Dialogue model training method, artificial intelligence interview method, computing device, storage medium and program product

Abstract

The embodiment of the invention provides a dialogue model training method, an artificial intelligence interview method, computing equipment, a computer readable storage medium and a computer program product. The dialogue model training method comprises the steps of determining a first dialogue model and a second dialogue model; taking the first dialogue model as a student model and the second dialogue model as a teacher model; acquiring first sample data; inputting the first sample data into a student model; splicing the target element data with the first sample data to obtain second sample data; respectively inputting the second sample data into a student model and a teacher model; determining difference information of the first probability distribution and the second probability distribution corresponding to different element data; and adjusting the student model based on the difference information to obtain a target dialogue model. According to the technical scheme provided by the embodiment of the invention, the training difficulty of the student model is reduced, and the performance of the target dialogue model generated by training is improved.

Description

Technical Field

[0001] Embodiments of the present application relate to the field of artificial intelligence technology, and in particular, to a method for training a dialogue model, an artificial intelligence interview method, a computing device, a computer-readable storage medium, and a computer program product. Background Art

[0002] In today's digital age, artificial intelligence technology is increasingly widely used in the recruitment field, especially the AI (Artificial Intelligence) interview function, which brings great convenience to both enterprises and job seekers.

[0003] AI interviews usually need to be implemented using a dialogue model. The dialogue model can ask questions to job seekers, and after obtaining the response data of the job seekers, continue to ask questions to the job seekers according to the response data, and complete the interview through dialogue.

[0004] In related technologies, knowledge distillation can be used to train the dialogue model. Knowledge distillation, simply speaking, is a model compression and transfer learning technology. Its core idea is to transfer the knowledge of a complex and powerful teacher model to a relatively simple student model, that is, the dialogue model.

[0005] The inventor found in the process of implementing the concept of the present application that in the training method based on knowledge distillation in related technologies, by processing the sample data respectively by the teacher model and the student model, the probability distributions of different element data are obtained respectively, and then the student model is trained according to the difference of the probability distributions. It requires the student model to strictly fit according to the output distribution of the teacher model, and this forced fitting may exceed the learning ability range of the student model, resulting in poor performance of the trained dialogue model. Summary of the Invention

[0006] Embodiments of the present application provide a method for training a dialogue model, an artificial intelligence interview method, a computing device, a computer-readable storage medium, and a computer program product.

[0007] In a first aspect, a method for training a dialogue model provided in an embodiment of the present application includes:

[0008] Determine a first dialogue model and a second dialogue model, where the number of parameters of the second dialogue model is greater than the number of parameters of the first dialogue model;

[0009] Use the first dialogue model as the student model and the second dialogue model as the teacher model;

[0010] Obtain first sample data; the first sample data includes at least one round of sample dialogue data;

[0011] Input the first sample data into the student model to perform an inference operation based on the first sample data using the student model, calculate the first probability distribution of different element data, and select target element data according to the first probability distribution; the element data is the smallest data unit.

[0012] Concatenate the target element data with the first sample data to obtain second sample data.

[0013] Input the second sample data into the student model and the teacher model respectively, to perform an inference operation based on the second sample data using the student model to calculate the first probability distribution of different element data, and perform an inference operation based on the second sample data using the teacher model to calculate the second probability distribution of different element data.

[0014] Determine the difference information between the first probability distribution and the second probability distribution corresponding to different element data.

[0015] Adjust the student model based on the difference information to transfer the knowledge of the second dialogue model to the first dialogue model, and obtain a target dialogue model; the target dialogue model is used to generate target question data from a question list corresponding to a target position in combination with at least one round of historical dialogue data.

[0016] In a second aspect, an artificial intelligence interview method provided in an embodiment of the present application is applied to a server, and the method includes:

[0017] Obtain an interview request sent by a client; the interview request is generated by the client in response to an interview operation triggered by a target user for a target position.

[0018] Determine a question list corresponding to the target position; the question list includes a plurality of question data.

[0019] Determine a target dialogue model; the target dialogue model is obtained by transferring the knowledge of a second dialogue model to a first dialogue model; the first dialogue model is used as a student model, and the second dialogue model is used as a teacher model; the student model is used to be adjusted based on the difference information between the first probability distribution and the second probability distribution of different element data generated by inference; the first probability distribution is generated by the student model based on the second sample data by inference; the second probability distribution is generated by the teacher model based on the second sample data by inference; the second sample input model is obtained by concatenating the first sample data with the target element data; the target element data is generated by the student model based on the first sample data by inference.

[0020] Generate target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data;

[0021] Send the target question data to the client;

[0022] Obtain the response data fed back by the client, and return to the step of generating target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data and continue to execute.

[0023] Thirdly, an apparatus for training a dialogue model is provided in an embodiment of the present application, including:

[0024] A first determination module, configured to determine a first dialogue model and a second dialogue model, where the number of parameters of the second dialogue model is greater than that of the first dialogue model;

[0025] A second determination module, configured to use the first dialogue model as a student model and the second dialogue model as a teacher model;

[0026] A sample acquisition module, configured to acquire first sample data; the first sample data includes at least one round of sample dialogue data;

[0027] A first input module, configured to input the first sample data into the student model, so as to use the student model to perform an inference operation based on the first sample data, calculate a first probability distribution of different element data, and select target element data according to the first probability distribution; the element data is the smallest data unit;

[0028] A splicing module, configured to splice the target element data with the first sample data to obtain second sample data;

[0029] A second input module, configured to input the second sample data into the student model and the teacher model respectively, so as to use the student model to perform an inference operation based on the second sample data to calculate a first probability distribution of different element data, and use the teacher model to perform an inference operation based on the second sample data to calculate a second probability distribution of different element data;

[0030] A difference determination module, configured to determine difference information between the first probability distribution and the second probability distribution corresponding to different element data;

[0031] A training module, configured to adjust the student model based on the difference information to transfer the knowledge of the second dialogue model to the first dialogue model, and obtain a target dialogue model; the target dialogue model is used to generate target question data from a question list corresponding to a target position in combination with at least one round of historical dialogue data.

[0032] In a fourth aspect, an artificial intelligence interview device is provided in an embodiment of the present application, which is applied to a server side. The method includes:

[0033] A first request acquisition module, configured to acquire an interview request sent by a client; the interview request is generated by the client in response to an interview operation triggered by a target user for a target position;

[0034] A first question list determination module, configured to determine a question list corresponding to the target position; the question list includes a plurality of question data;

[0035] A dialogue model determination module, configured to determine a target dialogue model; the target dialogue model is obtained by migrating the knowledge of a second dialogue model to a first dialogue model; the first dialogue model is used as a student model, and the second dialogue model is used as a teacher model; the student model is used to adjust based on the difference information between a first probability distribution and a second probability distribution of different element data generated by inference; the first probability distribution is generated by the student model based on second sample data by inference; the second probability distribution is generated by the teacher model based on second sample data by inference; the second sample input model is obtained by splicing first sample data with target element data; the target element data is generated by the student model based on the first sample data;

[0036] A first question generation module, configured to generate target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data;

[0037] A first question sending module, configured to send the target question data to the client;

[0038] A first response acquisition module, configured to acquire response data fed back by the client, and return to continue executing the step of generating target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data.

[0039] In a fifth aspect, a computing device is provided in an embodiment of the present application, including a processing component and a storage component;

[0040] The storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement the training method of the dialogue model provided in the embodiment of the present application, or to implement the artificial intelligence interview method provided in the embodiment of the present application.

[0041] Sixth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processing component, it implements the training method of the dialogue model provided by the embodiment of the present application, or implements the artificial intelligence interview method provided by the embodiment of the present application.

[0042] Seventh aspect, an embodiment of the present application provides a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processing component, it implements the training method of the dialogue model provided by the embodiment of the present application, or implements the artificial intelligence interview method provided by the embodiment of the present application.

[0043] In an embodiment of the present application, a first dialogue model and a second dialogue model with different numbers of parameters are first determined. The first dialogue model is used as the student model, and the second dialogue model is used as the teacher model. Then, first sample data including at least one round of sample dialogue data is obtained and input into the student model for inference operations. The first probability distribution of different element data is calculated and the target element data is selected. The target element data is concatenated with the first sample data to obtain second sample data. Then, the second sample data is respectively input into the student model and the teacher model, and the first probability distribution and the second probability distribution of different element data are respectively calculated. Next, the difference information between the first probability distribution and the second probability distribution of different element data is determined. Finally, the student model is adjusted based on the difference information, and the knowledge of the second dialogue model is transferred to the first dialogue model to obtain the target dialogue model. By comparing the difference information between the outputs of the student model and the teacher model after the student model already has its own output, and guiding the student model to adjust in a better direction according to the difference information, the training difficulty of the student model is reduced, and the performance of the target dialogue model generated by training is improved.

