A data processing method and device, computer equipment and readable storage medium

Abstract

This application discloses a data processing method, apparatus, computer device, and readable storage medium. The method includes: inputting acquired data objects to be processed into a data processing model for processing to obtain a target processing result; the data processing model is obtained by training a label prediction model in conjunction with an uncertainty estimation model; during training, based on the uncertainty target value obtained from processing test samples by the uncertainty estimation model and the pseudo-labels obtained from processing test samples by the label prediction model, target test samples are determined from the test sample set; the training sample set is updated based on each target test sample and its corresponding pseudo-label; the label prediction model is trained using the updated training sample set; and the data processing model is determined based on the trained label prediction model. Using this scheme, test samples can be automatically labeled during model training, resulting in high labeling efficiency and improving model training efficiency.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a data processing method, apparatus, computer equipment, and readable storage medium. Background Technology

[0002] In recent years, with the gradual development of machine learning technology, various machine learning algorithms have emerged. Based on different learning methods, they can be divided into supervised learning, unsupervised learning, semi-supervised learning, reinforcement learning, etc. Thanks to the development of data, algorithms, and computing power, deep learning technology with large-scale training data has made breakthrough progress.

[0003] Typically, training a network using fully labeled training data allows for its application to data processing in relevant scenarios. However, the availability of fully labeled data also implies higher annotation costs, especially in specialized domains. Annotation and full-scale training become challenging when dealing with large datasets. Active learning, which trains models with as few manually labeled samples as possible and quickly learns useful information from highly uncertain samples, significantly alleviates these problems. However, it usually relies on a query strategy to select valuable samples from unlabeled data for manual annotation, and then adds these samples to the model's training to iterate. This training method still depends on human expert annotation, which is inefficient and negatively impacts model training efficiency. Summary of the Invention

[0004] This application provides a data processing method, apparatus, computer device, and readable storage medium that can automatically label test samples during model training, achieving high labeling efficiency and improving model training efficiency.

[0005] One embodiment of this application provides a data processing method, including:

[0006] Get the data object to be processed;

[0007] The data object to be processed is input into the data processing model for processing to obtain the target processing result;

[0008] The data processing model is obtained by training the label prediction model in conjunction with the uncertainty estimation model. During the training process, the target test samples are determined from the test sample set based on the uncertainty target value obtained by processing the test samples included in the test sample set by the uncertainty estimation model and the pseudo labels obtained by processing the test samples by the label prediction model. The training sample set is updated based on each target test sample and each target test sample's pseudo label. The label prediction model is then trained using the updated training sample set. Finally, the data processing model is determined based on the trained label prediction model.

[0009] One embodiment of this application provides a data processing apparatus, including:

[0010] The acquisition module is used to acquire data objects to be processed.

[0011] The processing module is used to input the data object to be processed into the data processing model for processing to obtain the target processing result. The data processing model is obtained by training the label prediction model in combination with the uncertainty estimation model. During the training process, the target test samples are determined from the test sample set based on the uncertainty target value obtained by processing the test samples included in the test sample set by the uncertainty estimation model and the pseudo labels obtained by processing the test samples by the label prediction model. The training sample set is updated based on each target test sample and each target test sample's pseudo label. The label prediction model is trained using the updated training sample set. The data processing model is determined based on the trained label prediction model.

[0012] One embodiment of this application provides a computer device, including: a processor, a memory, and a network interface; the processor is connected to the memory and the network interface, wherein the network interface is used to provide network communication functions, the memory is used to store program code, and the processor is used to call the program code to execute the data processing method of this application embodiment.

[0013] One aspect of this application provides a computer-readable storage medium storing a computer program, which includes program instructions. When executed by a processor, the program instructions perform the data processing method described in this application.

[0014] Accordingly, embodiments of this application provide a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform a data processing method provided in one aspect of embodiments of this application.

[0015] In this embodiment, training the label prediction model using an uncertainty estimation model yields a data processing model, i.e., the trained label prediction model. During training, the label prediction model is used to label test data, generating pseudo-labels. Based on a reasonable selection strategy, the pseudo-labels and the uncertainty target value obtained from processing the test data using the uncertainty estimation model are used to actively select relatively reliable test samples (i.e., target test samples) from the test sample set. The training sample set updated with these test samples and their pseudo-labels is then used to retrain the model until the stopping criterion is met, resulting in the data processing model. This process is an automated, proactive learning method. The training process actively selects valuable samples and uses the automatically generated pseudo-labels from the model to train the model. Pseudo-labels replace manually labeled samples, effectively avoiding manual intervention and saving labeling costs. Furthermore, automated labeling improves labeling efficiency compared to manual labeling, thereby enhancing model training efficiency. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a network architecture diagram of a data processing system provided in an embodiment of this application;

[0018] Figure 2 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0019] Figure 3 This is a schematic diagram of an application interface for a reading aloud question type provided in an embodiment of this application;

[0020] Figure 4 This is a schematic diagram of an application interface for an open-ended question type provided in an embodiment of this application;

[0021] Figure 5 This is a business process diagram of a pseudo-label-based spoken language assessment model provided in an embodiment of this application;

[0022] Figure 6 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0023] Figure 7 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0024] Figure 8This is a schematic diagram of the functional framework of an automated active learning scheme provided in an embodiment of this application;

[0025] Figure 9 This is a schematic diagram comparing the test results of an automated active learning scheme provided in this application embodiment applied to an oral exam;

[0026] Figure 10 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application;

[0027] Figure 11 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] The data processing scheme provided in this application is an automated, proactive learning scheme, belonging to Machine Learning (ML) within the field of artificial intelligence. Artificial Intelligence (AI) is the theory, methods, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to have perception, reasoning, and decision-making functions. AI technology is a comprehensive discipline involving a wide range of fields, encompassing both hardware and software technologies. Basic AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operating / interactive systems, and mechatronics. AI software technologies mainly include computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0030] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0031] When the data to be processed in this solution is audio data, the solution provided in this application also involves speech technology and natural language processing technology in the field of artificial intelligence. Key technologies in speech technology include Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and voiceprint recognition. Enabling computers to hear, see, speak, and feel is the future direction of human-computer interaction, with speech being one of the most promising methods. Natural Language Processing (NLP) is an important direction in computer science and artificial intelligence. It studies various theories and methods that enable effective communication between humans and computers using natural language. NLP is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language, that is, the language people use in daily life, and thus it is closely related to linguistic research. Natural language processing technologies typically include text processing, semantic understanding, machine translation, robot question answering, and knowledge graphs.

[0032] Please see Figure 1 This is a network architecture diagram of a data processing system provided in an embodiment of this application, such as... Figure 1As shown, the system includes multiple terminal devices 101 and a server 100, which communicate with each other via wired or wireless means. The terminal devices 101 can be smartphones, tablets, smart wearable devices, personal computers, etc., while the server 100 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The main functions of the terminal devices 101 and the server 100 in this solution are as follows:

[0033] Terminal device 101 can install and run functional clients (such as learning and education applications or web application platforms). These clients can collect and store data objects to be processed through terminal device 101. These data objects can be audio data, image data, or other types of data, such as voice data collected by the terminal device 101's microphone or image data collected by its camera. Terminal device 101 sends these data objects to server 100. Server 100 can be considered as the functional client processing the data using its corresponding algorithm, obtaining the processing result, and then sending the result back to terminal device 101. Terminal device 101 can display the processing result in an appropriate format. Taking an oral exam application as an example, terminal device 101 can record the user's answers or recitations, obtaining audio data, and then submit the audio data to the backend server 100. The backend server scores the audio data and returns the final score to the terminal device, i.e., terminal device 101 outputs the exam score or assessment score. Or, taking the application of plant species identification as an example, when a user captures a picture containing a plant through terminal device 101, the terminal device 101 may display "Identifying". This process is that the picture is sent to the backend server 100, and the backend data processing algorithm is used to identify the plant species included in the picture. Finally, the terminal device 101 may output the name of the plant and a series of related information.

