In a specific implementation: the reward value of each filtered historical interview interaction log is calculated according to the historical semantic score, the historical quality score, and the historical keyword matching score in each filtered historical interview interaction log, including the following steps: The semantic score, the quality score, and the keyword matching score corresponding to the current historical interview question in the filtered historical interview interaction log, and the semantic score, the quality score, and the keyword matching score corresponding to the previous historical interview question are obtained; the reward value calculation formula is as follows: The reward value of the filtered historical interview interaction log corresponding to the current historical interview question is calculated, wherein, represents the reward value of the filtered historical interview interaction log corresponding to the current historical interview question, represents the weight corresponding to the score of the dimension, , represents the score of the dimension corresponding to the current historical interview question, represents the score of the dimension corresponding to the previous historical interview question, represents a penalty threshold, represents a penalty intensity coefficient, represents an indicator function.

[0067] The above dimension includes: semantic score dimension, quality score dimension, and keyword matching score dimension, and the sum of all weights is 1. Different weights can be set for different posts or types (such as technical posts paying more attention to depth scores, and customer service posts paying more attention to integrity scores). The above penalty threshold can be 0.2, the above penalty intensity coefficient ∈ [0.5, 2], t represents the current historical interview question, t-1 represents the previous historical interview question, and when the condition is met, 1, when the condition is not met, 0.

[0068] (4) Periodically offline train the hybrid scoring strategy and conservative Q-learning algorithm according to the filtered historical interview interaction logs and the corresponding reward values.

[0069] Specifically, tens of thousands of filtered historical interview interaction logs collected offline are used as a training data set D. The Q network is trained using the Adam optimizer, and the target network parameters are updated regularly (e.g., every 10,000 steps). Hyperparameters such as α (CQL conservative coefficient), γ (discount factor), learning rate, etc. are adjusted on the validation set through grid search or Bayesian optimization. The trained model parameters are saved and deployed through a model service (such as TF Serving or TorchServe).

[0070] In implementation, the model training pipeline (using Kubeflow or Airflow orchestration) is triggered regularly (e.g., every week) to train the NLU and conservative Q-learning algorithm. After training, offline evaluation is automatically performed to generate an evaluation report. If the evaluation passes, the new model is pushed to the model library, and the model version of the online service is updated.

[0071] In an optional embodiment of the present application, the preset rules include a preset mandatory rule, and the preset mandatory rule includes: if the time consumption of the current interview stage is greater than a preset time, or if the comprehensive score of a continuous preset number of times is less than a preset threshold, the interview is terminated. When the strategy optimization behavior is to ask follow-up questions, if the number of follow-up questions reaches a maximum number threshold, the strategy optimization behavior is not performed. In this way, the balance between compliance and experience can be ensured.

[0072] The above hybrid structure (rule engine first matches, then calls RL agent) not only prioritizes calling rules, but also directly limits behaviors with lower scores or non-compliance at the RL behavior selection level.

[0073] In an optional embodiment of the present application, the method further includes the following steps: If the next interview question is generated, the target interview question is obtained from the question bank, and the target interview question is used as the next interview question, wherein the target interview question is an interview question that meets the rule instruction or the strategy optimization behavior, and the target interview question is an interview question that has not been answered by the interviewee, and the target interview question has a correlation degree less than a preset correlation degree with the current interview question.

[0074] Specifically, the question bank is stored in MongoDB and supports multi-dimensional index queries by type, difficulty, tag, and other dimensions. When receiving a rule instruction or a policy optimization action (such as ask_question(difficulty='hard', type='coding')), MongoDB is queried to obtain a list of questions that meet the conditions and have not been answered by the interviewee. The interview question can be selected and returned based on a specific strategy (such as randomness or low correlation with the previous question).

[0075] In an optional embodiment of the present invention, the interview report includes: information about the interviewee, comprehensive scores and interview answers corresponding to each interview question, the total interview score corresponding to the entire current interview process, and interview conclusions.

[0076] Specifically, when all interview questions in each stage are completed or the conditions for terminating the interview are triggered, the DST module marks the current state of the interview session as "Ended." The system summarizes the scores of the interview responses at each stage and generates a detailed interview report. The interview conclusions can include: "Failed the interview," "Passed the interview," and "Proceed to the next round of interviews."