[0044] These aspects or other aspects of the present application will be more clearly understood in the following description of the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following briefly introduces the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0046] Figure 1 Shows a system architecture diagram to which a technical solution of an embodiment of the present application can be applied;

[0047] Figure 2 Shows a flowchart of a training method of a dialogue model provided by an embodiment of the present application;

[0048] Figure 3 The flowchart of an artificial intelligence interview method provided in an embodiment of the present application is shown;

[0049] Figure 4 The schematic diagram of the artificial intelligence interview method provided in the embodiment of the present application is shown;

[0050] Figure 5 The block diagram of a training device for a dialogue model provided in an embodiment of the present application is shown;

[0051] Figure 6 The block diagram of an artificial intelligence interview device provided in the embodiment of the present application is shown;

[0052] Figure 7 The block diagram of a computing device provided in an embodiment of the present application is shown. Detailed implementation manners

[0053] In order to enable those skilled in the art to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application.

[0054] In some processes described in the specification and claims of the present application and the above-mentioned drawings, a plurality of operations appear in a specific order. However, it should be clearly understood that these operations may not be executed in the order in which they appear in this article or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions such as "first" and "second" in this article are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, nor do they limit that "first" and "second" are of different types.

[0055] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or fully authorized by all parties. And the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of relevant countries and regions, and corresponding operation entrances are provided for users to choose to authorize or refuse.

[0056] It should be noted that the technical solution of the embodiment of the present application is applicable to a network virtual environment. Generally, the described user refers to a "virtual user". A real user can register a user account on the server through registration to obtain a user identity in the network environment. In the embodiment of the present application, the same user account can be logged in to the server through different types of clients, so that the server can identify the same user. Of course, different user accounts can also be logged in to the server through different types of clients. The server stores the binding relationships of different user accounts, so that different user accounts with binding relationships can be considered as the same user.

[0057] In today's digital age, the application of artificial intelligence technology in the recruitment field is becoming increasingly widespread, especially the AI (Artificial Intelligence) interview function, which brings great convenience to both enterprises and job seekers.

[0058] AI interviews usually need to be implemented using a dialogue model. The dialogue model can ask questions to job seekers and, after obtaining the response data of the job seekers, continue to ask questions to the job seekers based on the response data to complete the interview through a dialogue.

[0059] In the related art, a method based on knowledge distillation of forward KL divergence can be used to train the dialogue model. Knowledge distillation, simply speaking, is a model compression and transfer learning technology. Its core idea is to transfer the knowledge of a complex and powerful teacher model to a relatively simple student model, that is, the dialogue model.

[0060] In the training process using forward KL divergence, the sample data can be processed by the teacher model and the student model respectively to obtain their probability distributions for different element data. The probability distribution of the teacher model can be denoted as P(x), and that of the student model as Q(x). Then, for each element data, the forward KL divergence can be calculated based on P(x) and Q(x). The forward KL divergence value can be used to quantify the difference between P(x) and Q(x), and the larger its value, the greater the difference between the two distributions. In model training, the forward KL divergence value can be used as a loss function, and the parameters of the student model can be adjusted by minimizing the forward KL divergence value to make the output distribution of the student model as close as possible to the output distribution of the teacher model.

[0061] In the process of implementing the concept of this application, the inventor found that when there is a large difference in the data output distribution between the student model and the teacher model, the forward KL divergence makes it extremely difficult for the student model to learn. Because it requires the student model to strictly fit the output distribution of the teacher model, and this forced fitting may exceed the learning ability range of the student model. For example, in complex natural language processing tasks, the teacher model may have trained a very fine output distribution based on large-scale data, but due to its relatively limited own structure and training data, it is difficult for the student model to directly learn such a complex distribution, resulting in the obstruction of the learning process.

[0062] To solve the technical problems existing in the related art, in the application embodiment, first, a first dialogue model and a second dialogue model with different numbers of parameters are determined. The first dialogue model is used as the student model, and the second dialogue model is used as the teacher model. Then, first sample data including at least one round of sample dialogue data is obtained and input into the student model for inference operations. The first probability distribution of different element data is calculated and target element data is selected. The target element data is concatenated with the first sample data to obtain second sample data. Then, the second sample data is respectively input into the student model and the teacher model, and the first probability distribution and the second probability distribution of different element data are respectively calculated. Next, the difference information between the first probability distribution and the second probability distribution of different element data is determined. Finally, based on the difference information, the student model is adjusted, and the knowledge of the second dialogue model is transferred to the first dialogue model to obtain the target dialogue model. Through the technical solution provided by the application embodiment, training can be performed based on the output of the student model. The teacher model does not provide a probability distribution for the student model to fit as in the forward KL divergence. Instead, after the student model already has its own output, by comparing the differences between the two, the student model is guided to adjust in a better direction, reducing the training difficulty of the student model and improving the performance of the trained target dialogue model.

[0063] Next, the technical solutions in the embodiments of the present application will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0064] Figure 1 A system architecture diagram to which the technical solution of an embodiment of the present application can be applied is shown. The system architecture may include a client 101 and a server 102.

[0065] Among them, a connection is established between the client 101 and the server 102 through a network. The network provides the medium for the communication link between the client 101 and the server 102. The network can include various connection types, such as wired, wireless communication links, or fiber optic cables, etc. Optionally, the server can communicate with the client through a mobile network. Optionally, the client can also establish a communication connection with the server through methods such as Bluetooth, WiFi, infrared, etc.

[0066] The client 101 can interact with the server 102 through the network to receive or send messages, etc.

[0067] Among them, the client 101 can be a browser, an APP (Application), or a web application such as an H5 (HyperText Markup Language 5) application, or a light application (also known as a mini-program, a lightweight application program), or a cloud application, etc. The client 101 can be deployed in an electronic device and needs to rely on the device or certain apps in the device to run, etc. The electronic device can, for example, have a display screen and support information browsing, etc., such as a personal mobile terminal such as a mobile phone, a tablet computer, a personal computer, a desktop computer, a smart speaker, a smart watch, etc. For the sake of easy understanding, Figure 1 the client is mainly represented by the image of the device in the text. Various other types of applications can usually be configured in the electronic device, such as human-computer dialogue applications, model training applications, text processing applications, web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc. The electronic device can refer to a device used by a user and having functions such as computing, Internet access, and communication required by the user, such as a mobile phone, a tablet computer, a personal computer, a wearable device, etc. The electronic device usually can include at least one processing component and at least one storage component. The electronic device may also include basic configurations such as a network card chip, an IO bus, audio-video components, etc., which are not limited in this application. Optionally, according to the implementation form of the electronic device, some peripheral devices can also be included, such as a keyboard, a mouse, a stylus, a printer, etc., which are not limited in this application.

[0068] The server 102 can include servers that provide various services, such as a server for background training that provides support for the models used on the client 101, or a server that processes the interaction information sent by the client, etc.

[0069] It should be noted that the server 102 can be implemented as a distributed server cluster composed of multiple servers, or as a single server. The server can also be a server of a distributed system, or a server combined with a blockchain. The server can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms, or an intelligent cloud computing server or intelligent cloud host with artificial intelligence technology.

[0070] It should be noted that the training method of the dialogue model and the artificial intelligence interview method provided in the embodiments of the present application are generally executed by the server 102, and the corresponding training test device and artificial intelligence interview device of the dialogue model are generally set in the server 102. However, in other embodiments of the present application, the client 101 can also have a similar function to the server 102, so as to execute the training method of the dialogue model and the artificial intelligence interview method provided in the embodiments of the present application. In other embodiments, the training method of the dialogue model and the artificial intelligence interview method provided in the embodiments of the present application can also be jointly executed by the client 101 and the server 102.

[0071] It should be understood that Figure 1 the number of clients and servers in

[0072] The implementation details of the technical solutions of the embodiments of the present application are elaborated in detail below.

[0073] Figure 2 The flowchart of a training method of a dialogue model provided by an embodiment of the present application is shown, as Figure 2 shown, the method may specifically include the following steps:

[0074] 201: Determine a first dialogue model and a second dialogue model, where the number of parameters of the second dialogue model is greater than that of the first dialogue model.

[0075] 202: Use the first dialogue model as the student model and the second dialogue model as the teacher model.

[0076] Among them, the first dialogue model can be a model with relatively few parameters and more focused on specific dialogue tasks. In the embodiments of the present application, the specific dialogue tasks may include dialogue tasks in an interview scenario.