[0034] Server 100 can be equipped with a data processing model to process received data objects and obtain processing results. Accordingly, the data processing model is constructed using relevant algorithms, which may include functions for extracting features from the data objects. For example, audio data can be recognized using Automatic Speech Recognition (ASR) technology, converting the audio data into text data, and then extracting text features by extracting keywords. Simultaneously, acoustic features of the audio data can be obtained using voiceprint recognition technology. These feature data are input into a trained network for processing to obtain the processing results, which are then sent to terminal device 101. Typically, due to business requirements, server 100 can process multiple data objects sent by terminal devices 101 in parallel to obtain processing results; there is no limit to the number of data objects processed concurrently by server 100. The algorithm corresponding to the training process of the data processing model can also be mounted on the server 100. During training, it includes a label prediction model and an uncertainty estimation model. The server 100 trains by receiving labeled sample data sent by the terminal device. It uses the label prediction model that has reached the expected level of training to predict unlabeled sample data. It also combines the predicted pseudo-labels with the uncertainty target value obtained from the uncertainty estimation model to filter out valuable sample data. Then, it updates the training sample set and trains the model again until the stopping condition is met. The optimal label prediction model obtained from multiple trainings is determined as the data processing model. The relevant calculations involved in this process are supported by the computing power of the server 100.

[0035] It can be observed that the data processing model mounted on server 100 can process the data objects collected by terminal device 101 and obtain accurate processing results. This is because during the training process of the data processing model, pseudo-labels and uncertain target values ​​are used to filter target test samples, continuously iterating and optimizing the model, and then applying the model with the best training effect to the actual data processing scenario. Furthermore, the addition of pseudo-labels during the training process in this embodiment enables automated active learning, avoiding the consumption of labeling resources by manual annotation, and at the same time, efficiently optimizing the model.

[0036] For ease of understanding, the following embodiments are provided to further illustrate the data processing scheme of this application. It is understood that the data processing method provided in the embodiments of this application can be executed by a computer device (such as the terminal device 101 or server 100 described above). For ease of understanding, the embodiments provided below are all described with the server as the execution subject.

[0037] Please see Figure 2 , Figure 2This is a flowchart illustrating a data processing method provided in an embodiment of this application. The data processing method may include at least the following steps S101 to S102:

[0038] S101, Obtain the data object to be processed.

[0039] In one embodiment, the data object to be processed can be audio, text, image, video, or other data. The server can receive the data object to be processed automatically sent by the terminal device online or offline, or it can actively pull these data objects to be processed from the terminal device. There are no restrictions on the type of data object to be processed, the specific acquisition method, or the scale.

[0040] S102, input the data object to be processed into the data processing model for processing, and obtain the target processing result.

[0041] In one embodiment, the target processing result obtained after the data object to be processed by the data processing model may also be different depending on the application scenario. In this embodiment, the processing process involved in this solution is described using an English oral exam as an application scenario. An application (APP) with English oral exam functionality is installed and running on the terminal device. The types of questions in the oral exam cover objective questions (such as reading aloud) and subjective questions (such as picture description and oral composition). Correspondingly, the data object to be processed includes the voice data entered by the user through the application client, such as reading a passage of English in the reading aloud question type, or an open-ended answer to an English question in the subjective question type, as well as question information and corresponding reference answers. After collecting the user's voice data, the terminal device sends it to the server along with the question information. The server then combines the received voice data, question information (which is text data), and stored audio data (or text data) of the reference answers with the data processing model. This model includes a feature extraction module, which extracts acoustic and text features from the voice and text data. In the actual processing, these two features are used as input to the data processing model to obtain the target processing result. In this application scenario, the target processing result is the user's oral exam score for a specific exam question. The oral exam score can be displayed to the user as a percentage, or as a grade or star rating. The format in which the target processing result is displayed on the terminal device is not limited here.

[0042] For example, please see Figure 3 This is a schematic diagram of an application interface for a reading aloud question type provided in an embodiment of this application. Users can click... Figure 3In sub-image (a), the "Start Reading Aloud" button allows the terminal device to begin recording, and the user can start reading along with the sentence. When the user no longer wants to continue reading or has finished reading, they can click the button. Figure 3 The "End Reading" button in sub-diagram (b) stops the recording on the terminal device. At this point, the backend server can retrieve the audio recorded by the user on the terminal, process it, and return the target processing result, i.e., the evaluation result displayed on the terminal device's screen to the user. Figure 3 Subgraph (c) shows the sentence evaluation results for 4 stars. Please refer to [link / reference]. Figure 4 This is a schematic diagram of an application interface for an open-ended question type provided in an embodiment of this application, and Figure 3 Similarly, when a user clicks... Figure 4 In sub-figure (a), the "Start Recording" button allows the terminal device to begin collecting voice data. When the user clicks it... Figure 4 The "End Recording" button in sub-diagram (b) indicates that recording is complete, and the terminal device screen will display the scoring results processed and returned by the server. Figure 4 The final evaluation result for subgraph (c) is a score of 81 out of 100.

[0043] In one embodiment, the data processing model described above is obtained by training a label prediction model in conjunction with an uncertainty estimation model. During training, target test samples are determined from the test sample set based on the uncertainty target values ​​obtained by processing the test samples included in the test sample set using the uncertainty estimation model, and pseudo-labels obtained by processing the test samples using the label prediction model. The training sample set is then updated based on each target test sample and its pseudo-label. The label prediction model is trained using the updated training sample set, and the data processing model is determined based on the trained label prediction model. Therefore, the data processing model described above, applied to the oral exam score evaluation scenario, can also be called an evaluation model trained with pseudo-labels (or an oral evaluation model based on pseudo-labels). The method for obtaining the data processing model can be found in the following embodiments, and will not be elaborated upon here.

[0044] Please see Figure 5 This is a business process diagram of a pseudo-label-based oral assessment model provided in an embodiment of this application. The user opens an oral exam app on their terminal device, and the screen displays the oral exam questions, as described above. Figure 3 or Figure 4 The content shown in sub-image (a) is then displayed. Users click the button in the app that indicates "start recording" to answer questions. The app sends the audio and the read-aloud text (or question information) to the server via the terminal device. Of course, in... Figure 5The diagram showing the main content transmitted between the app and the server is audio data. The server can also retrieve the text or question information from its database. After feature extraction, the server sends the audio and reference answers (for reading-aloud questions, the reference answer is the text being read, i.e., the question information; for open-ended questions, it's the corresponding answer) to a pseudo-label-based evaluation model. The model returns the score to the server, which then sends the final score back to the app, allowing the user to view their final exam score. It should be noted that... Figure 5 The feature extraction module and evaluation model are separated from the server as a separate part to better illustrate the overall processing logic. These modules can be deployed on, for example, Figure 5 The functional modules shown in the server can also be deployed in addition to those shown in the server. Figure 5 There are no restrictions on the other one or more servers shown.