[0077] The method of the present invention can combine deterministic preset rules with probabilistic intelligent strategies, which not only ensures that the core links and boundary conditions of the interview process are effectively controlled, but also can demonstrate high adaptability, flexibility and intelligence according to the real-time performance of the interviewee during the interaction process, thereby improving interview efficiency, evaluation depth and interviewee experience while ensuring standardization. It is a new AI interview method that can intelligently adapt to the interviewee's status, conduct in-depth evaluation and dynamic interaction while ensuring the standardization, controllability and safety of the core process. The focus is on the hierarchical decision-making of "rule priority judgment and RL adaptive supplement", as well as safe offline RL strategy learning based on hybrid NLU evaluation results and configurable reward functions.

[0078] Example 2: An embodiment of the present invention also provides an adaptive AI interview device, which is mainly used to execute the adaptive AI interview method provided in Example 1 of the present invention. The adaptive AI interview device provided in the embodiment of the present invention is specifically introduced below.

[0079] Figure 2 is a schematic diagram of an adaptive AI interview device according to an embodiment of the present invention, such as Figure 2 As shown, the device mainly includes: an acquisition and comprehensive scoring unit 10, an updating unit 20, a sequence checking unit 30, a first execution unit 40, and a second execution unit 50, wherein: The acquisition and comprehensive scoring unit is configured to acquire an interview answer input by the interviewee for a current interview question and comprehensively score the interview answer by using a mixed scoring strategy to obtain a comprehensive scoring vector, wherein the mixed scoring strategy comprises a semantic similarity scoring strategy, a fine-tuning model scoring strategy, and a keyword matching scoring strategy. The updating unit is configured to update a current state of a current interview session based on the comprehensive scoring vector, wherein the current interview session corresponds to a current entire interview process, and the current state comprises a current interview stage, a historical interview question feature, a comprehensive scoring sequence feature, an interview time consumption, an interviewee dynamic evaluation label, and an interview warning label. The sequential checking unit is configured to sequentially check whether the current state satisfies a trigger condition in a preset rule according to a priority corresponding to the trigger condition in the preset rule. The first execution unit is configured to execute a rule instruction corresponding to the trigger condition if the current state satisfies the trigger condition. The second execution unit is configured to intelligently decide the current state by using an offline reinforcement learning agent if the current state does not satisfy the trigger condition, to obtain a policy optimization behavior, and to execute the policy optimization behavior to generate a next interview question according to the rule instruction or the policy optimization behavior or to terminate the interview. If the next interview question is generated, the next interview question is taken as the current interview question, and the step of acquiring the interview answer input by the interviewee for the current interview question is executed until interview questions of all interview stages are completed or the interview is terminated. Then, an interview report of the current interview session is generated. The offline reinforcement learning agent comprises a conservative Q-learning algorithm. During training, the conservative Q-learning algorithm suppresses overestimation of a Q value of a behavior outside a distribution of a training data set by using a conservative regularization term.

[0080] In the embodiment of the present application, an adaptive AI interview device is provided, comprising: obtaining an interview answer input by an interviewee for a current interview question, and comprehensively scoring the interview answer by using a mixed scoring strategy to obtain a comprehensive scoring vector, wherein the mixed scoring strategy comprises: a semantic similarity scoring strategy, a fine-tuning model scoring strategy, and a keyword matching scoring strategy; updating a current state of a current interview session based on the comprehensive scoring vector, wherein the current interview session corresponds to a current entire interview process, and the current state comprises: a current interview stage, a historical interview question feature, a comprehensive scoring sequence feature, an interview time consumption, an interviewee dynamic evaluation label, and an interview warning label; sequentially checking whether the current state satisfies a trigger condition in a preset rule according to a priority of each trigger condition in the preset rule; if satisfied, executing a rule instruction corresponding to the trigger condition; if not satisfied, intelligently deciding the current state by using an offline reinforcement learning agent to obtain a strategy optimization behavior, and executing the strategy optimization behavior to generate a next interview question according to the rule instruction or the strategy optimization behavior, or terminate the interview, if the next interview question is generated, taking the next interview question as the current interview question, returning to the step of obtaining the interview answer input by the interviewee for the current interview question, until all interview questions of all interview stages are completed or the interview is terminated, thereby generating an interview report of the current interview session, wherein the offline reinforcement learning agent comprises a conservative Q-learning algorithm, and during training, the conservative Q-learning algorithm suppresses overestimation of Q values of behaviors outside a distribution of a training data set by using a conservative regularization term. As can be seen from the above description, in the adaptive AI interview device of the present application, the current state of the current interview session is automatically updated based on the comprehensive scoring vector of the interview answer of the interviewee, and then the rule instruction or the strategy optimization behavior corresponding to the current state is determined, that is, the method of the present application can dynamically adjust the next operation (i.e. the rule instruction or the strategy optimization behavior, specifically the next interview question or the termination of the interview) according to the interview answer, has good dynamic adaptability, is more flexible, and in addition, the mixed scoring strategy is used during comprehensive scoring, which is more scientific, when the current state does not satisfy the trigger condition in the preset rule, the offline reinforcement learning agent is also used to intelligently decide the current state to obtain the strategy optimization behavior, which has better accuracy, furthermore, the preset rule is pre-configured, has good controllability, and the offline reinforcement learning agent is the conservative Q-learning algorithm, which suppresses overestimation of Q values of behaviors outside a distribution of a training data set by using a conservative regularization term during training, so that irregular behaviors are suppressed, that is, more controllable and compliant, thereby solving the technical problems of traditional interview methods lacking dynamic adaptability, poor flexibility, poor accuracy, and uncontrollability.