[0077] The second dialogue model can have a large number of parameters. Thus, the second dialogue model can be a more complex and powerful model, which can be trained on a wider range of datasets, have more abundant knowledge, and stronger processing capabilities.

[0078] In the embodiments of the present application, the Knowledge Distillation technology can be used to train the first dialogue model. In the framework of knowledge distillation, the first dialogue model can be regarded as a student model, and the student model can learn knowledge and skills from another more powerful model, that is, the teacher model. In the embodiments of the present application, the second dialogue model with a larger number of parameters can be used as the teacher model.

[0079] Through knowledge distillation, the teacher model can use its rich parameters and extensive training knowledge to provide guidance for the student model, thereby improving the performance of the student model on specific tasks while maintaining the advantages of a relatively simple structure, such as faster inference speed and lower resource consumption.

[0080] 203: Obtain the first sample data; the first sample data includes at least one round of sample dialogue data.

[0081] Among them, in the first sample data, at least one round of sample dialogue data can be included, and the at least one round of sample dialogue data can be arranged in the dialogue order.

[0082] In a possible implementation manner, the at least one round of sample dialogue data can be obtained by collecting the dialogue data generated during at least one historical interview process. The sample dialogue data can record the dialogue situation between the interviewer and the job-seeking user in the real interview scenario.

[0083] In the actual dialogue scenario, the order of the dialogue will affect the semantic understanding and logical coherence. For example, in the interview scenario, the interviewer may further ask relevant questions based on the previous answers of the job-seeking user, or adjust the direction and depth of the questions according to the progress of the interview. This sequentiality enables the first dialogue model to learn the natural flow of the dialogue and the context relevance.

[0084] 204: Input the first sample data into the student model to use the student model to perform an inference operation based on the first sample data, calculate the first probability distribution of different element data, and select the target element data according to the first probability distribution; the element data is the smallest data unit.

[0085] After obtaining the first sample data, the first sample data can be input into the student model.

[0086] After receiving the first sample data, the student model can perform an inference operation. In an embodiment of the present application, the inference operation may refer to the process in which the student model processes the first sample data based on its internal neural network structure and training parameters to calculate the first probability distribution of different element data.

[0087] In an embodiment of the present application, the element data may be the smallest data unit. For example, in a natural language processing task, the element data may be a single character, word, morpheme, or other smallest semantic unit. The student model can calculate the first probability distribution of these element data.

[0088] Among them, the first probability distribution can represent the possibility of each element data appearing under the current input. For example, for a dialogue generation task, if the element data is a word, the student model can calculate the probability distribution of the next possible word based on the first sample data. For example, assuming that the input of the student model is "Please describe your in the project", the student model may calculate the probability distribution of a series of words, such as the probability of "experience" being 0.3, the probability of "role" being 0.25, the probability of "contribution" being 0.2, etc. This probability distribution is obtained based on the language patterns and semantic relationships learned by the student model during training, and reflects what the student model thinks the next element data may be under the current input.

[0089] According to the calculated first probability distribution, the student model can select the target element data. Among them, when the student model selects the target element data from different element data, for example, selection strategies such as a greedy strategy and a random sampling strategy can be adopted. In the greedy strategy, the student model can select the element data with the highest probability as the target element data. For example, in the above example, if the greedy strategy is adopted, the student model can select "experience" as the target element data because it has the highest probability of 0.3; in the random sampling strategy, the student model can perform random sampling according to the probability distribution, which can increase the diversity of the generated content.

[0090] 205: Concatenate the target element data with the first sample data to obtain the second sample data.

[0091] By concatenating the target element data with the first sample data, a continuous dialogue process can be simulated. In a natural language processing task, especially in a scenario involving dialogue generation, the model needs to continuously update its input based on the previous content in order to continue generating subsequent content.

[0092] For example, the process of concatenating the target element data with the first sample data to obtain the second sample data can be as follows:

[0093] Suppose the first sample data is a sequence of conversations. For example, in an interview scenario, the first sample data might be "Interviewer: Please describe your experience in the project. Job-seeking user: I have participated in a software project,". According to the previous reasoning operation, the target element data that the student model can select from the calculated first probability distribution might be "development". Through the concatenation operation, the target element data is added to the end of the first sample data to obtain the second sample data. At this time, the second sample data becomes "Interviewer: Please describe your experience in the project. Applicant: I have participated in a software project, development".

[0094] By adding the newly generated element data to the original sample data to form a new input (i.e., the second sample data), it is possible to provide updated information for the model, enabling it to continue learning and reasoning based on the updated input.

[0095] After generating the second sample data, the second sample data can be re-input into the student model and the teacher model, allowing the models to further calculate the probability distribution based on the updated information and perform new reasoning operations. This helps the model learn the coherence and logic of the conversation or text, enabling the model to gradually improve its performance during continuous iteration, better master the structure and semantics of the language, and provide a basis for transferring the knowledge of the teacher model to the student model and generating high-quality conversations or texts.

[0096] 206: Input the second sample data into the student model and the teacher model respectively, so as to use the student model to perform reasoning operations based on the second sample data to calculate the first probability distribution of different element data, and use the teacher model to perform reasoning operations based on the second sample data to calculate the second probability distribution of different element data.

[0097] By inputting the second sample data into the student model and the teacher model respectively, the student model and the teacher model can process and analyze the input second sample data according to their respective parameters and structures, thereby calculating the probability distribution of different element data.

[0098] 207: Determine the difference information between the first probability distribution and the second probability distribution corresponding to different element data respectively.

[0099] After inputting the second sample data, the student model and the teacher model may calculate the probability distribution of the next word respectively. By comparing these two probability distributions, it is possible to find the differences in the predictions of the student model in some cases from those of the teacher model, and then adjust the parameters of the student model to enable it to more accurately predict the next word. In this way, the student model can learn more accurate language patterns and knowledge from the teacher model, improving the quality and accuracy of the text it generates.

[0100] 208: Adjust the student model based on the difference information to transfer the knowledge of the second dialogue model to the first dialogue model, obtaining a target dialogue model; the target dialogue model is used to generate target question data from the question list corresponding to the target position in combination with at least one round of historical dialogue data.

[0101] After determining the difference information, the training parameters of the student model can be adjusted with the goal of making the first probability distribution output by the student model closer to the second probability distribution output by the teacher model. Thus, the student model can continuously learn the performance of the teacher model when processing the same input, learn better knowledge and experience from the teacher model, and thereby transfer the knowledge of the teacher model to the student model.

[0102] After multiple adjustments and optimizations, the finally obtained student model can be the target dialogue model. This target dialogue model combines the structure of the student model and part of the knowledge of the teacher model. When processing tasks such as interview conversations, it can utilize the knowledge advantages of the teacher model while maintaining the characteristics of the student model itself. For example, the target dialogue model may inherit the teacher model's capabilities in language patterns, semantic understanding, and logical reasoning, while also having certain advantages in resource utilization and computational efficiency like the student model, making it suitable for application in actual dialogue systems, such as intelligent interview systems, which can generate more appropriate and fluent target question data from the question list corresponding to the target position based on at least one round of historical dialogue data.

[0103] In some embodiments, the method may further include:

[0104] Obtain pre-training data, where the pre-training data includes third sample data and sample question data corresponding to the third sample data; the third sample data includes at least one round of sample dialogue data; each round of sample dialogue data includes sample question data and sample response data corresponding to the question data;

[0105] Use the pre-training data to pre-train the first dialogue model and the second dialogue model respectively.

[0106] By inputting the obtained pre-training data into the first dialogue model and the second dialogue model, the first dialogue model and the second dialogue model can learn the language patterns, semantic understanding, and logical relationships of the dialogue according to the sample question data and the sample response data corresponding to the question data in the pre-training data. For example, it will learn the common answer patterns corresponding to different questions and how to make reasonable responses according to the context. By continuously processing and learning the pre-training data, the first dialogue model and the second dialogue model can gradually adjust their own parameters to meet the requirements of this dialogue task, thereby enhancing their capabilities in dialogue-related tasks.

[0107] In the application embodiment, first, a first dialogue model and a second dialogue model with different numbers of parameters are determined. The first dialogue model is used as the student model, and the second dialogue model is used as the teacher model. Then, first sample data containing at least one round of sample dialogue data is obtained and input into the student model for inference operations. The first probability distribution of different element data is calculated and the target element data is selected. The target element data is concatenated with the first sample data to obtain second sample data. Then, the second sample data is input into the student model and the teacher model respectively, the first probability distribution and the second probability distribution of different element data are calculated respectively, and then the difference information between the first probability distribution and the second probability distribution of different element data is determined. Finally, based on the difference information, the student model is adjusted, and the knowledge of the second dialogue model is transferred to the first dialogue model to obtain the target dialogue model. By comparing the difference information between the outputs of the student model and the teacher model after the student model already has its own output and guiding the student model to adjust in a better direction according to the difference information, the training difficulty of the student model is reduced, and the performance of the trained target dialogue model is improved.