[0045] It should be noted that the above data processing model can serve as a general-purpose model applicable to various scenarios. However, the training samples, test samples, and data objects to be processed during the model's training must be tailored to the specific application scenario. For example, if the data processing model has image recognition capabilities, then the training samples, test samples, and data objects to be processed will all be image data. This allows for the separation of framework and resources in the training of the data processing model. In other words, in different application scenarios, only the corresponding sample data needs to be used to train the model, without changing the general training process.

[0046] In summary, the embodiments of this application have at least the following advantages:

[0047] During model training, a label prediction model is used to automatically label test samples, i.e., predict pseudo-labels, to train the model efficiently. At the same time, not all pseudo-label samples are learned, but reliable target test samples are actively selected for learning. This can reduce the error caused by inaccurate pseudo-labels, ensure the accuracy of model learning, and make model training more efficient. Using a data processing model that combines pseudo-labels to automatically and actively learn data objects to process the data objects, reliable target processing results can be obtained.

[0048] Please see Figure 6 , Figure 6 This is a flowchart illustrating a data processing method provided in an embodiment of this application. The data processing method may include at least the following steps S201 to S204:

[0049] S201, Obtain the training sample set and the test sample set.

[0050] In one embodiment, the training sample set includes multiple training samples and reference labels for each training sample, and the test sample set includes multiple test samples. This training sample set is the original training sample set, which will be updated subsequently to train the model. As mentioned in the previous embodiments, due to different application scenarios, the training and test samples required during training should also be matched according to the specific application scenario. Taking the oral exam application scenario as an example, the training samples here are audio data with reference labels, i.e., labeled data, while the test samples are audio data without reference labels, i.e., unlabeled data. The reference label refers to the score manually labeled for the audio data, which can be considered as the real label. Of course, in other application scenarios, such as image recognition, the reference label refers to the category to which a certain image belongs, the training samples are image data with category labels, and the test samples are image data without category labels. Typically, to save on the cost of manual labeling, the number of training samples is often less than the number of test samples. Using a small number of training samples to label more unlabeled samples, achieving time-saving and labor-saving training and efficient model performance improvement, is also a point that the field of machine learning strives to improve. There is no limit to the number of training samples and test samples. For ease of understanding, the following examples all use the application scenario of oral examination as an illustration.

[0051] S202, the initial label prediction model is trained using the training sample set to obtain the first label prediction model after initial training, and the uncertainty estimation model is trained using the training sample set to obtain the uncertainty estimation model after training.

[0052] In one embodiment, the initial label prediction model is an initial model that has not been trained with any data. The initial model is trained for the first time using training samples and reference labels in the training sample set to obtain the initially trained label prediction model, which is referred to here as the first label prediction model. Whether it is the initial label prediction model, the first label prediction model, or a label prediction model that has been trained a second or multiple times with an updated training sample set, in this embodiment, they are collectively referred to as the label prediction model or the base model.

[0053] Optionally, the initial label prediction model includes a feature extraction module. The step of training the initial label prediction model using a training sample set to obtain the first label prediction model after initial training may include: inputting the training samples included in the training sample set into the initial label prediction model; using the feature extraction module included in the initial label prediction model to extract features from the training samples to obtain feature data for each training sample; during the training of the initial label prediction model using the feature data of each training sample, using the initial label prediction model to perform label prediction processing on the feature data of each training sample to obtain the predicted label for each training sample; determining loss information based on the predicted label and reference label of each training sample, and adjusting the model parameters of the initial label prediction model based on the loss information; and determining the first label prediction model after initial training based on the initial label prediction model after the model parameter adjustment.

[0054] The feature extraction module extracts effective feature data based on training data (i.e., training samples). In the oral exam application scenario, the training data consists of audio and corresponding artificial labels (i.e., reference labels). The feature data extracted by the feature extraction module includes text features and acoustic features. Text features mainly include semantic features, pragmatic features, keyword features, and text fluency features. Keyword features mainly include extracting keywords from the standard answer and keywords from the answer content, calculating precision and recall, etc. Pragmatic features include the diversity of words and sentence structures in the answer content, as well as the grammatical accuracy of the answer content based on language model analysis. Semantic features include thematic features of the answer content, TF-IDF (term frequency-inverse document frequency) features, etc. Acoustic features are mainly divided into pronunciation accuracy, pronunciation fluency, and pronunciation prosody. Pronunciation accuracy refers to pronunciation scores at the phoneme, word, and sentence levels. Pronunciation fluency includes speech rate features during pronunciation and features based on duration statistics, such as the average duration of pronunciation segments and the average pause duration between pronunciation segments. Pronunciation prosody includes assessments of pronunciation rhythm, word stress accuracy in sentences, and sentence boundary intonation.

[0055] The construction of a label prediction model based on label data (i.e., reference labels) and the extracted acoustic and text features involves training an initial label prediction model using feature data and label data from training samples, and evaluating the trained label prediction model. This initial label prediction model can be a regression model or a classification model, fitting human scores. Optionally, the regression model can be a traditional regression model, such as KNN (k-Nearest Neighbor), SVR (support vector regression), GBT (Gradient Boosting Tree), etc., or a deep neural network model that obtains the final score through multi-layer forward propagation. The type of initial label prediction model is not limited here. The process of training the initial label prediction model using feature data mainly involves using the initial label prediction model to perform label prediction processing on the feature data extracted by the feature extraction module, obtaining the predicted labels of the training samples to which the feature data belongs. If we only focus on the model's input and output, then all training samples in the training sample set are input into the initial label prediction model, and the output is the predicted label of all training samples. For example, if 100 training samples are input into the initial label prediction model, then the output is the predicted label of all 100 training samples. Unlike the reference label, which is the true value manually labeled for the training samples, the predicted label is a predicted value obtained by the initial label prediction model through processing the training samples. It may not be the same as the reference label, and since the predicted label is only changed once due to subsequent adjustments to the model parameters, the predicted label can be determined based on the reference label and the predicted label. This loss information (such as the loss value) is used to measure the degree of inconsistency between the model's predicted value and the true value. Based on this loss information, the model parameters of the initial label prediction model are adjusted once to obtain the adjusted initial label prediction model. However, one training session may not yield the optimal label prediction model. Therefore, after the model parameters are adjusted once, the label prediction model is trained again. The convergence of the loss information is used to determine whether further adjustments to the model parameters are needed until the label prediction model achieves the highest prediction accuracy for the training samples or the loss information is within the error range, thus obtaining the first label prediction model. Alternatively, the training process can involve dividing the labeled data into two parts: one part serves as the training sample set for training the model, and the other part serves as the validation sample set for testing the model after parameter adjustments. In other words, the model is tested by inputting sample data with real labels (i.e., the validation sample set) to measure whether the prediction effect meets the expected results. If not, the above training steps are repeated. Through continuous training and parameter adjustment, the optimal label prediction model is obtained through continuous iteration and updates, and the optimal label prediction model is used as the first label prediction model.