[0081] Optionally, the obtaining and synthesizing score unit is further configured to: calculate semantic similarity between the interview answer and the preset reference answer by using a pre-trained language model, and then obtain a semantic score; score the interview answer by using a fine-tuned scoring model to obtain a quality score, wherein the fine-tuned scoring model is obtained by fine-tuning a BERT-based classification model on an interview data set, and the interview data set includes: questions, answers, and star-level quality corresponding to each answer; match the interview answer with a preset keyword list to obtain a keyword matching score; and calculate the semantic score, the quality score, and the keyword matching score by weighting to obtain a comprehensive score of the interview answer, and determine a comprehensive score vector according to the comprehensive score of the interview answer and the comprehensive score corresponding to the historical interview question.

[0082] Optionally, the device is further configured to: obtain a historical interview interaction log, wherein the historical interview interaction log includes: a historical semantic score, a historical quality score, a historical keyword matching score, a historical state, a rule instruction corresponding to the historical state, or a strategy optimization behavior; filter the historical interview interaction log based on the historical state in each historical interview interaction log to obtain a filtered historical interview interaction log; calculate a reward value of each filtered historical interview interaction log according to the historical semantic score, the historical quality score, and the historical keyword matching score in each filtered historical interview interaction log; and periodically train a hybrid scoring strategy and a conservative Q-learning algorithm offline according to the filtered historical interview interaction log and the reward value corresponding thereto.

[0083] Optionally, the device is further configured to: obtain the semantic score, the quality score, and the keyword matching score corresponding to the current historical interview question in the filtered historical interview interaction log, and the semantic score, the quality score, and the keyword matching score corresponding to the previous historical interview question; calculate a reward value of the current historical interview question according to the following formula:

[0084] ​​​​​​​​​​​​Optionally, the preset rules include: a preset mandatory rule, the preset mandatory rule includes: if an interview time consumption of a current interview stage is greater than a preset time, or if a comprehensive score of a continuous preset number of times is less than a preset threshold, then the interview is terminated; and when the strategy optimization behavior is asking a question, if a number of times of asking a question reaches a maximum number threshold, then the strategy optimization behavior is not performed.

[0085] Optionally, the device is further configured to: if a next interview question is generated, then obtaining a target interview question from the question bank, and taking the target interview question as the next interview question, wherein the target interview question is an interview question that meets the rule instruction or the strategy optimization behavior, and the target interview question is an interview question that has not been answered by the interviewee, and the target interview question has a correlation degree with the current interview question that is less than a preset correlation degree.

[0086] Optionally, the interview report includes: information of the interviewee, the comprehensive score and the interview answer corresponding to each interview question, an interview total score corresponding to the current entire interview process, and an interview conclusion.

[0087] The device provided by the embodiment of the application has the same implementation principle and generated technical effects as the method embodiment, and for brevity of description, the part not mentioned in the device embodiment can be referred to the corresponding content in the method embodiment.

[0088] As shown in Figure 3 The electronic device 600 provided by the embodiment of the application includes: a processor 601, a memory 602 and a bus, the memory 602 stores machine readable instructions executable by the processor 601, when the electronic device is running, the processor 601 and the memory 602 communicate through the bus, and the processor 601 executes the machine readable instructions to perform the steps of the method of the adaptive AI interview.

[0089] Specifically, the memory 602 and the processor 601 can be a general memory and processor, which are not specifically limited here, when the processor 601 runs the computer program stored in the memory 602, the method of the adaptive AI interview can be executed.

