A data augmentation method and related apparatus

Abstract

This application discloses a data augmentation method and related apparatus, relating to the field of computer technology. The method includes: acquiring a sample set to be processed; determining dirty data labels and sparse data labels for each sample in the sample set based on preset sample quality assessment conditions; selecting samples that are sparse data and not dirty data from the sample set as augmentation seed samples based on the dirty data labels and sparse data labels; and generating synthetic samples corresponding to the augmentation seed samples based on the category labels and core features contained in the augmentation seed samples. This application, based on the dual screening of sparse and dirty data and the controlled generation of core features, achieves accurate screening of augmentation seed samples and high-quality generation of synthetic samples, effectively suppressing model performance degradation.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a data augmentation method and related apparatus. Background Technology

[0002] With the rapid development of artificial intelligence technology, data is widely recognized as a core element driving model performance improvement. Data annotation, as a crucial step in obtaining high-quality training data, directly determines the success or failure of the entire artificial intelligence project in terms of efficiency and quality. Currently, numerous data annotation platforms (such as Labelbox and Scale AI) have emerged in the industry, offering a wealth of annotation tools and project management functions to assist in large-scale data annotation.

[0003] When existing labeled data is insufficient to support model training, the industry typically uses data augmentation techniques to expand the training dataset. This includes randomly selecting samples and randomly replacing or deleting words from the selected text to obtain augmented synthetic data. However, this method of randomly selecting samples and randomly replacing or deleting words lacks precise guidance for specific samples and is prone to generating low-quality synthetic data that does not match the distribution of the actual labeled data, leading to model performance degradation. Summary of the Invention

[0004] In view of the above problems, this application provides a data augmentation method and related apparatus, which can accurately select augmentation seed samples with both augmentation value and label reliability from labeled samples, and generate synthetic samples accordingly. This avoids contamination of synthetic samples due to the use of dirty data and reduces the risk of model performance degradation. The specific solution is as follows:

[0005] The first aspect of this application provides a data augmentation method, comprising:

[0006] Obtain a sample set to be processed. Each sample in the sample set to be processed includes: labeled data, category labels corresponding to the labeled data, and core features. The core features are the discriminative basis in the labeled data used to support the labeling decision of the category labels.

[0007] Based on preset sample quality assessment conditions, dirty data labels and sparse data labels are determined for each sample in the sample set to be processed. The dirty data labels are used to characterize whether the corresponding sample is dirty data with incorrect class label, and the sparse data labels are used to characterize whether the target model to be trained is uncertain about the prediction classification result of the corresponding sample.

[0008] Based on the dirty data label and the sparse data label, samples that are sparse data and not dirty data are selected from the sample set to be processed and used as enhancement seed samples.

[0009] Based on the category labels and core features contained in the enhanced seed sample, a synthetic sample corresponding to the enhanced seed sample is generated.

[0010] In one possible implementation, obtaining the sample set to be processed includes:

[0011] Obtain an initial dataset, which consists of multiple labeled data points;

[0012] The preset user annotation interface sequentially displays each labeled data in the initial dataset, and performs the following operations for each displayed labeled data:

[0013] Lock the submit control;

[0014] In response to the core feature region defined by the user in the labeled data, obtain the core features corresponding to the labeled data;

[0015] In response to a user's category label classification operation on the labeled data, obtain the category label corresponding to the labeled data;

[0016] After obtaining the core features and category labels corresponding to the labeled data, unlock the submission control;

[0017] In response to the user's triggering operation on the submission control, the labeled data, the core features and category labels corresponding to the labeled data are treated as a sample and stored in the sample set to be processed.

[0018] One possible implementation also includes:

[0019] Select one initial labeled data from the initial dataset, store it in a pre-built data retention set, and iterate through each labeled data in the initial dataset except for the initial labeled data:

[0020] Calculate the similarity information between the currently traversed labeled data and the labeled data in the data retention set, wherein the similarity information includes at least one similarity value;

[0021] If all similarity values ​​in the similarity information are not greater than the preset similarity threshold, then the currently traversed labeled data is stored in the data retention set, and the next labeled data is traversed; otherwise, the next labeled data is traversed directly until the traversal ends.

[0022] Use the data set obtained at the end of the traversal as the updated initial dataset;

[0023] The step of sequentially displaying each labeled data item in the initial dataset on a preset user annotation interface includes:

[0024] The user annotation interface sequentially displays each annotation data in the updated initial dataset.