[0056] Furthermore, training the uncertainty estimation model using the training sample set to obtain the trained uncertainty estimation model can be similar to the process of training the initial label prediction model, which will not be elaborated further. The uncertainty estimation model in this embodiment can also be called an uncertainty analysis model or uncertainty module. Based on extracted feature data (e.g., text features and acoustic features), there are various methods for constructing uncertainty analysis models, such as typical Gaussian process regression, Monte Carlo dropout, and deep mixture density networks. Gaussian process regression uses a Gaussian distribution to model the output and determine the mean and variance of each prediction result. This method uses variance as a measure of uncertainty; the larger the variance, the greater the uncertainty. Monte Carlo dropout uses multiple models to analyze the uncertainty of the model. It assumes that for uncertain data, the output of each model is diverse; the more diverse the output, the greater the uncertainty. Deep mixture density networks are similar to Gaussian process modeling, modeling the mean and variance of the results. This method also uses variance as a measure of uncertainty; the larger the variance, the greater the uncertainty. In the embodiments of this application, the uncertainty analysis model based on the hybrid density network and Gaussian process can be applied to the aforementioned oral examination scenario (such as an oral examination system).

[0057] It should be noted that the feature extraction module mentioned above can be a shared module of the label prediction model and the uncertainty analysis model. The resulting feature data will be used by both the uncertainty estimation model and the label prediction model.

[0058] S203, input the test samples included in the test sample set into the first label prediction model for label prediction processing to obtain the pseudo label of each test sample, and input the test samples into the trained uncertainty estimation model for uncertainty analysis processing to obtain the uncertainty target value of each test sample.

[0059] In one embodiment, the test sample set includes unlabeled test samples. The first label prediction model, trained on labeled training samples, can predict the test samples relatively accurately because it is trained to its optimal state using the original training sample set. However, the model's accuracy may not be 100% due to potential errors in the labeling of training samples and the possibility that some test samples in the test sample set were not learned by the first label prediction model in the original training set. Therefore, the accuracy of the predicted labels is uncertain, and the results of the first label prediction model's predictions for the test samples are referred to as "pseudo-labels." From the perspective of the label prediction model's processing, it is a pseudo-label algorithm, belonging to semi-supervised learning (SSL). This is a learning method that combines supervised and unsupervised learning. Semi-supervised learning uses a large amount of unlabeled data, along with labeled data, to perform pattern recognition. However, determining the reliability of the obtained pseudo-labels requires using an uncertainty estimation model to perform uncertainty analysis on the test samples and obtain the uncertainty target value for the test samples. For example, the mean and variance output from the Gaussian process regression model can be used as the uncertainty target value. This uncertainty target value describes the uncertainty of the prediction result, that is, how high the reliability or credibility (confidence level) of the prediction result is when it is correct. The higher the uncertainty, the lower the confidence level. In addition, it should be noted that the input test sample here can be all test samples in the test sample set, or it can be a partial test sample; there is no restriction here.

[0060] In one embodiment, the step of processing test samples using an uncertainty estimation model to obtain the corresponding uncertainty target value may include: performing uncertainty analysis on the test samples using the trained uncertainty estimation model to obtain the data uncertainty prediction value and the model uncertainty prediction value corresponding to the test samples; and determining the uncertainty target value of the test samples based on the data uncertainty prediction value and the model uncertainty prediction value. Here, the test samples are also the samples processed by the first label prediction model. In this embodiment, the uncertainty estimation model uses a mixture density network and a Gaussian process to model data uncertainty and model uncertainty respectively. Simply put, data uncertainty is caused by noise in the training sample data, and model uncertainty is caused by incomplete model training. By modeling these two types of uncertainty, the uncertainty of the prediction result can be described more accurately. Applied to the oral examination system, the corresponding data uncertainty prediction value includes the mean and variance of the predicted score of any test sample by the mixture density network, and the model uncertainty prediction value includes the mean and variance of the output of the Gaussian process. By weighted summing of the two uncertainty prediction values, the final uncertainty target value can be obtained. Both model uncertainty modeling and data uncertainty modeling satisfy a Gaussian distribution, that is, the outputs of the mixture density network and the Gaussian process are both possible distributions of the target value, expressed as follows:

[0061] A~N(μ1,σ1 2 ), B~N(μ2,σ2) 2 (1)

[0062] Where A represents the model uncertainty, μ1, σ1 2 Let μ² and σ² represent the mean and variance of the Gaussian process output, respectively, and B represent the data uncertainty. 2 These represent the mean and variance of the mixed density network output, respectively.

[0063] The weighted sum of the two uncertain predictions yields the following expression:

[0064] λA+(1-λ)B~N(λμ1+(1-λ)μ2,λ 2 σ1 2 +(1-λ) 2 σ2 2 (2)

[0065] Here, λ represents the uncertainty weight, which ranges from 0 to 1. When the model has high uncertainty and modeling accuracy, λ can be set to a value greater than 0.5.

[0066] The uncertainty target value includes the mean and variance as shown in equation (2), where the variance represents the uncertainty outcome, denoted as uncertainty = λ. 2 σ12 +(1-λ) 2 σ2 2 It should be noted that other networks or models that describe uncertainty, such as Bayesian neural networks, can also be used to construct uncertainty analysis models for uncertainty estimation models. There are no restrictions on the type of uncertainty estimation model.

[0067] S204. Determine target test samples from the test sample set based on the uncertainty target value and pseudo-label of each test sample, update the training sample set using each target test sample and its pseudo-label, retrain the first label prediction model using the updated training sample set, and determine the data processing model based on the second label prediction model obtained from the retraining.

[0068] In one embodiment, based on reasonable screening rules, a label prediction model can predict the pseudo-label of a test sample, and an uncertainty target value (which may include the mean and variance mentioned above) can be output by an uncertainty estimation model. Test samples that meet the conditions in the test sample set are then used as target test samples. Subsequent updates to the training sample set using the selected target test samples can be achieved by adding the target test samples and their corresponding pseudo-labels to the training sample set to form a new training sample set. This new training sample set is the original training sample set (or initial training sample set), and the labels of the training samples are manually labeled reference labels. In this case, the new training sample set includes both reference labels and pseudo-labels. Alternatively, multiple target test samples can be used as separate new training sample sets to replace the original training sample set. The new training sample set is then used to retrain the first label prediction model to obtain the second label prediction model. The above describes the first update of the training samples. In the Nth (N greater than 2) training iteration, the determined target test samples and their pseudo-labels can be used to update the training sample set after the previous training update, or the original training sample set can be updated. For example, during the third training iteration, the update could be based on the initial training sample set or on the training samples updated during the second training iteration; there is no restriction on this. It should be noted that this single training iteration of the model refers to the process of training the label prediction model to its optimal state using a newly updated training sample set.

[0069] Optionally, the model can be updated through incremental learning or relearning to incorporate pseudo-labels into the machine learning model. Compared to incorporating manually labeled labels into the machine learning model in active learning, this method saves significant manual labeling costs and improves training efficiency, making it an excellent training approach. It should be noted that the number of times the label prediction model is trained using the training sample set (e.g., an updated training sample set) is not limited to two. Training stops when the base model meets a stopping condition. This stopping condition could be an evaluation of the prediction performance on the test samples; if the corresponding evaluation metrics are met, training can cease. For example, if the pseudo-label prediction performance of the second label prediction model and the first label prediction model is roughly the same for the test samples, then the second label prediction model can be used as the final data processing model. Conversely, if the prediction performance of the second label prediction model is significantly improved compared to the first label prediction model, then the second label prediction model needs to be trained again using the updated training sample set to obtain the third label prediction model, which is then evaluated again. In short, as long as the prediction performance of the currently trained label prediction model for the test samples converges (e.g., the prediction performance is the same as or only slightly different from that of the label prediction models obtained in the previous one or more training sessions), training can stop. Otherwise, iterative model training is required. It is important to note that the label prediction model trained using the updated training samples should only be used to process the test data after the expected performance has been achieved. In other words, it should be trained thoroughly in one go to ensure the comprehensiveness and accuracy of each learning iteration.