[0090] The processor 601 can be an integrated circuit chip having a processing capability of signals. In the implementation process, each step of the above method can be completed by the integrated logic circuit of hardware or the instruction in the form of software in the processor 601. The processor 601 described above can be a general processor, including a central processing unit (CPU), a network processor (NP), etc.; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. Each method, step and logic block diagram disclosed in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register or other mature storage medium in the art. The storage medium is located in the memory 602, and the processor 601 reads the information in the memory 602 and combines the hardware to complete the steps of the above method.

[0091] Corresponding to the above adaptive AI interview method, the embodiments of the present application also provide a computer readable storage medium, the computer readable storage medium stores computer executable instructions, when the processor calls and runs the computer executable instructions, the computer executable instructions make the processor run the steps of the above adaptive AI interview method.

[0092] The adaptive AI interview device provided by the embodiments of the present application can be specific hardware on the device or software or firmware installed on the device. The device provided by the embodiments of the present application has the same implementation principle and generated technical effects as the foregoing method embodiments. For the sake of brevity, the part of the device embodiment not mentioned in the foregoing method embodiment can be referred to the corresponding content in the foregoing method embodiment. Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above can be referred to the corresponding process in the foregoing method embodiment, which will not be described here.

[0093] In the embodiments of the present application, it should be understood that the disclosed apparatus and method can be implemented in other manners. The embodiments described above are merely specific implementation manners of the present application, and for example, the division of the units is only a logical function division, and there can be another division manner in actual implementation; for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different units, or the among different units, can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electric, mechanical or other forms.

[0094] For another example, the flowcharts and block diagrams in the drawings show the possible implementation architectures, functions and operations of the apparatus, method and computer program product according to the embodiments of the present application. In this regard, each block in the flowcharts or block diagrams can represent a module, a program segment or a part of code, which contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in different orders from that shown in the drawings. For example, two consecutive blocks can actually be executed substantially in parallel, and sometimes they can be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and the combination of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system for performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0095] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. they can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the present embodiment.

[0096] In addition, each functional unit in the embodiments of the present application can be integrated into one processing unit, or each unit can exist physically as a separate unit, or two or more units can be integrated into one unit.

[0097] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing an electronic device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the adaptive AI interview method described in various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0098] It should be noted that similar reference numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. In addition, the terms "first", "second", "third" and the like are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0099] Finally, it should be noted that the above-described embodiments are only specific implementations of the present application, which are used to illustrate the technical solutions of the present application, but not to limit them. The protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily think of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed by the present application, or make equivalent replacements to some technical features. These modifications, changes or replacements do not cause the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application. They should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for adaptive AI interviewing, characterized in that: include: Obtaining the interview answer input by the interviewee for the current interview question, and comprehensively scoring the interview answer using a hybrid scoring strategy to obtain a comprehensive scoring vector, wherein the hybrid scoring strategy includes: a semantic similarity scoring strategy, a fine-tuning model scoring strategy, and a keyword matching scoring strategy; Updating the current state of the current interview session based on the comprehensive score vector, wherein the current interview session corresponds to the entire current interview process, and the current state includes: the current interview stage, historical interview question features, comprehensive score sequence features, interview duration, interviewer dynamic evaluation tags, and interview warning tags; According to the priority corresponding to each trigger condition in the preset rule, sequentially check whether the current state meets the trigger condition in the preset rule; If it is satisfied, the rule instruction corresponding to the trigger condition is executed; If it is not satisfied, an offline reinforcement learning agent is used to make an intelligent decision on the current state, obtain a policy optimization behavior, and execute the policy optimization behavior to generate the next interview question according to the rule instruction, or the policy optimization behavior, or terminate the interview. If the next interview question is generated, the next interview question is used as the current interview question, and the step of obtaining the interview answer input by the interviewee for the current interview question is returned to execute until the interview questions of all interview stages are completed, or the interview is terminated, and then an interview report for the current interview session is generated. The offline reinforcement learning agent includes: a conservative Q learning algorithm. During training, the conservative Q learning algorithm suppresses the overestimation of Q values ​​of behaviors outside the distribution of the training data set through a conservative regularization term.

2. The method according to claim 1, characterized in that The interview responses were scored comprehensively using a mixed scoring strategy, including: Using a pre-trained language model to calculate the semantic similarity between the interview answer and the preset reference answer, and then obtain a semantic score; Scoring the interview answers using a fine-tuned scoring model to obtain a quality score, wherein the fine-tuned scoring model is obtained by fine-tuning a BERT-based classification model on an interview dataset, wherein the interview dataset includes: questions, answers, and the star quality corresponding to each answer; Matching the interview answers with the corresponding preset keyword list to obtain a keyword matching score; The semantic score, the quality score and the keyword matching score are weightedly calculated to obtain a comprehensive score of the interview answer, and the comprehensive score vector is determined according to the comprehensive score of the interview answer and the comprehensive score corresponding to the historical interview question.