[0108] In some embodiments, the pre-training of the first dialogue model and the second dialogue model respectively using the pre-training data can be specifically implemented as follows:

[0109] For any one of the first dialogue model and the second dialogue model, the third sample data is used as the input data of the model, and the sample question data is used as the training label;

[0110] The model is trained using the input data and the training label.

[0111] After any one of the first dialogue model and the second dialogue model receives the input third sample data, the third sample data can perform forward propagation in the neural network structure of the model. Thus, the model can perform operations such as feature extraction and semantic understanding on the input data according to its own parameters and structure, and finally generate an output. This output can include a prediction of the next possible dialogue content or a preliminary answer to a related question.

[0112] After the model generates an output, the output of the model can be compared with the training label to calculate the difference between the output of the model and the training label using a loss function. The loss function can include, for example, the cross-entropy loss function, etc. The loss function can be used to measure the distance between the probability distribution of the model output and the true label. If the output of the model is quite different from the training label, the loss value will be higher, and vice versa.

[0113] Based on the loss value calculated using the loss function, the backpropagation algorithm can be used to calculate the gradient of the loss function with respect to the model parameters. The gradient represents the direction and degree of the impact of parameter changes on the loss value. Then, through optimization algorithms (such as Stochastic Gradient Descent, Adam, etc.), the model parameters are updated according to the gradient, so that when the model encounters a similar input next time, it can produce an output closer to the training label. This process is repeated continuously. Through multiple iterative trainings, the model parameters are gradually adjusted to continuously improve the performance of the model, and finally it can better handle dialogue-related tasks.

[0114] In some embodiments, obtaining the first sample data can be specifically implemented as follows:

[0115] Determine at least one round of sample dialogue data and the requirements for question output;

[0116] Based on at least one round of sample dialogue data and the requirements for question output, generate the first sample data.

[0117] Among them, the requirements for question output can be the specifications and expectations for the model output, used to clarify the conditions that the output generated by the model after processing the sample dialogue data needs to meet. In the scenario of generating interview questions, the requirements for question output can, for example, include the language style of the questions (formal, concise, etc.), the types of questions (technical, behavioral, etc.), the relevance of the questions to the interview position, the difficulty level of the questions, etc. In addition, the requirements for question output can also include the requirements for the model's capabilities. For example, it can be required that the model selects a question from a preset list of questions and outputs it in sequence.

[0118] Among them, the first sample data can be implemented as a prompt. A prompt is a form of input used to prompt or guide the model to give an output that meets expectations. A prompt is a natural language input, similar to a command or instruction, to let the model know what it needs to do.

[0119] In the embodiments of the present application, in order to facilitate the generation of prompt information, a prompt template can be pre-configured. By adding at least one round of sample dialogue data and the requirements for question output to the prompt template, the corresponding prompt can be generated.

[0120] Among them, the prompt template can include the following parts:

[0121] At least one round of sample dialogue data: Interviewer: "Please share the work content you were responsible for in your recent project."

[0122] Job-seeking user: "I mainly XXXX."

[0123] Requirements for question output: Generate a formal, professional question targeting the technical capabilities of this job-seeking user.

[0124] In some embodiments, the second sample data is input into the student model and the teacher model respectively, so as to use the student model to perform an inference operation based on the second sample data to calculate the first probability distribution of different element data, which can be specifically implemented as follows:

[0125] The second sample data is input into the student model and the teacher model respectively, so as to use the student model to perform an inference operation based on the second sample data, and select target element data according to the first probability distribution;

[0126] In some embodiments, the method may further include:

[0127] The target element data is concatenated with the second sample data to update the second sample data, and the step of inputting the second sample data into the student model and the teacher model respectively is returned to continue execution.

[0128] In the embodiments of the present application, after the concatenation is completed, the updated second sample data is input into the student model and the teacher model again, and the above steps are repeated, realizing an iterative training process. By continuously updating the sample data and letting the model process it, the model can gradually learn how to generate more reasonable and coherent text or dialogue, and can continuously optimize its calculation of the probability distribution of element data and the selection of target element data, which helps to improve the performance of the model and make it perform better in tasks such as dialogue generation and text continuation.

[0129] In practical applications, this iterative training process can be continued multiple times until a certain stopping condition is met. The stopping condition may include, for example, reaching a predetermined number of iterations, the generated text length reaching the requirement, or the generated content meeting a certain evaluation index (such as the generated dialogue conforming to human language habits and logic). Through this iterative method, the student model can continuously learn from the input data, and gradually adjust its own parameters and prediction ability, and finally achieve learning and absorbing knowledge from the teacher model to achieve better performance. At the same time, the teacher model can also provide a reference for the student model to help the student model adjust its prediction of element data in each iteration, so that the student model gradually approaches the performance of the teacher model.

[0130] Figure 3 The flowchart of an artificial intelligence interview method provided in an embodiment of the present application is shown, which is applied to a server, such as Figure 3 As shown, the method may specifically include the following steps:

[0131] 301: Obtain an interview request sent by the client; the interview request is generated by the client in response to an interview operation triggered by a target user for a target position.

[0132] In an embodiment of the present application, the server can detect an interview operation triggered by the client. When the target user triggers an interview operation for a certain target position on the display interface provided by the client, the client can generate an interview request. The interview operation can be, for example, that the target user clicks the "Start Interview" button, or selects to apply for an interview on a specific recruitment page, etc.

[0133] Then, the client can send this interview request to the server. After the server obtains this request, it can start the interview process. For example, on an online recruitment platform, after a user browses a desired waiter position and clicks the "Apply for Interview" button, the client will generate a corresponding interview request and send it to the server.

[0134] 302: Determine the question list corresponding to the target position; the question list includes multiple question data.

[0135] After the server obtains the interview request, it can obtain a series of questions related to the target position. The server can find the question list corresponding to the target position according to the internal database or configuration information. This question list contains multiple question data, and these question data can be prepared in advance according to the characteristics, requirements, and common examination points of the target position. For example, for a software engineer position, the question list may include questions in dimensions such as programming language knowledge, project experience, and algorithm understanding.

[0136] 303: Determine the target dialogue model; the target dialogue model is obtained by migrating the knowledge of the second dialogue model to the first dialogue model; the first dialogue model is used as the student model, and the second dialogue model is used as the teacher model; the student model is used to adjust based on the difference information between the first probability distribution and the second probability distribution of different element data generated by inference; the first probability distribution is generated by the student model based on the second sample data; the second probability distribution is generated by the teacher model based on the second sample data; the second sample input model is obtained by splicing the first sample data with the target element data; the target element data is generated by the student model based on the first sample data.

[0137] The target dialogue model can be trained and generated by the Figure 2 shown dialogue model training method. The training method of the target dialogue model can refer to the relevant embodiments of the Figure 2 shown dialogue model training method and will not be elaborated here.

[0138] 304: Use the target dialogue model to combine at least one round of historical dialogue data to generate target question data from the question list.

[0139] The target dialogue model can generate target question data from a question list based on at least one round of historical dialogue data. This historical dialogue data records the content that has been exchanged during the previous interview process. The target dialogue can screen and generate the target question data most suitable for the current interview stage from the question list by analyzing this historical dialogue data and the question list.

[0140] Among them, the question list contains multiple pre-set question data, and these questions can be arranged in a certain logical order. The target dialogue model can be trained to have the ability to generate target question data from the question list following the predefined question order.

[0141] The target dialogue model can, according to the input historical dialogue data, combine its own training results and decide which question should be output currently as the target question data from the question list in order.

[0142] 305: Send the target question data to the client.

[0143] 306: Obtain the response data feedback by the client, and return to continue the execution of the step of generating target question data from the question list by using the target dialogue model combined with at least one round of historical dialogue data.

[0144] After generating the target question data, the server may send the target question data to the client. After receiving the target question data, the client can display it to the target user for the target user to answer. In this way, the target user can see the interview questions and prepare the corresponding responses.

[0145] The client can collect the response data given by the target user for the target question and feedback it to the server. After obtaining this response data, the server will return to step 304 again, that is, use the target dialogue model combined with the new at least one round of historical dialogue data (the historical dialogue data at this time includes the just user's response) to generate new target question data from the question list. This process will loop continuously until the interview ends, and in this way, a dynamic interview process based on the user's answers and continuously advancing is realized.