[0070] In summary, the embodiments of this application have at least the following advantages:

[0071] By combining active learning with pseudo-label training—specifically, combining the pseudo-label training method of semi-supervised learning with the method of selecting valuable test samples based on uncertain results in active learning—pseudo-labels are assigned to unlabeled test data based on the trained label prediction model. The model is then trained a second time or multiple times using effective pseudo-label test samples. Valid pseudo-labels are the label data of target test samples selected from the test samples based on the uncertainty target value output by the uncertainty estimation model. These target test samples are uncertain samples with high information content, i.e., samples that the current model finds difficult to distinguish. Updating the training sample set using these uncertain samples makes the model training more targeted, more efficient, and results in a more significant improvement. Simultaneously, the fully automated training process is achieved, further reducing the cost of manual annotation. The entire process aims to efficiently train a data processing model with the highest possible performance using fewer manually annotated samples, saving annotation and training resource costs.

[0072] Please see Figure 7 , Figure 7This is a flowchart illustrating a data processing method provided in an embodiment of this application. The data processing method may include at least the following steps S301 to S306:

[0073] S301, Obtain the training sample set and the test sample set.

[0074] S302, the initial label prediction model is trained using the training sample set to obtain the first label prediction model after initial training, and the uncertainty estimation model is trained using the training sample set to obtain the uncertainty estimation model after training.

[0075] S303, input the test samples included in the test sample set into the first label prediction model for label prediction processing to obtain the pseudo label of each test sample, and input the test samples into the trained uncertainty estimation model for uncertainty analysis processing to obtain the uncertainty target value of each test sample.

[0076] For the specific implementation methods of steps S301 to S303 above, please refer to the above. Figure 2 The steps S201 to S203 in the corresponding embodiments will not be described again here.

[0077] S304, determine the mean uncertainty based on the first uncertainty target value of each test sample.

[0078] In one embodiment, the uncertainty target value includes a first uncertainty target value and a second uncertainty target value, which are the values ​​output by the uncertainty module (i.e., the uncertainty estimation model) for the prediction result of a certain test sample. The first uncertainty target value can correspond to the variance (i.e., the uncertainty result or simply uncertainty) in equation (2), and can be denoted as uncertainty_i = λ. 2 σ1 2 +(1-λ) 2 σ2 2 The second uncertainty target value corresponds to the mean in equation (2), which can be denoted as p. uncertain_i =λμ1+(1-λ)μ2. Taking the example of all test samples in the test sample set undergoing uncertainty estimation modeling, assuming the test sample set includes n test samples, the first uncertainty target value for all test samples in the test sample set is denoted as Uncertain = [uncertainty_1, uncertainty_2, ..., uncertainty_n]. The expression for calculating the mean uncertainty of this batch of test data is as follows:

[0079]

[0080] Where, mean uncertainLet n represent the mean of uncertainty, n represent the number of test samples, and uncertainty_i represent the first uncertainty target value.

[0081] S305, determine the target test sample from the test sample set based on the mean uncertainty, the first and second uncertainty target values ​​of each test sample, and the pseudo-label of each test sample.

[0082] In one embodiment, this step may include: sorting the test samples according to the first uncertainty target value from largest to smallest to obtain sorted test samples; selecting test samples whose second uncertainty target value is greater than or equal to the uncertainty mean as candidate test samples; and determining the candidate test samples whose second uncertainty target value and the corresponding pseudo-label have a difference less than a difference threshold as target test samples. Determining target test samples from the test sample set is essentially the process of selecting relatively reliable samples from high uncertainty samples. Similar to the pseudo-label algorithm using pseudo-label samples with high confidence, this embodiment selects pseudo-label samples with low confidence, i.e., high uncertainty samples. Since these samples often contain a lot of information, incorporating them into secondary training can significantly improve the model. However, since the pseudo-labels of high uncertainty samples often have large errors and may be incorrect labels, a reasonable selection strategy is used to select effective pseudo-labels from high uncertainty samples, combining multiple uncertainty modeling results. Assume the base model predicts a pseudo-label of p for a certain test sample. base_i The uncertainty module outputs a first uncertainty target value of uncertaint_i and a second uncertainty target value of p. uncertain_i The above screening strategy corresponds to the following screening process:

[0083] (1) Sort the uncertainty (i.e. the first uncertainty target value) of all samples in the test set from largest to smallest;

[0084] (2) Iterate through the uncertain samples sorted from high to low using a For loop:

[0085] If p uncertain_i <mean uncertain :

[0086] Break stops the loop;

[0087] If p base_i -p uncertain_i <threshold:

[0088] Add this sample to the filtered samples;

[0089] (3) The final selected test samples (i.e., target test samples) and their pseudo-label records p pseudo , Where k is the number of target test samples.

[0090] In the above screening process, the threshold, or difference threshold, is used to measure the difference between the pseudo-label and the second uncertainty target value of a test sample. Taking the oral exam application scenario as an example, the pseudo-label of the test sample is the score, and the second uncertainty target value is the mean of the predicted score of a test sample, which is the point with the highest probability in the possible distribution of predicted scores. Therefore, theoretically, the closer the difference between these two values ​​is to 0, the higher the accuracy of the pseudo-label. Test samples processed by the uncertainty estimation model can all be regarded as uncertain samples. Test samples with greater uncertainty are also called high uncertainty samples. Since active learning selects high uncertainty samples for learning based on a certain query strategy, sorting them from high to low uncertainty is more convenient for the subsequent screening process. The sorted test samples are iterated through, starting with the test sample with the highest uncertainty. Further filtering is performed by comparing the second uncertainty target value (i.e., the mean of the prediction results) with the uncertainty mean. That is, if the second uncertainty target value is less than the uncertainty mean during the iteration, the loop is broken using a break statement, and the current loop is interrupted, meaning that subsequent uncertain samples are no longer iterated. Conversely, if the second uncertainty target value is greater than or equal to the uncertainty mean (i.e., candidate test samples), it is necessary to continue to determine whether the difference between the second uncertainty target value and the pseudo-label is within the difference threshold. If so, the test sample can be added to the filtered samples. However, the existence of the break statement may lead to incomplete coverage. Therefore, it is also possible to first iterate through all uncertain samples, filter out those that meet the first requirement (i.e., the first uncertainty target value is greater than or equal to the uncertainty mean), and then iterate through all samples that meet the first requirement, filter out those that meet the second requirement (i.e., the difference between the second uncertainty target value and the pseudo-label is within the difference threshold). By employing a dual screening process—combining the first and second uncertainty target values ​​obtained from uncertainty modeling to select reliable test samples (i.e., target test samples) from high-uncertainty samples and applying them to the next training process—the model can learn more comprehensively, ensuring the accuracy and effectiveness of its learning.

[0091] S306, update the training sample set using each target test sample and its pseudo-label, retrain the first label prediction model using the updated training sample set, and determine the data processing model based on the second label prediction model obtained from the retraining.