[0025] In one possible implementation, before traversing each labeled data in the initial dataset other than the initial labeled data, the method further includes:

[0026] A feature vector is generated for each labeled data in the initial dataset using the first pre-trained model, so as to obtain the feature vector of each labeled data in the initial dataset;

[0027] The calculation of similarity information between the currently traversed labeled data and the labeled data in the data retention set includes:

[0028] Construct an approximate nearest neighbor index of the feature vectors of all labeled data in the data retention set;

[0029] Based on the approximate nearest neighbor index, retrieve the k candidate labeled data that are most similar to the currently traversed labeled data from the data retention set, where k is a preset positive integer;

[0030] Calculate the similarity between the feature vector of the currently traversed labeled data and the feature vectors of the k candidate labeled data respectively, and obtain k similarity values ​​as the similarity information.

[0031] In one possible implementation, determining dirty data labels and sparse data labels for each sample in the sample set to be processed, based on preset sample quality assessment conditions, includes:

[0032] The pre-trained classification model is used to classify and predict each labeled data in the sample set to be processed, and the prediction probability vector corresponding to each labeled data in the sample set to be processed is obtained. The prediction probability vector includes the prediction probability of all categories in the preset category set.

[0033] The predicted probability vector and category label corresponding to each labeled data in the sample set to be processed are input into the confidence learning module to obtain the label quality score corresponding to each labeled data in the sample set to be processed.

[0034] Based on the label quality score corresponding to each labeled data in the sample set to be processed, determine the dirty data label of the sample in which each labeled data in the sample set to be processed is located;

[0035] The maximum predicted probability value of the predicted probability vector corresponding to each labeled data in the sample set to be processed is determined as the category prediction confidence of each labeled data in the sample set to be processed.

[0036] Based on the category prediction confidence of each labeled data in the sample set to be processed, the sparse data label of the sample containing each labeled data in the sample set to be processed is determined.

[0037] In one possible implementation, determining the dirty data label of the sample containing each labeled data in the sample set to be processed, based on the label quality score corresponding to each labeled data in the sample set to be processed, includes:

[0038] For each piece of labeled data in the sample set to be processed, if the label quality score corresponding to the labeled data is less than a preset score threshold, then the dirty data label of the sample containing the labeled data is determined to be true; otherwise, the dirty data label of the sample containing the labeled data is determined to be false.

[0039] Wherein, the dirty data label being true indicates that the sample is dirty data, and the dirty data label being false indicates that the sample is not dirty data.

[0040] In one possible implementation, determining the sparse data label of the sample containing each labeled data in the sample set to be processed based on the category prediction confidence corresponding to each labeled data in the sample set to be processed includes:

[0041] For each labeled data in the sample set to be processed, if the confidence of the category prediction corresponding to the labeled data is less than the preset confidence threshold, then the sparse data label of the sample containing the labeled data is determined to be true; otherwise, the sparse data label of the sample containing the labeled data is determined to be false.

[0042] Wherein, the sample corresponding to the sparse data label being a true representation is sparse data, and the sample corresponding to the sparse data label being a false representation is not sparse data.

[0043] In one possible implementation, generating a synthetic sample corresponding to the enhanced seed sample based on the category labels and core features contained in the enhanced seed sample includes:

[0044] Based on the category labels and core features contained in the enhanced seed samples, prompt instructions are generated. These prompt instructions are used to instruct the generative large language model to generate semantically similar but different sentence structures and expressions based on the category labels and core features.

[0045] The prompting instructions are input into the generative large language model to obtain the augmented data output by the model;

[0046] The augmented data and the category labels and core features contained in the augmented seed sample are combined to form the synthetic sample.

[0047] One possible implementation also includes:

[0048] The dirty data samples in the sample set to be processed are pushed to the manual review interface. After receiving the reviewer's correction result on the category label of the dirty data sample, the category label of the dirty data sample is updated according to the correction result. The dirty data sample refers to the sample whose dirty data label is characterized as dirty data.

[0049] And / or, store at least one of the following data into the training dataset: samples that are neither dirty nor sparse data in the sample set to be processed, the synthetic samples, and dirty data samples whose category labels have been updated according to the correction results.

[0050] A second aspect of this application provides a data enhancement apparatus, comprising:

[0051] The data acquisition unit is used to acquire a sample set to be processed. Each sample in the sample set to be processed includes: labeled data, category labels corresponding to the labeled data, and core features. The core features are the discriminative basis in the labeled data used to support the labeling decision of the category labels.

[0052] The label determination unit is used to determine dirty data labels and sparse data labels for each sample in the sample set to be processed based on preset sample quality assessment conditions. The dirty data labels are used to characterize whether the corresponding sample is dirty data with an incorrect class label, and the sparse data labels are used to characterize whether the target model to be trained is uncertain about the prediction classification result of the corresponding sample.