[0146] In some embodiments, selecting the target question data from the question list by using the target dialogue model combined with at least one round of historical dialogue data can be specifically implemented as:

[0147] Combined with at least one round of historical dialogue data, determine whether the target user meets the question generation condition;

[0148] If the target user meets the question generation condition, use the target dialogue model, combined with at least one round of historical dialogue data, to generate target question data from the question list.

[0149] Among them, the historical dialogue data of at least one round contains the previous interview interaction information, and the previous interview interaction information can be used to help the target dialogue model determine whether new question data needs to be generated. If new question data needs to be generated, the target question data can be generated from the question list by using the target dialogue model in combination with the historical dialogue data of at least one round.

[0150] In some embodiments, if the target user does not meet the question generation condition, the method may further include:

[0151] Determining the target question data from the historical question data in the historical dialogue data.

[0152] If the target user does not meet the question generation condition, that is, new question data does not need to be generated at present, the target question data can be determined from the historical question data that has been asked.

[0153] In some embodiments, determining whether the target user meets the question generation condition by combining the historical dialogue data of at least one round can be specifically implemented as:

[0154] Determining the previous round of question data and the previous round of response data in the previous round of dialogue data;

[0155] Judging whether the previous round of response data matches the previous round of question data; if so, determining that the target user meets the question generation condition, otherwise not meeting the question generation condition;

[0156] In a possible implementation, a first recognition model can be used to determine whether the previous round of response data matches the previous round of question data. The first recognition model can be a trained large model. A large model refers to a machine learning model with a large number of parameters and a complex structure, which can process massive amounts of data and complete various complex tasks, such as natural language processing, computer vision, speech recognition, etc. It is an AI (Artificial Intelligence) model. Among them, the large model can be implemented using, for example, a large language model (LLM) or a multimodal large model (MLM), etc. For example, it can be implemented using GPT-3 (Generative Pre-Trained Transformer-3, the third generation of generative pre-trained model), GPT-4 (Generative Pre-Trained Transformer-4, the fourth generation of generative pre-trained model), BERT (Bidirectional Encoder Representation from Transformers, a bidirectional encoder model based on Transformers), Turing NLG (Turing Natural language Generation), and so on. This application does not limit this. Such large models can perform well in various natural language processing tasks.

[0157] The first recognition model can be trained based on the sample dialogue data and the matching labels between the sample question data and the sample response data in the sample dialogue data. For example: for the sample dialogue data, it may contain multiple sets of sample question data and sample response data, as well as the corresponding matching labels (indicating whether the response matches the question), such as "Question: Please list the programming languages you master. Response: I master Java and Python. Matching label: Yes". By training on a large amount of such sample data, the first recognition model learns how to determine whether a response reasonably answers the corresponding question.

[0158] When the previous round of question data and response data are received, they can be input into the first recognition model. The first recognition model can, based on the knowledge it has learned through training, evaluate whether the previous round of response data is a reasonable answer to the previous round of question data.

[0159] In the case where the first recognition model determines that the previous round of response data matches the previous round of question data, it means that more questions need to be asked to further understand the user's situation, or enter the next interview session, which can indicate that the target user meets the question generation conditions.

[0160] In the case where the first recognition model determines that the previous round of response data does not match the previous round of question data, it may mean that the user's answer is not accurate enough or the question is not understood, and the current question needs to be asked again. At this time, it can be determined that the target user does not meet the question generation conditions.

[0161] In some embodiments, the method may further include:

[0162] If the target user does not meet the question generation conditions, use the previous round of question data as the target question data.

[0163] In some embodiments, if the target user does not meet the question generation conditions, using the previous round of question data as the target question data can be specifically implemented as:

[0164] If the target user does not meet the question generation conditions, determine whether the retry count of the previous round of question data meets the limit conditions;

[0165] If not, use the previous round of question data as the target question data and accumulate the retry count;

[0166] If so, use the target dialogue model, in combination with at least one round of historical dialogue data, to generate the target question data from the question list.

[0167] Among them, the retry count may refer to the number of times of repeatedly asking the same question data to the target user, and the limit condition may be a preset number threshold for controlling the upper limit of the number of times of asking the same question data. For example, in an interview scenario, the limit condition can be set to a maximum of 3 retries for each question. If it is the first or second time to ask a certain question to the target user, then it is necessary to determine whether the current retry count has reached 3 times at this time. If it has not reached 3 times, the limit condition is not met and the question can be asked again; if it has reached 3 times, the limit condition is met.

[0168] If it is determined that the retry count of the previous round of question data does not meet the limit conditions, then use the previous round of question data as the target question data again and send it to the target user. In the case where the previous round of response data does not match the previous round of question data, it may be that the target user does not understand the question or answers incorrectly for a moment. By asking again, a more appropriate answer may be obtained.

[0169] At the same time, after sending the previous round of question data as the target question data, it is necessary to accumulate the retry count of this question.

[0170] When it is determined that the retry count of the previous round of question data has met the limit condition, it indicates that after multiple inquiries about this question, the target user still has not given a suitable answer. At this time, in order not to get stuck in this question data and prevent the interview or conversation from proceeding, this question data can be temporarily skipped to ensure the smooth progress of the interview or conversation.

[0171] In some embodiments, if the target user does not meet the question generation condition, taking the previous round of question data as the target question data can be specifically implemented as follows:

[0172] If the target user does not meet the question generation condition, determine whether the previous round of response data is a question data;

[0173] If not, take the previous round of question data as the target question data;

[0174] If so, use the conversation model to generate response data corresponding to the question data and send the response data to the client.

[0175] In a possible implementation manner, the second recognition model can be used to determine whether the previous round of response data is a question data. The second recognition model can be a trained large model.

[0176] If the second recognition model determines that the previous round of response data is not a question data, it means that the user is answering a question, but their answer may not meet the requirements, resulting in the target user not meeting the question generation condition. In this case, the previous round of question data can be taken as the target question data.

[0177] When the second recognition model determines that the previous round of response data is a question data, it indicates that the target user has asked a question. At this time, the conversation model needs to be used to generate response data for this question data according to the position information of the target position.

[0178] In a possible implementation manner, the conversation model can be implemented as a target dialogue model.

[0179] In some embodiments, combining at least one round of historical conversation data to determine whether the target user meets the question generation condition can be specifically implemented as follows:

[0180] Determine the previous round of question data and the previous round of response data in the previous round of conversation data;

[0181] Determine whether the previous round of response data is a question data; if so, determine that the target user meets the question generation condition.

[0182] In some embodiments, the method may further include:

[0183] If the target user does not meet the question generation condition, take the previous round of question data as the target question data.

[0184] In some embodiments, if the historical dialogue data meets the question generation condition, using the target dialogue model, combining at least one round of historical dialogue data to generate target question data from the question list can be specifically implemented as follows:

[0185] If the target user meets the question generation condition, use the conversation model to generate a reply data corresponding to the question data, and send the reply data to the client;

[0186] Use the target dialogue model, combine the historical dialogue data to generate target question data from the question list.

[0187] In some embodiments, sending the target question data to the client can be specifically implemented as follows:

[0188] Judge whether the target question data meets the constraint condition;

[0189] If so, send the target question data to the client;

[0190] If not, return to execute using the target dialogue model to combine at least one round of historical dialogue data to generate target question data from the question list.

[0191] In some embodiments, judging whether the target question data meets the constraint condition can be specifically implemented as follows:

[0192] Determine whether at least one round of historical dialogue data contains the precondition information of the target question data;

[0193] If so, it is determined that the constraint condition is met; if not, it is determined that the constraint condition is not met.

[0194] Among them, in the question list, at least one question data can be configured with a constraint condition. A question data configured with a constraint condition means that the raising of this question data needs to be based on a certain precondition information. In the case that the precondition information appears in at least one round of historical dialogue data, it can be judged that the target question data meets the constraint condition; otherwise, it does not meet the constraint condition. For example, in an interview scenario, for the question data "Please elaborate in detail on the specific problems and solutions you encountered when using Python for data analysis", then its precondition information can be that in the previous historical dialogue, the target user mentioned that he has experience in using Python.

[0195] After selecting the target question data configured with constraints, the target dialogue model can check at least one round of historical dialogue data to see if there is any prerequisite information related to the target question data. If it exists, it indicates that the target question data meets the constraints because it is a natural extension based on the previous dialogue content and conforms to the logical coherence of the dialogue. For example, if the applicant mentioned in the previous dialogue that "I often use Python to process data at work", then the above question about Python data analysis has a reasonable prerequisite and meets the constraints.