[0092] In one embodiment, after the initial label prediction model is trained for the first time using the original training sample set, a first label prediction model is obtained. The target test samples and pseudo-labels of each target test sample obtained through the above steps are used to update the training sample set. For example, these target test samples and model prediction values ​​(i.e., pseudo-labels) can be incorporated into the original training data to expand the original training set and train the first label prediction model for the second time. Since the labels of the selected samples are pseudo-labels, the prediction results of the base model (first label prediction model) may still have some errors. Therefore, the scheme provided in this embodiment combines the uncertainty results, assigns less weight to the pseudo-label samples and higher weight to the original training samples, and performs a second training by mixing them.

[0093] Optionally, retraining the first label prediction model using the updated training sample set and determining the data processing model based on the retrained second label prediction model may include: normalizing the uncertainty target value of each target test sample to obtain the first weight coefficient of each target test sample; retraining the first label prediction model using each target test sample, the first weight coefficient of each target test sample, the training samples included in the training sample set, and the second weight coefficient of each training sample; and determining the data processing model based on the retrained second label prediction model; wherein the first weight coefficient is less than the second weight coefficient. The value obtained by normalizing the uncertainty target value of each target test sample, including the first uncertainty target value, is used as the first weight coefficient, which is the weight assigned to the pseudo-label based on the uncertainty result. Since the uncertainty result after normalization is a value in the range of 0 to 1, the second weight coefficient of the original training sample is 1, and the first weight coefficient is always less than the second weight coefficient. Optionally, the uncertainty results of all test samples can also be normalized by maximum and minimum values, so that each test sample corresponds to a normalized uncertainty result, and the first weight coefficient of the target test sample can use the normalized uncertainty result corresponding to that test sample. The expression for normalizing all test samples in the test sample set is as follows:

[0094]

[0095] Among them, uncertain i Let represent the first uncertainty target value of the i-th test sample among all test samples, min(uncertain) represents the minimum value of the first uncertainty target value among all test samples, and max(uncertain) represents the maximum value of the first uncertainty target value among all test samples.

[0096] Assume the original training samples are y = [y1, y2, ..., y].m The final secondary training sample is Y_COMBINE. Therefore, combining the above, the secondary training sample can be represented as follows: in, This represents the pseudo-label of the i-th target test sample among the selected target test samples, i = 1, ..., k. The first weight coefficient of each target test sample is Norm(uncertain). i ), i = 1, ..., k. y i Let represent the reference label of the i-th training sample in the original training sample set, i = 1, ..., m, where m represents the number of original training samples, and the second weight coefficient of the training sample is 1.

[0097] It can be observed that using the normalized uncertainty result as the first weight coefficient results in inconsistent weights for each target test sample due to the varying uncertainty. This reduces errors caused by inaccurate pseudo-labels, while giving greater weights to more accurate pseudo-labels allows them to play a larger role in model training. Furthermore, this approach assigns less weight to the target test samples than to the original training samples to minimize error. Since pseudo-labels are predictions from the base model, accuracy cannot be guaranteed, and errors still exist. Giving less weight to pseudo-labeled test samples than to relatively accurate training samples minimizes the accumulation of errors from pseudo-labels, thus ensuring model accuracy. It should be noted that other methods can be used to assign weights to the target test samples and the second weight coefficients to the original training samples, but the principle remains the same: samples with high label accuracy should have greater weights than samples with low label accuracy.

[0098] In the representation of the above-mentioned secondary training samples, weights and labels are directly associated. These samples and their corresponding weights are input into the first label prediction model for secondary training to obtain the second label prediction model. The data processing model is determined based on the second label prediction model. If the second label prediction model achieves the expected results, it can be used as the data processing model. However, if the evaluation results of the second label prediction model do not meet expectations, further training of the second label prediction model is required. For details, please refer to the aforementioned embodiments; further elaboration is not provided here.

[0099] Based on the data processing scheme provided in the above embodiments, please participate. Figure 8This diagram illustrates the functional framework of an automated active learning scheme provided in this application, including feature extraction, base model training, model prediction, uncertainty modeling, sample selection, and secondary training by fusing uncertainties. The relevant content of the entire process has been described in detail in the above embodiments, and will be briefly described here.

[0100] First, the training data (i.e., training samples and artificial labels) is input into the feature extraction module to extract the feature data of the training samples. For multiple feature data, vector representation is usually used. These vectors are then input separately into the base model training and uncertainty modeling modules. In the base model training module, the model is trained using the feature data of the training samples and the artificial labels. In uncertainty modeling, a similar method is used to train the uncertainty estimation model. Next, unlabeled test data (i.e., test samples) is input into the trained base model to predict pseudo-labels. Similarly, in uncertainty modeling, the same unlabeled test data is input to obtain the uncertainty result (i.e., the uncertainty target value) for the corresponding test samples, including variance and mean. Based on the pseudo-labels and uncertainty results output from these two parts, and with appropriate filtering strategies, samples can be selected from the test data. Finally, the selected samples are fused with the training data, different weights are assigned to the samples, and then fed into the feature extraction module to obtain the corresponding feature data. Combined with the artificial labels and pseudo-labels, the entire model is trained a second time. Through iterative modeling, the optimal base model can be obtained and applied to the corresponding scenario. It can be observed that this scheme combines active learning with pseudo-label training, utilizing high-uncertainty samples with abundant information and effective pseudo-labels for secondary training. Since the pseudo-labels for high-uncertainty samples often have significant errors and may be incorrect, effective pseudo-labels are selected from the high-uncertainty samples based on various uncertainty modeling results, replacing manual intervention for secondary training. Simultaneously, because the labels of the selected samples are pseudo-labels, meaning the predictions of the base model may still have some error, the scheme further combines uncertainty results, assigning fewer weights to pseudo-label samples and higher weights to original training samples, mixing them for secondary training. Each step of the design aims to minimize the expected errors during model training, gradually improving model performance and ensuring the reliability and effectiveness of the training.

[0101] The effectiveness of this solution in relevant scenarios was tested using different question types in oral exams, including situational questioning, rapid response, oral composition, and semi-open-ended questions. Each question type included 250 training samples and 1400 test samples. Evaluation metrics included the Pearson correlation coefficient (a measure of the correlation (linear correlation) between two variables X and Y, with values ​​between -1 and 1, where X can be the true label and Y refers to the model's predicted value, i.e., the pseudo-label), and the consistency rate (the probability that the true label and the model's predicted value are less than a certain threshold). Results Figure 9 The image shows a comparison between the results of a single training iteration (base) and the results of an automated active learning approach (improve). The results demonstrate that the automated active learning approach effectively improves model performance. It should be noted that, to evaluate the model's predictive performance, the test samples used here are samples with true labels; however, the test samples are not labeled when inputting into the base model for prediction.

[0102] In summary, the embodiments of this application have at least the following advantages:

[0103] Based on a reasonable and effective screening strategy, valid test samples are selected by combining the uncertainty results output by the uncertainty estimation model. These samples are then integrated into the original training samples as the next training samples. During training, training samples carrying pseudo-labels are given lower weights and training samples carrying reference labels are given higher weights, taking into account the uncertainty results. This ensures that the samples have a certain amount of information while fully considering the error caused by the pseudo-labels of uncertain samples, thus reducing the error caused by pseudo-labels to a controllable range. By fusing samples with two different weights to train and update the model, the effectiveness of the training and the prediction accuracy of the final data processing model are guaranteed.

[0104] Please see Figure 10 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application. The data processing device can be a computer program (including program code) running on a computer device, for example, the data processing device is an application software; the data processing device can be used to execute corresponding steps in the method provided in the embodiment of this application. Figure 10 As shown, the data processing device 1000 may include: an acquisition module 1001 and a processing module 1002.