[0053] The sample screening unit is used to select samples that are sparse data and not dirty data from the sample set to be processed based on the dirty data label and the sparse data label, as enhancement seed samples.

[0054] The data augmentation unit is used to generate a synthetic sample corresponding to the augmented seed sample based on the category labels and core features contained in the augmented seed sample.

[0055] A third aspect of this application provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the data enhancement method described in the first aspect or any implementation thereof.

[0056] A fourth aspect of this application provides an electronic device, including at least one processor and a memory connected to the processor, wherein:

[0057] The memory is used to store computer programs;

[0058] The processor is used to execute the computer program to enable the electronic device to implement the data enhancement method of the first aspect or any implementation thereof.

[0059] The fifth aspect of this application provides a computer storage medium carrying one or more computer programs, which, when executed by an electronic device, enable the electronic device to implement the data enhancement method described in the first aspect or any implementation thereof.

[0060] Using the above technical solution, the data augmentation method provided in this application first obtains a sample set to be processed. In addition to labeled data and category labels, each sample in the sample set contains an additional core feature. This core feature is the discriminative basis used in the labeled data to support the labeling decision of the category label. By recording this core feature, the key evidence on which the labeling decision is based during the labeling process can be preserved, providing traceable constraint information for subsequent data augmentation.

[0061] To select truly valuable high-quality seed samples for enhancement from the sample set to be processed, this application further assigns dirty data labels and sparse data labels to each sample in the sample set based on preset sample quality assessment conditions. Dirty data labels reflect whether the corresponding sample's category label is incorrect, while sparse data labels reflect whether the target model being trained is sufficiently confident in its classification prediction results for the corresponding sample. Since the model's classification prediction results for the sample are uncertain because the model lacks sufficient training material for similar samples, and when the sample category label is incorrect, the synthetic samples generated with the incorrect category label will amplify the error, causing the synthetic samples to deviate further and further from the true distribution. Therefore, to obtain high-quality seed samples that have both enhancement value and label reliability, this application can select samples that are sparse data and not dirty data from the sample set to be processed, based on dirty data labels and sparse data labels, as enhancement seed samples. Finally, synthetic samples are generated based on the category labels and core features in the enhanced seed samples. Since the category labels can provide semantic boundaries and the core features can provide content constraints, synthetic samples that do not deviate from the true distribution can be generated under the dual constraints of category labels and core features. The synthesis quality is more reliable and does not introduce additional noise, effectively suppressing the degradation of model performance. Attached Figure Description

[0062] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0063] Figure 1 A schematic diagram of a system architecture provided for this application;

[0064] Figure 2A flowchart illustrating a data augmentation method provided in this application;

[0065] Figure 3 A schematic diagram of the structure of a data enhancement device provided in this application;

[0066] Figure 4 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0067] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0068] As will be known to those skilled in the art, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0069] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0070] This application provides a data augmentation method and related apparatus that can be applied to scenarios that require labeled training data to train models, such as text classification, sentiment analysis, and named entity recognition in natural language processing, image classification, object detection, and semantic segmentation in computer vision, as well as the data preparation stage for artificial intelligence tasks such as speech recognition and multimodal understanding.

[0071] It should be noted that the above scenarios are merely examples and are not intended to limit this application.

[0072] Optionally, the data augmentation method provided in this application can be applied to, for example... Figure 1 The system architecture shown includes a terminal 100 and a server 200. The server 200 may include one or more servers (…). Figure 1 (This example uses a server as an illustration).

[0073] Either terminal 100 or server 200 can be used independently to execute the data enhancement method provided in the embodiments of this application. Alternatively, terminal 100 and server 200 can also be used collaboratively to execute the data enhancement method provided in the embodiments of this application.

[0074] The following description Figure 1 The product form of the mid-terminal 100;

[0075] The terminal 100 in this application embodiment can be a mobile phone, tablet computer, wearable device, vehicle device, augmented reality (AR) / virtual reality (VR) device, laptop computer, ultra-mobile personal computer (UMPC), netbook, personal digital assistant (PDA), etc., and this application embodiment does not impose any restrictions on it.

[0076] To enable those skilled in the art to better understand this application, the data augmentation method of the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0077] Reference Figure 2 , Figure 2 This is a flowchart illustrating a data augmentation method provided in an embodiment of this application, as shown below. Figure 2 As shown, the data augmentation method may include:

[0078] Step S201: Obtain the sample set to be processed.