3. The method according to claim 2, characterized in that The training process of the conservative Q learning algorithm includes: Obtaining historical interview interaction logs, wherein the historical interview interaction logs include: historical semantic scores, historical quality scores, historical keyword matching scores, historical states, rule instructions corresponding to the historical states, or strategy optimization behaviors; Filtering the historical interview interaction logs based on the historical status in each of the historical interview interaction logs to obtain filtered historical interview interaction logs; Calculate the reward value of each filtered historical interview interaction log according to the historical semantic score, historical quality score, and historical keyword matching score in each filtered historical interview interaction log; The hybrid scoring strategy and the conservative Q learning algorithm are periodically trained offline according to the filtered historical interview interaction logs and their corresponding reward values.

4. The method according to claim 3, characterized in that The reward value of each filtered historical interview interaction log is calculated based on the historical semantic score, historical quality score, and historical keyword matching score in each filtered historical interview interaction log, including: Obtain the semantic score, quality score, and keyword matching score corresponding to the current historical interview question and the semantic score, quality score, and keyword matching score corresponding to the previous historical interview question in the filtered historical interview interaction log; Calculation formula based on reward value: Calculate the reward value of the filtered historical interview interaction log corresponding to the current historical interview question, where: Indicates the reward value of the filtered historical interview interaction log corresponding to the current historical interview question, express The weight corresponding to the dimension score, , Indicates the current historical interview question corresponding to The score of the dimension, Indicates the corresponding historical interview question The score of the dimension, represents the penalty threshold, represents the penalty intensity coefficient, Represents the indicator function.

5. The method according to claim 1, wherein The preset rules include: preset mandatory rules, wherein the preset mandatory rules include: if the interview duration of the current interview stage exceeds the preset time, or the comprehensive score of the preset number of consecutive times is less than the preset threshold, then the interview is terminated; When the strategy optimization behavior is asking follow-up questions, if the number of follow-up questions reaches a maximum threshold, the strategy optimization behavior is not performed.

6. The method according to claim 1, characterized in that The method further comprises: If the next interview question is generated, the target interview question is obtained from the question bank, and the target interview question is used as the next interview question, wherein the target interview question is an interview question that complies with the rule instructions, or the strategy optimization behavior, and the target interview question is an interview question that the interviewee has not answered, and the correlation between the target interview question and the current interview question is less than the preset correlation.

7. The method according to claim 1, characterized in that The interview report includes: the interviewee's information, the comprehensive score and interview answers corresponding to each interview question, the total interview score corresponding to the entire current interview process, and the interview conclusion.

8. An adaptive AI interview device, characterized in that: include: An acquisition and comprehensive scoring unit is used to acquire the interview answer input by the interviewee for the current interview question, and comprehensively score the interview answer using a hybrid scoring strategy to obtain a comprehensive scoring vector, wherein the hybrid scoring strategy includes: a semantic similarity scoring strategy, a fine-tuning model scoring strategy, and a keyword matching scoring strategy; an updating unit, configured to update a current state of a current interview session based on the comprehensive scoring vector, wherein the current interview session corresponds to the entire current interview process, and the current state includes: a current interview stage, historical interview question features, comprehensive scoring sequence features, interview duration, interviewer dynamic evaluation tags, and interview warning tags; A sequential checking unit, configured to sequentially check whether the current state satisfies the triggering conditions in the preset rules according to the priorities corresponding to the triggering conditions in the preset rules; A first execution unit, configured to execute a rule instruction corresponding to the trigger condition if the trigger condition is met; The second execution unit is used to use an offline reinforcement learning agent to make intelligent decisions on the current state if the conditions are not met, obtain policy optimization behavior, and execute the policy optimization behavior to generate the next interview question according to the rule instructions, or the policy optimization behavior, or terminate the interview. If the next interview question is generated, the next interview question is used as the current interview question, and the step of obtaining the interview answer input by the interviewee for the current interview question is returned to execute until the interview questions of all interview stages are completed, or the interview is terminated, and then an interview report for the current interview session is generated. The offline reinforcement learning agent includes: a conservative Q learning algorithm. During training, the conservative Q learning algorithm suppresses the overestimation of Q values ​​of behaviors outside the distribution of the training data set through a conservative regularization term.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to execute the method according to any one of claims 1 to 7.

Images