[0105] Module 1001 is used to acquire data objects to be processed;

[0106] The processing module 1002 is used to input the data object to be processed into the data processing model for processing to obtain the target processing result. The data processing model is obtained by training the label prediction model in combination with the uncertainty estimation model. During the training process, the target test samples are determined from the test sample set based on the uncertainty target value obtained by processing the test samples included in the test sample set by the uncertainty estimation model and the pseudo labels obtained by processing the test samples by the label prediction model. The training sample set is updated based on each target test sample and the pseudo labels of each target test sample. The label prediction model is trained using the updated training sample set. The data processing model is determined based on the trained label prediction model.

[0107] In one embodiment, the data processing device 1000 further includes a training module 1003 and a determination module 1004, wherein:

[0108] The acquisition module 1001 is also used to acquire a training sample set and a test sample set. The training sample set includes multiple training samples and reference labels for each training sample, and the test sample set includes multiple test samples.

[0109] The training module 1003 is used to train the initial label prediction model using the training sample set to obtain the first label prediction model after initial training, and to train the uncertainty estimation model using the training sample set to obtain the uncertainty estimation model after training.

[0110] The processing module 1002 is also used to input the test samples included in the test sample set into the first label prediction model for label prediction processing to obtain the pseudo label of each test sample, and to input the test samples into the trained uncertainty estimation model for uncertainty analysis processing to obtain the uncertainty target value of each test sample.

[0111] The determination module 1004 is used to determine the target test sample from the test sample set based on the uncertainty target value and pseudo label of each test sample, update the training sample set using each target test sample and its pseudo label, retrain the first label prediction model using the updated training sample set, and determine the data processing model based on the second label prediction model obtained from the retraining.

[0112] In one embodiment, the initial label prediction model includes a feature extraction module, and the training module 1003 is used to: input the training samples included in the training sample set into the initial label prediction model; use the feature extraction module included in the initial label prediction model to extract features from the training samples to obtain feature data of each training sample; during the training of the initial label prediction model using the feature data of each training sample, use the initial label prediction model to perform label prediction processing on the feature data of each training sample to obtain the predicted label of each training sample; determine loss information based on the predicted label and reference label of each training sample, and adjust the model parameters of the initial label prediction model based on the loss information; and determine the first label prediction model after initial training based on the initial label prediction model after the model parameter adjustment.

[0113] In one embodiment, the processing module 1002 is specifically used to: perform uncertainty analysis processing on the test sample using the trained uncertainty estimation model to obtain the data uncertainty prediction value and the model uncertainty prediction value corresponding to the test sample; and determine the uncertainty target value of the test sample based on the data uncertainty prediction value and the model uncertainty prediction value.

[0114] In one embodiment, the uncertainty target value of each test sample includes a first uncertainty target value and a second uncertainty target value. The determination module 1004 is specifically used to: determine the uncertainty mean based on the first uncertainty target value of each test sample; and determine the target test sample from the test sample set based on the uncertainty mean, the first uncertainty target value and the second uncertainty target value of each test sample, and the pseudo label of each test sample.

[0115] In one embodiment, the determining module 1004 is further configured to: sort each test sample in descending order of the first uncertainty target value to obtain sorted test samples; select test samples in the sorted test samples whose second uncertainty target value is greater than or equal to the uncertainty mean as candidate test samples; and determine candidate test samples in each candidate test sample whose difference between the second uncertainty target value and the corresponding pseudo label is less than the difference threshold as target test samples.

[0116] In one embodiment, the determining module 1004 is further configured to: normalize the uncertainty target value of each target test sample to obtain the first weight coefficient of each target test sample; retrain the first label prediction model using each target test sample, the first weight coefficient of each target test sample, the training samples included in the training sample set, and the second weight coefficient of each training sample; and determine the data processing model based on the second label prediction model obtained from the retraining; wherein the first weight coefficient is less than the second weight coefficient.

[0117] It is understood that the functions of each functional module of the data processing apparatus described in the embodiments of this application can be specifically implemented according to the methods in the above method embodiments, and the specific implementation process can be referred to the relevant descriptions in the above method embodiments, which will not be repeated here. In addition, the beneficial effects of using the same method will not be repeated here either.

[0118] Please see Figure 11 This is a schematic diagram of a server structure provided in an embodiment of this application. The server 110 may include a processor 1101, a memory 1102, a network interface 1103, and at least one communication bus 1104. The processor 1101 is used to schedule computer programs and may include a central processing unit, a controller, or a microprocessor. The memory 1102 is used to store computer programs and may include high-speed random access memory (RAM), non-volatile memory such as disk storage devices or flash memory devices. The network interface 1103 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface) to provide data communication functions. The communication bus 1104 is responsible for connecting the various communication components. This server 110 may correspond to the server 100 described above.

[0119] The processor 1101 can be used to call a computer program in memory to perform the following operations:

[0120] Get the data object to be processed;

[0121] The data to be processed is input into the data processing model for processing to obtain the target processing result. The data processing model is obtained by training the label prediction model in combination with the uncertainty estimation model. During the training process, the target test samples are determined from the test sample set based on the uncertainty target value obtained by processing the test samples included in the test sample set by the uncertainty estimation model and the pseudo labels obtained by processing the test samples by the label prediction model. The training sample set is updated based on each target test sample and each target test sample's pseudo label. The label prediction model is trained using the updated training sample set. The data processing model is determined based on the trained label prediction model.

[0122] In one embodiment, the processor 1101 is further configured to:

[0123] Obtain the training sample set and the test sample set. The training sample set includes multiple training samples and reference labels for each training sample. The test sample set includes multiple test samples.

[0124] The initial label prediction model is trained using the training sample set to obtain the first label prediction model after initial training; and the uncertainty estimation model is trained using the training sample set to obtain the uncertainty estimation model after training.

[0125] The test samples included in the test sample set are input into the first label prediction model for label prediction processing to obtain the pseudo label of each test sample. The test samples are also input into the trained uncertainty estimation model for uncertainty analysis processing to obtain the uncertainty target value of each test sample.

[0126] Target test samples are determined from the test sample set based on the uncertainty target value and pseudo-label of each test sample. The training sample set is then updated using each target test sample and its pseudo-label. The first label prediction model is then retrained using the updated training sample set. The data processing model is determined based on the second label prediction model obtained from the retraining.

[0127] In one embodiment, the processor 1101 is specifically configured to: input the training samples included in the training sample set into the initial label prediction model; extract features from the training samples using the feature extraction module included in the initial label prediction model to obtain feature data of each training sample; during the training of the initial label prediction model using the feature data of each training sample, perform label prediction processing on the feature data of each training sample using the initial label prediction model to obtain the predicted label of each training sample; determine loss information based on the predicted label and reference label of each training sample; adjust the model parameters of the initial label prediction model based on the loss information; and determine the first label prediction model after initial training based on the initial label prediction model after the model parameter adjustment.

[0128] In one embodiment, the processor 1101 is specifically used to: perform uncertainty analysis processing on the test sample using the trained uncertainty estimation model to obtain the data uncertainty prediction value and the model uncertainty prediction value corresponding to the test sample; and determine the uncertainty target value of the test sample based on the data uncertainty prediction value and the model uncertainty prediction value.

[0129] In one embodiment, the processor 1101 is specifically configured to: determine the mean uncertainty based on the first uncertainty target value of each test sample; and determine the target test sample from the test sample set based on the mean uncertainty, the first uncertainty target value and the second uncertainty target value of each test sample, and the pseudo label of each test sample.