[0196] When the target question data meets the constraints, the server can send it to the client. This can ensure that the questions received by the client are relevant and reasonable to the previous dialogue, which helps the smooth progress of the dialogue. If the target question data does not meet the constraints, the operation of generating the target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data may be returned for execution.

[0197] In a possible implementation, a third recognition model can be used to determine whether the target question data meets the constraints. The third recognition model can be a trained large model.

[0198] In actual application scenarios, different groups of people, such as those with differences in age, occupation, educational background, etc., and those with different experiences, will have different questions to be asked in the dialogue scenario. For example, in a recruitment interview, for fresh graduates and job seekers with many years of work experience, the interviewer's focus of attention is different, and the questions asked are naturally different. Under different recruitment positions, the questions to be asked are also different.

[0199] By defining prerequisite constraints for question data, it is like setting a "switch" for the questions. Only when specific conditions are met will the corresponding questions be asked. When new categories need to be expanded, such as adding new job categories in recruitment, only need to define all the questions to be asked for that category and clarify under what conditions different questions are asked. The target dialogue model can then automatically propose appropriate questions according to the actual situation in the new category dialogue scenario, thus realizing the dialogue ability to automatically support new categories without large-scale modification of the overall dialogue system architecture or logic, reducing the development and maintenance costs, and improving the flexibility and scalability of the interview.

[0200] In some embodiments, generating the target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data can be specifically implemented as:

[0201] Using the target dialogue model in combination with at least one round of historical dialogue data, generate candidate question data from the unlabeled question data in the question list, and determine whether there is corresponding response data for the candidate question data in at least one round of historical dialogue data. If so, label the candidate question data and regenerate the candidate question data. If not, use the candidate question data as the target question data and label the target question data.

[0202] Among them, the unlabeled question data means that these question data have not been used in the previous dialogue process. The target dialogue model can generate candidate question data from these unlabeled question data according to the information provided by the historical dialogue data.

[0203] If there is already a relevant answer to a certain question data, then it may not be necessary to ask this question data again. Therefore, when generating candidate question data, search in at least one round of historical dialogue data to see if there is already corresponding response data for this candidate question.

[0204] If the corresponding response data for the candidate question data is found in the historical dialogue data, then label the candidate question data. By labeling this question data, it can be indicated that this question already has relevant information in the current dialogue and cannot be used as the current valid question. Then, candidate question data may be regenerated, and again look for a suitable question from the unlabeled questions in the question list, repeating the above judgment process.

[0205] If the corresponding response data for the candidate question data is not found in the historical dialogue data, this means that this question has not been involved in the current dialogue and is a new question that can be asked. At this time, the candidate question data can be used as the target question data and the target question data is labeled. Labeling the target question data is to record that this question has been used and will not be misselected again in subsequent dialogues.

[0206] In some embodiments, using the target dialogue model in combination with at least one round of historical dialogue data to generate target question data from the question list can be specifically implemented as:

[0207] Using the target dialogue model in combination with at least one round of historical dialogue data, generate target question data from the question list according to at least one question output requirement;

[0208] The at least one question output requirement includes one or more of the following options:

[0209] The response data in the previous round of historical dialogue data matches the question data;

[0210] The retry count of the question data in the previous round of historical dialogue data exceeds the limit condition;

[0211] There is no response data in at least one round of historical conversation data that matches the target question data;

[0212] And,

[0213] The target question data meets its corresponding constraint conditions.

[0214] As mentioned in the above embodiments, in the embodiments of the present application, the first recognition model, the second recognition model, and the third recognition model can be used to identify whether the response data matches the question data, whether the number of retry attempts exceeds the limit conditions, and whether the constraint conditions are met. Thus, these complex tasks that the target dialogue model needs to handle can be split out, thereby reducing the processing pressure on the target dialogue model and lowering the training cost of the target dialogue model. Moreover, for each recognition model, more targeted data can be used for training, enabling different recognition models to focus more on their respective tasks. Through specialized training and optimization, their performance in specific tasks can be improved.

[0215] In some embodiments, after obtaining the response data fed back by the client, the method may further include:

[0216] Determine whether there is unselected question data in the question list;

[0217] If so, continue to execute the step of selecting question data from the question list; if not, generate an interview end prompt message.

[0218] If it is found after checking the question list that there is no unselected question data in the question list, that is, all questions have been asked, this indicates that the interview has covered all the preset aspects of the investigation and the interview can end. At this time, the server can generate an interview end prompt message. This prompt message will be sent to the client to inform the applicant that the interview has ended. For example, the client may display "This interview has ended. Thank you for your participation!".

[0219] In some embodiments, the method may further include:

[0220] Receive the resume submission operation of the client for the target position;

[0221] In response to the resume submission operation, send an interview invitation notice to the client;

[0222] Obtaining the interview request sent by the client includes:

[0223] Obtain the interview request sent by the client by triggering the interview invitation notice.

[0224] Figure 4 Shows a schematic diagram of the artificial intelligence interview method provided by the embodiments of the present application.

[0225] In Figure 4 Figure 4 , 401 may represent the server, and 402 may represent the client. The target dialogue model 4011 may be deployed in the server 401, and the first recognition model 4012, the second recognition model 4013, and the third recognition model 4014 may be deployed in the server 401.

[0226] After the server 401 receives the response data for the target question data sent by the client 402, it may input at least one round of historical dialogue data into the first recognition model 4012, the second recognition model 4013, and the third recognition model 4014 respectively, so that the first recognition model 4012 can identify whether the response data of the target user matches the target question data, and whether there is a question in the response data of the target user; the second recognition model 4013 may be used to identify the target question data of this round of Q&A, and the retry times of the target question data; the third recognition model 4014 may be used to identify whether the target question data meets the constraint conditions.

[0227] The first recognition model 4012, the second recognition model 4013, and the third recognition model 4014 may respectively output recognition results, and then the recognition results may be input into the target dialogue model, so that the target dialogue model can generate the question data to be asked in the next round of dialogue according to the recognition results. The question data to be asked in the next round of dialogue may be to repeat asking the target question data, or to reselect question data from the question list.

[0228] In the embodiments of the present application, in order to improve the interaction experience, after the server determines the target question data, it may convert the target question data into voice data, so that the digital human can broadcast the target question data using the generated voice data. Broadcasting by the digital human can create an atmosphere closer to real face-to-face communication for the target user.

[0229] Figure 5 shows a block diagram of a training device for a dialogue model provided by an embodiment of the present application, as Figure 5 shown, the device may include:

[0230] The first determination module 501 is used to determine the first dialogue model and the second dialogue model, wherein the number of parameters of the second dialogue model is greater than that of the first dialogue model;

[0231] The second determination module 502 is used to use the first dialogue model as the student model and the second dialogue model as the teacher model;

[0232] The sample acquisition module 503 is used to acquire first sample data; the first sample data includes at least one round of sample dialogue data;

[0233] The first input module 504 is configured to input the first sample data into the student model, so as to use the student model to perform an inference operation based on the first sample data, calculate a first probability distribution of different element data, and select target element data according to the first probability distribution; the element data is the smallest data unit.

[0234] The splicing module 505 is configured to splice the target element data and the first sample data to obtain second sample data.

[0235] The second input module 506 is configured to input the second sample data into the student model and the teacher model respectively, so as to use the student model to perform an inference operation based on the second sample data to calculate a first probability distribution of different element data, and use the teacher model to perform an inference operation based on the second sample data to calculate a second probability distribution of different element data.

[0236] The difference determination module 507 is configured to determine difference information between the first probability distribution and the second probability distribution corresponding to different element data.

[0237] The training module 508 is configured to adjust the student model based on the difference information, so as to transfer the knowledge of the second dialogue model to the first dialogue model to obtain a target dialogue model; the target dialogue model is used to generate target question data from a question list corresponding to a target position in combination with at least one round of historical dialogue data.

[0238] In some embodiments, the apparatus may further include:

[0239] The pre-training data acquisition module is configured to acquire pre-training data, where the pre-training data includes third sample data and sample question data corresponding to the third sample data; the third sample data includes at least one round of sample dialogue data; each round of sample dialogue data includes sample question data and sample response data corresponding to the question data.

[0240] The pre-training module is configured to perform pre-training on the first dialogue model and the second dialogue model respectively using the pre-training data.

[0241] In some embodiments, the pre-training module is specifically configured to:

[0242] For any one of the first dialogue model and the second dialogue model, use the third sample data as the input data of the model, and use the sample question data as the training label.

[0243] Use the input data and the training label to train the model.

[0244] In some embodiments, the sample acquisition module 503 is specifically configured to:

[0245] Determine at least one round of sample conversation data and question output requirements;

[0246] Generate the first sample data based on the at least one round of sample conversation data and the question output requirements.