[0079] The sample set to be processed includes multiple samples. Each sample includes a piece of labeled data and its corresponding category label. Here, labeled data refers to the raw data being labeled, such as a piece of text, an image, or an audio clip; the category label is the classification result assigned to the labeled data from a predefined set of categories. For example, for a text classification task, the category label could be "artificial intelligence," "blockchain," etc.; for an image classification task, the category label could be "animal," "plant," "vehicle," "pedestrian," etc.; for an audio classification task, the category label could be "abnormal" or "normal." Of course, the above category labels are merely examples and are not intended to limit this application.

[0080] Understandably, there are certain criteria for making labeling decisions on any piece of labeled data. For example, a text classified as "artificial intelligence" might be categorized as such because it contains the key phrase "neural network," while an image classified as "pedestrian" might be categorized as such because it contains a face. To fully preserve these criteria and provide strong data support for subsequent data augmentation and synthesis processes, each sample in this embodiment also includes core features corresponding to the labeled data. Here, core features are the discriminative criteria in the labeled data used to support labeling decisions on category labels, i.e., the key evidence upon which the labeler bases their labeling decisions.

[0081] In a preferred embodiment, each sample in the sample set to be processed includes not only the content of the core feature itself, but also the position coordinates of that content in the labeled data, such as the start and end indices of text, the coordinates of the bounding box of an image, etc. Thus, by recording the position coordinates of the core feature, a precise location and discrimination basis is provided for the subsequent data augmentation and synthesis process.

[0082] Step S202: Based on the preset sample quality assessment conditions, determine dirty data labels and sparse data labels for each sample in the sample set to be processed.

[0083] Here, the dirty data label is used to characterize whether the corresponding sample is dirty data with an incorrect class label. If the dirty data label of a sample is true, it indicates that the sample is dirty data with an incorrect class label (also called a dirty data sample), which means that the annotator may have selected the wrong class label for the sample, that is, the class label of the sample has a potential error and is unreliable; conversely, if the dirty data label of a sample is false, it indicates that the sample is non-dirty data with a correct class label (also called a clean data sample), which means that the annotator has selected the correct class label for the sample.

[0084] Sparse data labels are used to characterize whether the target model being trained is uncertain about the prediction classification result of the corresponding sample. If the sparse data label of a sample is true, it indicates that the sample is sparse data whose category prediction confidence of the target model does not meet the preset requirements. This means that the target model is uncertain about the classification result of the sample, which also means that the target model lacks sufficient training material for the feature space region corresponding to the sample. This results in a knowledge blind spot in the target model's learning of the sample, requiring retraining to improve the performance of the target model.

[0085] The aforementioned preset sample quality assessment conditions include various specific algorithms, such as confidence learning and category prediction confidence threshold judgment, etc. Their specific forms are related to the process of determining dirty data labels and sparse data labels, and this application does not impose specific limitations.

[0086] Step S203: Based on the dirty data label and sparse data label, select samples that are sparse data and not dirty data from the sample set to be processed, and use them as seed samples for enhancement.

[0087] As mentioned earlier, each sample in the sample set to be processed has a dirty data label and a sparse data label. Therefore, this embodiment can select samples with true sparse data labels but false dirty data labels from the sample set to be processed, and use these samples as seed samples for data augmentation as follows.

[0088] As mentioned earlier, sparse data consists of samples whose class prediction confidence is very low (i.e., the model is hesitant about the classification result of the sample). The model lacks sufficient training material in the spatial feature regions corresponding to these samples. Therefore, using sparse data as one of the screening conditions to obtain enhanced seed samples and performing targeted data synthesis can enable the model to obtain more training samples in the corresponding spatial feature regions, thereby enabling targeted training of the model and improving model performance.

[0089] Considering that if the class label of the augmented seed sample is incorrect, the generated synthetic sample will inherit and amplify this error, resulting in model performance degradation due to the large amount of noise mixed into the synthetic sample. Therefore, to avoid generating synthetic samples with incorrect classes, this embodiment uses non-dirty data (i.e., clean data) as another screening criterion, thereby improving the reliability of the class label of the augmented seed sample.

[0090] Therefore, based on the above two screening conditions, this embodiment can accurately locate high-quality enhanced seed samples that have both enhancement value and label reliability.

[0091] Step S204: Based on the category labels and core features contained in the enhanced seed sample, generate the synthetic sample corresponding to the enhanced seed sample.

[0092] Specifically, this embodiment utilizes the category labels in the enhanced seed samples to provide semantic boundaries and the core features to provide content constraints, generating new synthetic data that is semantically similar to the labeled data in the enhanced seed samples. This results in a synthetic sample containing the synthetic data, its corresponding category labels, and core features. This synthetic sample shares the same category labels and core feature constraints as the enhanced seed samples, effectively expanding the model's training samples in sparse regions. Through training, this effectively improves the model's classification performance.