[0130] In one embodiment, the processor 1101 is specifically configured to: sort each test sample in descending order of the first uncertainty target value to obtain sorted test samples; select test samples in the sorted test samples whose second uncertainty target value is greater than or equal to the uncertainty mean as candidate test samples; and determine the candidate test samples in the candidate test samples whose difference between the second uncertainty target value and the corresponding pseudo label is less than the difference threshold as target test samples.

[0131] In one embodiment, the processor 1101 is specifically configured to: normalize the uncertainty target value of each target test sample to obtain a first weight coefficient for each target test sample; retrain the first label prediction model using each target test sample, the first weight coefficient of each target test sample, the training samples included in the training sample set, and the second weight coefficient of each training sample; and determine the data processing model based on the second label prediction model obtained from the retraining; wherein the first weight coefficient is less than the second weight coefficient.

[0132] It should be understood that the computer device 110 described in the embodiments of this application can perform the foregoing... Figure 2 The description of the data processing method in the corresponding embodiment can also be performed as described above. Figure 10 The description of the data processing apparatus 1000 in the corresponding embodiments will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated here.

[0133] Furthermore, it should be noted that embodiments of this application also provide a storage medium storing a computer program for the aforementioned data processing method. This computer program includes program instructions. When one or more processors load and execute these program instructions, the data processing method described in one embodiment can be implemented, as will not be repeated here. The beneficial effects of using the same method will also not be repeated here. It is understood that the program instructions can be deployed on one or more computer devices capable of communicating with each other for execution.

[0134] This application also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps described in the embodiments of the above methods.

[0135] Finally, it should be noted that the terminology in the specification, claims, and accompanying drawings of this application, such as relational terms like "first" and "second," is used merely to distinguish one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0136] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A data processing method, characterized by, The method comprises: obtaining a training sample set and a test sample set, the training sample set comprising a plurality of training samples and reference labels of the training samples, and the test sample set comprising a plurality of test samples; training an initial label prediction model by using the training sample set to obtain a first label prediction model after initial training, and training an uncertainty estimation model by using the training sample set to obtain a trained uncertainty estimation model; inputting the test samples in the test sample set into the first label prediction model for label prediction processing to obtain pseudo labels of the test samples; performing uncertainty analysis processing on the test samples by using the trained uncertainty estimation model to obtain data uncertainty prediction values and model uncertainty prediction values corresponding to the test samples, and determining uncertainty target values of the test samples according to the data uncertainty prediction values and the model uncertainty prediction values; wherein the data uncertainty prediction values are used to indicate uncertainty of the pseudo labels of the test samples caused by data noise of the training samples, and the model uncertainty prediction values are used to indicate uncertainty of the pseudo labels of the test samples caused by incomplete model training; determining target test samples from the test sample set according to the uncertainty target values and the pseudo labels of the test samples, updating the training sample set by using the target test samples and the pseudo labels of the target test samples, retraining the first label prediction model by using the updated training sample set, and determining the data processing model according to a second label prediction model obtained through retraining. obtaining a to-be-processed data object and inputting the to-be-processed data object into the data processing model for processing to obtain a target processing result.

2. The method of claim 1, wherein, The initial label prediction model comprises a feature extraction module, and the training of the initial label prediction model by using the training sample set to obtain the first label prediction model after initial training comprises: inputting the training samples in the training sample set into the initial label prediction model, performing feature extraction on the training samples by using the feature extraction module in the initial label prediction model to obtain feature data of the training samples; in the process of training the initial label prediction model by using the feature data of the training samples, performing label prediction processing on the feature data of the training samples by using the initial label prediction model to obtain prediction labels of the training samples; determining loss information according to the prediction labels and the reference labels of the training samples, adjusting model parameters of the initial label prediction model according to the loss information, and determining the first label prediction model after initial training according to the initial label prediction model after the adjustment of the model parameters.

3. The method of claim 1 or 2, wherein, The uncertainty target values of the test samples comprise first uncertainty target values and second uncertainty target values, and the determination of the target test samples from the test sample set according to the uncertainty target values and the pseudo labels of the test samples comprises: determining an uncertainty mean value according to the first uncertainty target values of the respective test samples; determining a target test sample from the test sample set according to the uncertainty mean value, the first uncertainty target values and the second uncertainty target values of the respective test samples, and the pseudo labels of the respective test samples.

4. The method of claim 3, wherein, The determining a target test sample from the test sample set according to the uncertainty mean value, the first uncertainty target values and the second uncertainty target values of the respective test samples, and the pseudo labels of the respective test samples, comprises: performing sorting processing on the respective test samples in a descending order of the first uncertainty target values, to obtain sorted respective test samples; taking a test sample with a second uncertainty target value greater than or equal to the uncertainty mean value as a candidate test sample from the sorted respective test samples; taking a candidate test sample with a difference between the second uncertainty target value and the corresponding pseudo label less than a difference threshold as a target test sample.

5. The method of claim 1, wherein, The re-training the first label prediction model by using the updated training sample set, and determining the data processing model according to a second label prediction model obtained by the re-training, comprises: performing normalization processing on the uncertainty target values of the respective target test samples, to obtain first weight coefficients of the respective target test samples; re-training the first label prediction model by using the respective target test samples, the first weight coefficients of the respective target test samples, training samples included in the training sample set, and second weight coefficients of the respective training samples, and determining the data processing model according to a second label prediction model obtained by the re-training; wherein the first weight coefficients are less than the second weight coefficients.

6. A data processing apparatus, characterized by, comprises: The obtaining module is further configured to obtain a training sample set and a test sample set, the training sample set including a plurality of training samples and reference labels of the respective training samples, and the test sample set including a plurality of test samples; The training module is configured to train an initial label prediction model by using the training sample set, to obtain a first label prediction model after initial training, and to train an uncertainty estimation model by using the training sample set, to obtain a trained uncertainty estimation model; The processing module is further configured to input test samples included in the test sample set into the first label prediction model for label prediction processing, to obtain pseudo labels of the respective test samples; The processing module is further configured to perform uncertainty analysis processing on the test samples by using the trained uncertainty estimation model, to obtain data uncertainty prediction values and model uncertainty prediction values corresponding to the test samples, and to determine uncertainty target values of the test samples according to the data uncertainty prediction values and the model uncertainty prediction values; wherein the data uncertainty prediction values are used to indicate uncertainty of the pseudo labels of the test samples caused by data noise of the training samples, and the model uncertainty prediction values are used to indicate uncertainty of the pseudo labels of the test samples caused by incomplete model training. The determining module is configured to determine target test samples from the test sample set according to the uncertainty target values and the pseudo labels of the test samples, update the training sample set by using the target test samples and the pseudo labels of the target test samples, retrain the first label prediction model by using the updated training sample set, and determine the data processing model according to a second label prediction model obtained through the retraining. The acquisition module is further configured to acquire the to-be-processed data object. The processing module is further configured to input the to-be-processed data object into the data processing model for processing to obtain a target processing result.

7. A computer device, comprising: Comprise: A processor, a memory and a network interface; The processor is connected with the memory and the network interface, wherein the network interface is configured to provide network communication function, the memory is configured to store program code, and the processor is configured to call the program code to execute the data processing method in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program comprises program instructions. When the program instructions are executed by the processor, the data processing method in any one of claims 1-5 is executed.

Images