[0247] In some embodiments, the second input module 506 is specifically configured to:

[0248] Input the second sample data into the student model and the teacher model respectively, so that the student model performs an inference operation based on the second sample data and selects target element data according to the first probability distribution.

[0249] In some embodiments, the device further includes:

[0250] A data splicing module, configured to splice the target element data with the second sample data to update the second sample data, and return to continue executing the step of inputting the second sample data into the student model and the teacher model respectively.

[0251] Figure 5 The training device of the dialogue model can execute Figure 2 The method of the dialogue model described in the illustrated embodiments, and its implementation principle and technical effects will not be elaborated. For the device of the dialogue model in the above embodiments, the specific manners in which each module and unit perform operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0252] Figure 6 The block diagram of an artificial intelligence interview device provided in an embodiment of the present application is shown. This device can be applied to a server side, such as Figure 6 As shown, the device may include:

[0253] A first request acquisition module 601, configured to acquire an interview request sent by a client; the interview request is generated by the client in response to an interview operation triggered by a target user for a target position;

[0254] A first question list determination module 602, configured to determine a question list corresponding to the target position; the question list includes a plurality of question data;

[0255] A dialogue model determination module 603, configured to determine a target dialogue model; the target dialogue model is obtained by migrating the knowledge of a second dialogue model to a first dialogue model; the first dialogue model is used as a student model, and the second dialogue model is used as a teacher model; the student model is used to adjust based on the difference information between the first probability distribution and the second probability distribution of different element data generated by reasoning; the first probability distribution is generated by the student model based on second sample data; the second probability distribution is generated by the teacher model based on second sample data; the second sample input model is obtained by concatenating first sample data and target element data; the target element data is generated by the student model based on the first sample data.

[0256] A first question generation module 604, configured to generate target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data.

[0257] A first question sending module 605, configured to send the target question data to the client.

[0258] A first response acquisition module 606, configured to acquire the response data fed back by the client, and return to continue executing the step of generating target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data.

[0259] In some embodiments, the first question generation module 604 is specifically configured to:

[0260] In combination with at least one round of historical dialogue data, determine whether the target user meets the question generation condition;

[0261] If the target user meets the question generation condition, use the target dialogue model to generate target question data from the question list in combination with at least one round of historical dialogue data.

[0262] In some embodiments, if the target user does not meet the question generation condition, the apparatus may further include:

[0263] A first question determination module, configured to determine target question data from the historical question data in the historical dialogue data.

[0264] In some embodiments, the first question generation module 604 is specifically configured to:

[0265] Determine the previous round of question data and the previous round of response data in the previous round of dialogue data;

[0266] Determine whether the previous round of response data matches the previous round of question data; if so, determine that the target user meets the question generation condition, otherwise does not meet the question generation condition.

[0267] In some embodiments, the device may further include:

[0268] A second question determination module, configured to use the previous round of question data as the target question data when the target user does not meet the question generation condition.

[0269] In some embodiments, the second question determination module is specifically configured to:

[0270] If the target user does not meet the question generation condition, determine whether the retry count of the previous round of question data meets the limit condition;

[0271] If not, use the previous round of question data as the target question data and accumulate the retry count;

[0272] If so, use the target dialogue model, in combination with at least one round of historical dialogue data, to generate target question data from the question list.

[0273] In some embodiments, the second question determination module is specifically configured to:

[0274] If the target user does not meet the question generation condition, determine whether the previous round of response data is question data;

[0275] If not, use the previous round of question data as the target question data;

[0276] If so, use the session model to generate a response data corresponding to the question data and send the response data to the client.

[0277] In some embodiments, the second question determination module is specifically configured to:

[0278] Determine the previous round of question data and the previous round of response data in the previous round of dialogue data;

[0279] Determine whether the previous round of response data is question data; if so, determine that the target user meets the question generation condition.

[0280] In some embodiments, the device further includes:

[0281] A third question determination module, configured to use the previous round of question data as the target question data if the target user does not meet the question generation condition;

[0282] In some embodiments, the first question generation module 604 is specifically configured to:

[0283] If the target user meets the question generation condition, use the session model to generate a response data corresponding to the question data and send the response data to the client;

[0284] Generate target question data from the question list by using the target dialogue model in combination with the historical dialogue data.

[0285] In some embodiments, the first question sending module 605 is specifically configured to:

[0286] Determine whether the target question data meets the constraint conditions;

[0287] If so, send the target question data to the client;

[0288] If not, return to execute generating target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data.

[0289] In some embodiments, the first question sending module 605 is specifically configured to:

[0290] Determine whether the pre-information of the target question data is included in the at least one round of historical dialogue data;

[0291] If so, it is determined that the constraint conditions are met, and if not, it is determined that the constraint conditions are not met.

[0292] In some embodiments, the first question generating module 604 is specifically configured to:

[0293] Generate candidate question data from the unmarked question data in the question list by using the target dialogue model in combination with at least one round of historical dialogue data, and determine whether there is corresponding response data for the candidate question data in the at least one round of historical dialogue data. If so, mark the candidate question data and regenerate the candidate question data. If not, use the candidate question data as the target question data and mark the target question data.

[0294] In some embodiments, the first question generating module 604 is specifically configured to:

[0295] Generate target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data according to at least one question output requirement;

[0296] The at least one question output requirement includes one or more of the following options:

[0297] The response data in the previous round of historical dialogue data matches the question data;

[0298] The retry times of the question data in the previous round of historical dialogue data exceed the limit conditions;

[0299] There is no response data in the at least one round of historical dialogue data that matches the target question data;

[0300] And,

[0301] The target problem data satisfies its corresponding constraint conditions.

[0302] In some embodiments, the device may further include:

[0303] A list determination module, configured to determine whether there is unselected problem data in the problem list;

[0304] A first selection module, configured to, when there is unselected problem data in the problem list, continue to execute the step of selecting problem data from the problem list; otherwise, generate an interview end prompt message.

[0305] In some embodiments, the device may further include:

[0306] A first operation receiving module, configured to receive a resume submission operation of the client for the target position;

[0307] A first notification module, configured to send an interview invitation notification to the client in response to the resume submission operation;

[0308] In some embodiments, the first request acquisition module 601 is specifically configured to:

[0309] Obtain the interview request sent by the client by triggering the interview invitation notification.

[0310] Figure 6 The described artificial intelligence interview device can execute Figure 3 The artificial intelligence interview method described in the illustrated embodiments, and its implementation principles and technical effects will not be elaborated further. For the artificial intelligence interview device in the above embodiments, the specific manners in which each module and unit perform operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0311] The embodiments of the present application further provide a computing device, as Figure 7 shown, the device may include a storage component and a processing component;

[0312] The storage component stores one or more computer instructions, where the one or more computer instructions are called and executed by the processing component to implement the training method of the dialogue model and the artificial intelligence interview method provided by the embodiments of the present application.

[0313] Of course, the computing device may necessarily further include other components, such as an input / output interface, a display component, a communication component, etc.

[0314] The input / output interface provides an interface between the processing component and the peripheral interface module, and the above-mentioned peripheral interface module can be an output device, an input device, etc. The communication component is configured to facilitate communication between the computing device and other devices in a wired or wireless manner, etc.

[0315] Among them, the processing component may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component can also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components for executing the above method.

[0316] The storage component is configured to store various types of data to support the operation of the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0317] The display component can be an electroluminescent (EL) element, a liquid crystal display or a microdisplay with a similar structure, or a retina direct display or a similar laser scanning display.

[0318] It should be noted that the above computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. It can be implemented as a distributed cluster composed of multiple servers or terminal devices, or can be implemented as a single server or a single terminal device.

[0319] The embodiment of the present application also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a computer, it can implement the training method of the dialogue model and the artificial intelligence interview method provided in the above embodiment. The computer-readable medium can be included in the electronic device described in the above embodiment; or it can exist alone without being assembled into the electronic device.

[0320] The embodiment of the present application also provides a computer program product, which includes a computer program carried on a computer-readable storage medium, and when the computer program is executed by a computer, it can implement the training method of the dialogue model and the artificial intelligence interview method provided in the above embodiment. In such an embodiment, the computer program can be downloaded and installed from the network, and / or installed from a removable medium. When the computer program is executed by a processor, it executes various functions defined in the system of the present application.

[0321] It should be noted that in the embodiments of the present application, the use of user data may be involved. In practical applications, user-specific personal data can be used in the solutions described herein within the scope permitted by applicable laws and regulations of the country where it is located (for example, with the user's explicit consent, giving the user a practical notice, etc.).