[0093] In summary, the data augmentation method provided in this application first obtains a sample set to be processed. In addition to labeled data and category labels, each sample in the sample set contains an additional core feature. This core feature is the discriminative basis used in the labeled data to support the labeling decision of the category label. By recording this core feature, the key evidence on which the labeling decision is based during the labeling process can be preserved, providing traceable constraint information for subsequent data augmentation.

[0094] To select truly valuable high-quality seed samples for enhancement from the sample set to be processed, this application further assigns dirty data labels and sparse data labels to each sample in the sample set based on preset sample quality assessment conditions. Dirty data labels reflect whether the corresponding sample's category label is incorrect, while sparse data labels reflect whether the target model being trained is sufficiently confident in its classification prediction results for the corresponding sample. Since the model's classification prediction results for the sample are uncertain because the model lacks sufficient training material for similar samples, and when the sample category label is incorrect, the synthetic samples generated with the incorrect category label will amplify the error, causing the synthetic samples to deviate further and further from the true distribution. Therefore, to obtain high-quality seed samples that have both enhancement value and label reliability, this application can select samples that are sparse data and not dirty data from the sample set to be processed, based on dirty data labels and sparse data labels, as enhancement seed samples. Finally, synthetic samples are generated based on the category labels and core features in the enhanced seed samples. Since the category labels can provide semantic boundaries and the core features can provide content constraints, synthetic samples that do not deviate from the true distribution can be generated under the dual constraints of category labels and core features. The synthesis quality is more reliable and does not introduce additional noise, effectively suppressing the degradation of model performance.

[0095] In some embodiments of this application, the process of step S201 "obtaining the sample set to be processed" described above is introduced.

[0096] This embodiment provides a method for generating a sample set to be processed based on forced interactive annotation, so as to obtain samples containing labeled data, category labels and core feature triples.

[0097] Specifically, the initial dataset is first obtained, which consists of multiple unlabeled or relabeled data points. Optionally, the initial dataset can be stored in a pool of tasks to be labeled, and retrieved from the pool when needed.

[0098] To enable manual annotation, this application pre-builds a user annotation interface, which is a graphical user interface (GUI) for annotators to perform annotation operations. The interface is equipped with a submission control, such as a submit button, to submit the annotated data to the backend storage.

[0099] Therefore, each labeled data in the initial dataset can be displayed sequentially on the user annotation interface for users to annotate. To ensure that manual annotation is subject to effective procedural constraints and to avoid random errors, this embodiment can lock the submission control for each displayed labeled data, and unlock it after the user annotates the category label and core features.

[0100] In other words, once a piece of labeled data is displayed on the user's labeling interface, the system automatically locks the submission controls on the interface, such as making the submit button grayed out and unclickable, and forces the labeler to complete the following two necessary labeling operations before submitting.

[0101] The first type of annotation operation is core feature annotation. Users can define a region on the annotated data (i.e., the core feature is part of the annotated data) using an input device (such as a mouse or touchscreen). For example, for text data, the definition can be made by highlighting a continuous sequence of characters; for image data, the definition can be made by drawing a rectangular bounding box. Thus, in response to the core feature region defined by the user in the annotated data, the core features corresponding to that annotated data can be obtained. In addition, the position coordinates of the region where the core feature is located can be recorded, such as the start and end indices of the text, the coordinates of the image bounding box, etc.

[0102] The second type of annotation operation is category label annotation. In this embodiment, a preset set of categories can be provided in the category label annotation area of ​​the user annotation interface and displayed to the annotator in the form of a list, so that the annotator can select category labels. Thus, in response to the user's category label classification operation on the annotation data, the category labels corresponding to the annotation data can be obtained.

[0103] After both types of annotation operations are completed, that is, after obtaining the core features and category labels corresponding to the annotation data, the submission control is unlocked, such as the submission button changing from gray to green. Then, in response to the user's triggering operation of the submission control, the annotation data, the core features and category labels corresponding to the annotation data are treated as a sample and stored in the sample set to be processed.

[0104] After that, the user annotation interface automatically loads the next annotation data and repeats the above process until all annotation data in the initial dataset has been processed.

[0105] This embodiment achieves proactive process assurance of annotation quality by constructing the sample set to be processed through mandatory interactive annotation. By locking the submission control, annotators are required to first define the core feature regions and then select category labels before unlocking and submitting. This ensures, from the interactive process perspective, that annotators have given sufficient attention and confirmed key evidence to the annotated data before submission, significantly reducing random annotation errors caused by negligence or fatigue. This shifts quality control from post-event correction to in-process prevention. Simultaneously, this embodiment transforms implicit human judgment knowledge into structured information that can be processed by machines, providing information input for subsequent confidence learning verification and generative augmentation based on core features, thus enabling in-depth mining of data value.