[0322] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein again.

[0323] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0324] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the essence of the above technical solutions, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0325] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not intended to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments or perform equivalent replacements for some of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of each embodiment of the present application.

Claims

1. A method for training a dialogue model, characterized in that: include: Determining a first dialogue model and a second dialogue model, wherein a parameter amount of the second dialogue model is greater than a parameter amount of the first dialogue model; using the first dialogue model as a student model and the second dialogue model as a teacher model; Acquire first sample data; the first sample data includes at least one round of sample conversation data; Inputting the first sample data into the student model, using the student model to perform an inference operation based on the first sample data to calculate a first probability distribution of different element data, and selecting target element data according to the first probability distribution; the element data is the smallest data unit; splicing the target element data with the first sample data to obtain second sample data; Inputting the second sample data into the student model and the teacher model respectively, so as to use the student model to perform an inference operation based on the second sample data to calculate a first probability distribution of different element data, and use the teacher model to perform an inference operation based on the second sample data to calculate a second probability distribution of different element data; Determine difference information of the first probability distribution and the second probability distribution corresponding to different element data respectively; The student model is adjusted based on the difference information to transfer the knowledge of the second dialogue model to the first dialogue model to obtain a target dialogue model; the target dialogue model is used to generate target question data from a question list corresponding to a target position in combination with at least one round of historical dialogue data.

2. The method according to claim 1, characterized in that: Also includes: Acquire pre-training data, where the pre-training data includes third sample data and sample question data corresponding to the third sample data; The third sample data includes at least one round of sample conversation data; Each round of sample conversation data includes sample question data and sample answer data corresponding to the question data; The first dialogue model and the second dialogue model are pre-trained respectively using the pre-training data.

3. The method according to claim 2, characterized in that The pre-training the first dialogue model and the second dialogue model respectively using the pre-training data comprises: For any one of the first dialogue model and the second dialogue model, using the third sample data as input data of the model, and using the sample question data as a training label; The model is trained using the input data and the training labels.

4. The method according to claim 1, characterized in that: The obtaining of the first sample data comprises: Determine at least one round of sample conversation data and question output requirements; The first sample data is generated based on the at least one round of sample dialogue data and the question output requirement.

5. The method according to claim 1, characterized in that The step of inputting the second sample data into the student model and the teacher model respectively, using the student model to perform an inference operation based on the second sample data, so as to calculate a first probability distribution of different element data, and selecting target element data according to the first probability distribution includes: Inputting the second sample data into the student model and the teacher model respectively, so as to use the student model to perform an inference operation based on the second sample data and select target element data according to the first probability distribution; The method further comprises: The target element data is concatenated with the second sample data to update the second sample data, and the step of inputting the second sample data into the student model and the teacher model is returned to continue the execution.

6. An artificial intelligence interview method, characterized in that: Applied to the server, the method includes: Obtaining an interview request sent by a client; the interview request is generated by the client in response to an interview operation triggered by a target user for a target position; Determine a question list corresponding to the target position; the question list includes a plurality of question data; Determine a target dialogue model; the target dialogue model is obtained by transferring the knowledge of the second dialogue model to the first dialogue model; the first dialogue model is used as a student model, and the second dialogue model is used as a teacher model; the student model is used to adjust the difference information of the first probability distribution of different element data and the second probability distribution generated based on reasoning; the first probability distribution is generated by the student model based on the second sample data; the second probability distribution is generated by the teacher model based on the second sample data; the second sample input model is obtained by splicing the first sample data with the target element data; the target element data is generated by the student model based on the first sample data; Generate target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data; Sending the target question data to the client; The response data fed back by the client is obtained, and the step of using the target dialogue model type in combination with at least one round of historical dialogue data to generate target question data from the question list is returned to continue execution.

7. The method according to claim 6, characterized in that The step of using the target dialogue model in combination with at least one round of historical dialogue data to select target question data from a question list includes: Based on at least one round of historical conversation data, determine whether the target user meets the question generation condition; If the target user meets the question generation condition, the target dialogue model is used to generate target question data from the question list in combination with at least one round of historical dialogue data.

8. The method according to claim 7, characterized in that If the target user does not meet the question generation condition, the method further includes: Target question data is determined from historical question data in the historical dialogue data.

9. The method according to claim 8, characterized in that The step of combining at least one round of historical conversation data to determine whether the target user meets the question generation condition includes: Determine the previous round of question data and the previous round of answer data in the previous round of dialogue data; Determine whether the previous round of answer data matches the previous round of question data; if so, determine that the target user meets the question generation condition, otherwise, the question generation condition is not met; The method further comprises: If the target user does not meet the question generation condition, the previous round of question data is used as the target question data.

10. The method according to claim 9, characterized in that If the target user does not meet the question generation condition, taking the previous round of question data as the target question data includes: If the target user does not meet the problem generation conditions, determine whether the number of retries of the previous round of problem data meets the restriction conditions; If not, the previous round of question data is used as the target question data, and the number of retries is accumulated; If so, the target dialogue model is used in combination with at least one round of historical dialogue data to generate target question data from the question list.

11. The method according to claim 9, characterized in that If the target user does not meet the question generation condition, taking the previous round of question data as the target question data includes: If the target user does not meet the question generation conditions, determine whether the previous round of answer data is question data; If not, the previous round of question data is used as the target question data; If so, the conversation model is used to generate reply data corresponding to the question data, and the reply data is sent to the client.

12. The method according to claim 7, characterized in that The step of combining at least one round of historical conversation data to determine whether the target user meets the question generation condition includes: Determine the previous round of question data and the previous round of answer data in the previous round of dialogue data; Determine whether the previous round of answer data is question data; if so, determine that the target user meets the question generation condition; The method further comprises: If the target user does not meet the question generation conditions, the previous round of question data is used as the target question data; If the historical dialogue data satisfies the question generation condition, generating target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data includes: If the target user meets the question generation condition, generate reply data corresponding to the question data using the conversation model, and send the reply data to the client; The target dialogue model is used to generate target question data from the question list in combination with the historical dialogue data.

13. The method according to claim 6, characterized in that The sending the target question data to the client comprises: Determine whether the target problem data satisfies the constraint conditions; If so, sending the target question data to the client; If not, return to execute the process of generating target question data from the question list by utilizing the target dialogue model in combination with at least one round of historical dialogue data.

14. The method according to claim 13, characterized in that The determining whether the target problem data satisfies the constraint condition comprises: Determining whether the at least one round of historical dialogue data contains the pre-information of the target question data; If so, it is determined that the constraint condition is satisfied; if not, it is determined that the constraint condition is not satisfied.

15. The method according to claim 6, characterized in that Using the target dialogue model in combination with at least one round of historical dialogue data to generate target question data from the question list includes: The target dialogue model is combined with at least one round of historical dialogue data to generate candidate question data from the unmarked question data in the question list, and it is determined from the at least one round of historical dialogue data whether there is response data corresponding to the candidate question data. If so, the candidate question data is marked and the candidate question data is regenerated. If not, the candidate question data is used as the target question data and the target question data is marked.

16. The method according to claim 6, characterized in that The step of generating target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data includes: Generate target question data from the question list by using the target dialogue model in combination with at least one round of historical dialogue data according to at least one question output requirement; The at least one question output requirement includes one or more of the following options: Match the answer data and question data in the previous round of historical dialogue data; The number of retries for the question data in the previous round of historical conversation data exceeds the limit; There is no answer data matching the target question data in at least one round of historical dialogue data; as well as, The target problem data satisfies its corresponding constraints.

17. The method according to claim 6, characterized in that After obtaining the response data fed back by the client, the method further includes: Determine whether there is unselected question data in the question list; If yes, select the question data step from the question list to continue, if no, generate an interview end prompt message.

18. The method according to claim 6, characterized in that Also includes: Receiving a resume submission operation from the client for the target position; In response to the resume submission operation, sending an interview invitation notification to the client; The obtaining of the interview request sent by the client includes: The interview request sent by the client by triggering the interview invitation notification is obtained.

19. A computing device, characterized in that including a processing component and a storage component; The storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement the training method of the dialogue model as described in any one of claims 1 to 5, or to implement the artificial intelligence interview method as described in any one of claims 6 to 18.

20. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by the processing component, it implements the training method of the dialogue model as described in any one of claims 1 to 5, or implements the artificial intelligence interview method as described in any one of claims 6 to 18.

21. A computer program product, characterized in that It includes a computer program / instruction, which, when executed by a processing component, implements the training method of the dialogue model as described in any one of claims 1 to 5, or implements the artificial intelligence interview method as described in any one of claims 6 to 18.

Images