[0106] In some other embodiments of this application, considering that there may be duplicate labeled data in the initial dataset, in order to avoid the duplicate labeled data interfering with the training process of the target model, this embodiment can first use the sequential anchor filtering method or the graph connected component clustering method to preprocess the initial dataset for data deduplication before labeling the category labels and core features of the initial dataset in the previous embodiments.

[0107] Taking the sequential anchor filtering method as an example, firstly, one initial labeled data is selected from the initial dataset and stored in a pre-constructed data retention set, which is initially empty. Optionally, the process of selecting initial labeled data from the initial dataset can be random selection, sequential selection, or targeted selection. For example, the data quality of each labeled data in the initial dataset can be roughly determined by a large language model, and then the labeled data with the highest data quality can be selected as the initial labeled data. Of course, there are other selection methods, which will not be described in detail in this embodiment.

[0108] Next, iterate through each labeled data in the initial dataset except for the initial labeled data, and perform the following deduplication check.

[0109] Calculate the similarity information between the currently traversed labeled data and the labeled data in the data retention set, and the similarity information includes at least one similarity value.

[0110] In one implementation, an exact traversal approach can be used, calculating the semantic similarity between the currently traversed labeled data and each labeled data in the data retention set to obtain similarity information. Optionally, the semantic similarity can be calculated using bidirectional cosine similarity, Euclidean distance, or other vector retrieval algorithms.

[0111] Considering that model training requires a large amount of training data, the data set may contain a large amount of labeled data. Calculating semantic similarity one by one would be time-consuming. To speed up similarity retrieval and reduce memory and computing resource consumption, a preferred implementation can use an approximate nearest neighbor (ANN) method, such as Faiss (Facebook AI Similarity Search, a vector-oriented retrieval algorithm library).

[0112] In order to accelerate the retrieval, this embodiment can generate an N-dimensional feature vector for each labeled data in the initial dataset before traversing each labeled data except the initial labeled data in the initial dataset, so as to obtain the feature vector of each labeled data in the initial dataset, that is, to obtain the feature vector of each labeled data in the data retention set.

[0113] Here, the first pre-trained model refers to a model that can convert labeled data into vectors, such as the Enhanced Representation through kNowledge IntEgration (ERNIE) model for text, or the Vision Transformer (ViT) model for images.

[0114] The process of "calculating the similarity information between the currently traversed labeled data and the labeled data in the data reserve" can be summarized as follows: First, construct an approximate nearest neighbor index for the feature vectors of all labeled data in the data reserve. Specifically, use the approximate nearest neighbor algorithm to build an approximate nearest neighbor index for the feature vectors of the labeled data in the data reserve. This approximate nearest neighbor index can divide the vector space into multiple regions or construct a graph structure, thereby achieving sublinear time complexity retrieval. Next, based on the approximate nearest neighbor index, retrieve the k candidate labeled data that are most similar to the currently traversed labeled data from the data reserve, where k is a preset positive integer. Then, calculate the similarity between the feature vector of the currently traversed labeled data and the feature vectors of the k candidate labeled data, obtaining k similarity values ​​as similarity information. Since ANN retrieval does not need to compare with all feature vectors in the data reserve, but only retrieves the k most similar candidate labeled data by region, the retrieval time is shorter and the efficiency is higher.

[0115] Understandably, if the similarity between the currently traversed labeled data and any previously retained labeled data in the data retention set is greater than a preset similarity threshold T1, it indicates that the currently traversed labeled data semantically overlaps with the labeled data in the data retention set, and therefore the currently traversed labeled data needs to be filtered out. Only when the similarity between the currently traversed labeled data and all labeled data in the data retention set is less than or equal to the preset similarity threshold T1, is it added to the data retention set as a new anchor point. Based on this, this embodiment can store the currently traversed labeled data into the data retention set and traverse the next labeled data only when all similarity values ​​in the above similarity information are not greater than the preset similarity threshold T1; otherwise, it directly traverses the next labeled data until the traversal ends. Finally, the data retention set obtained at the end of the traversal is used as the updated initial dataset.

[0116] For example, the initial dataset includes three labeled data points: a, b, and c. First, an empty data reserve is constructed, and a is selected as the initial labeled data and stored in the data reserve. Then, b is iterated over, and the similarity between the feature vectors of b and a in the data reserve is calculated. It is found that the similarity between the feature vectors of b and a is greater than T1, so b is marked as redundant and discarded. Next, c is iterated over, and the similarity between the feature vectors of c and a in the data reserve is calculated. It is found that the similarity between the feature vectors of c and a is less than T1, so c is stored in the data reserve. Finally, a and c form the updated initial dataset, and b is safely filtered out.

[0117] It is understandable that the initial labeled data may be unreasonable, resulting in most of the labeled data in the initial dataset being filtered out. In order to avoid insufficient training of the target model due to the small number of labeled data in the updated data retention set, this embodiment can optionally calculate the first number of labeled data in the updated data retention set and the second number of labeled data in the data retention set before the update after each updated data retention set is obtained. If the ratio of the first number to the second number is less than a preset ratio threshold, the initial labeled data is re-selected, and then the above process is repeated.

[0118] Correspondingly, the process of "sequentially displaying each labeled data in the initial dataset in the preset user annotation interface" mentioned earlier is as follows: sequentially displaying each labeled data in the updated initial dataset in the user annotation interface. The subsequent annotation process can be referred to the previous description and will not be repeated here.

[0119] This embodiment, through vectorized semantic similarity calculation and sequential filtering mechanism, can efficiently identify and filter semantically repetitive labeled data, ensuring that the work of annotators focuses on unique samples with the highest information entropy, avoiding ineffective human investment in low-value data, significantly reducing labeling costs and improving labeling efficiency.

[0120] In some other embodiments of this application, the process of step S202, "determining dirty data labels and sparse data labels for each sample in the sample set to be processed based on preset sample quality assessment conditions," is described.

[0121] First, a pre-trained classification model (second pre-trained model) is used to classify and predict each labeled data in the sample set to be processed, so as to obtain the prediction probability vector corresponding to each labeled data.

[0122] Optionally, the pre-trained classification model can be an existing deep learning model such as BERT (Bidirectional Encoder Representations from Transformers) or Residual Network (ResNet). This model performs forward inference on the input labeled data and outputs a predicted probability vector. The dimension of this predicted probability vector is equal to the total number of categories in the preset category set. Each component represents the predicted probability that the labeled data belongs to the corresponding category, and the sum of the predicted probabilities of all categories is 1.

[0123] Next, the predicted probability vector and class label (i.e., the true label) corresponding to each labeled data in the sample set to be processed are input into the confidence learning module to obtain the label quality score corresponding to the labeled data.

[0124] Optionally, this embodiment can use Cleanlab or its variants to implement the confidence learning module. The basic processing flow is as follows: A pre-trained classification model is used to predict the probability of each labeled data point in the sample set to be processed through cross-validation, obtaining the predicted probability of each labeled data point belonging to each category, forming a prediction probability matrix corresponding to the sample set to be processed; for each category, a confidence threshold is calculated based on the predicted probabilities of all labeled data points under that category; based on the confidence threshold and the prediction probability matrix, the joint distribution matrix between noise labels and true labels is statistically estimated; based on the joint distribution matrix, a label quality score is output for each labeled data point. The lower the label quality score, the greater the probability that the category label of the labeled data point is incorrect.

[0125] Therefore, this embodiment can determine the dirty data label of the sample containing the labeled data based on the label quality score corresponding to each labeled data in the sample set to be processed.

[0126] Specifically, for each labeled data in the sample set to be processed, if its label quality score is less than the preset score threshold T2 (e.g., 0.6), then the dirty data label of the sample containing the labeled data is determined to be true, that is, the sample containing the labeled data is dirty data (the category label is likely to be wrong); otherwise, the dirty data label is determined to be false, that is, the sample containing the labeled data is not dirty data.

[0127] Having obtained the predicted probability vector corresponding to each labeled data in the sample set to be processed as described above, this embodiment can also determine the maximum predicted probability value of the predicted probability vector corresponding to each labeled data as the category prediction confidence of the labeled data, and then determine the sparse data label of the sample where the labeled data is located based on the category prediction confidence of each labeled data.

[0128] Understandably, for a given piece of labeled data, the maximum value in the predicted probability vector output by the classification model represents the model's belief in the most likely category of the labeled data and its level of confidence. For ease of understanding, this embodiment defines this maximum value as the category prediction confidence score. The higher the category prediction confidence score, the more certain the classification model is in classifying the labeled data; the lower the category prediction confidence score, the more hesitant the classification model is in classifying the labeled data, which usually means that the labeled data is near the decision boundary or that the model lacks training material for the corresponding region of the labeled data.

[0129] In one possible implementation, for each piece of labeled data, if its category prediction confidence is less than a preset confidence threshold T3, then the sparse data label of the sample containing the labeled data is determined to be true, that is, the sample containing the labeled data is sparse data; otherwise, the sparse data label is determined to be false, indicating that the sample containing the labeled data is not sparse data.

[0130] Optionally, the preset reliability threshold can be a fixed value or can be dynamically set according to the model's performance on the validation set or business needs.

[0131] This embodiment outputs a quantifiable label quality score for each sample through a confidence learning module, transforming the original manual sampling process of label error detection into an automated quantitative evaluation, significantly improving the efficiency and accuracy of identifying dirty data. Simultaneously, by directly calculating the class prediction confidence using the predicted probability vector output by the classification model, it can quickly locate samples where the model hesitates in its decision-making, i.e., sparse data. Dirty data labels and sparse data labels provide a dual characterization of samples from the perspectives of label reliability and model uncertainty, respectively. The former ensures that the selected enhancement seeds do not carry incorrect labels, while the latter ensures that enhancement operations are precisely applied to the regions where the model most needs supplementary training material. This provides a comprehensive and quantitative decision-making basis for subsequent enhancement seed selection, thereby achieving automated selection of high-quality enhancement seeds without increasing manual costs.

[0132] In some other embodiments of this application, the process of step S204, "generating a synthetic sample corresponding to the enhanced seed sample based on the category label and core features contained in the enhanced seed sample," is described.

[0133] To fully leverage the generative advantages of current generative large language models, this embodiment provides a method for data synthesis based on generative large language models. First, it is necessary to generate prompt instructions based on the category labels and core features contained in the enhanced seed samples. These prompt instructions are used to instruct the generative large language model to generate semantically similar but different sentence structures and expressions based on the category labels and core features.

[0134] In this embodiment, the prompting instructions include: task instructions, category labels included in the enhanced seed samples, and core features. The task instructions tell the generative large language model the action to be performed, for example, "Give three new texts that are semantically similar but have different sentence structures and expressions, based on the core features and category labels provided below." The category labels provide semantic boundaries, limiting the generated content to conform to the context of that category label, for example, "The generated text must conform to the semantic category of 'artificial intelligence'." The core features provide strong constraints at the content level, for example, "The generated text must forcibly contain and embody the core concept of 'deep neural network'."

[0135] In one possible implementation, the prompt instruction can be constructed using a template-based approach, combining the aforementioned elements according to a fixed format. For example, a complete prompt instruction might look like this:

[0136] "

System Task Commands

[0137] You are a professional data generation assistant. Based on the labeled data and category labels provided in the input data below, please generate three new texts that are semantically highly similar but differ in sentence structure and expression.

[0138] Generate constraints

[0139] The generated new text must strictly conform to the semantic context corresponding to the [category label], and must not deviate from it;

[0140] The generated new text must include and reflect the core intent expressed by the [core features] (the use of adversarial variants such as homophones, word splitting, and special punctuation insertion is allowed and encouraged to simulate a real adversarial environment).

[0141] [Input Data]

[0142] Annotated data: Recent research shows that deep neural networks have made groundbreaking progress in natural language processing tasks;

[0143] Category tag: Artificial Intelligence;

[0144] Core feature: Deep neural network.

[0145] Output Format

[0146] Please directly output the generated new text list, one line per line, without any further explanation.

[0147] In this embodiment, the prompts constructed above can be input into a generative large language model, such as a large language model (LLM), to obtain the augmented data output by the model.

[0148] For example, the generative large language model in the above example might output variant texts such as: "A major breakthrough has recently been achieved in natural language processing models based on deep learning," "Deep neural networks help improve the performance of language understanding," and "The field of natural language processing benefits from new progress made by deep neural networks." Because the prompts contain strong constraints on core features, these augmented data are highly semantically related to the labeled data in the original augmentation seed samples, but they are diverse in sentence structure and expression, thus achieving high-quality data augmentation.

[0149] Finally, this embodiment combines the augmented data and the category labels and core features contained in the augmented seed samples to obtain a synthetic sample. That is, the augmented data output by the generative large language model, the category labels in the augmented seed samples, and the core features are associated to form a complete synthetic sample. This synthetic sample can be directly used for training the target model.

[0150] This embodiment generates synthetic samples by using prompts based on category labels and core features, achieving a fundamental shift from uncontrolled to controlled generation. The core features captured during the annotation stage are incorporated as strong constraints into the prompts, forcing the generative model to semantically expand around these core features. This ensures that the synthetic samples retain the most discriminative evidence from the original seed augmentation samples while possessing necessary sentence diversity. Furthermore, since the augmentation seed itself has undergone dual screening of sparse and non-dirty data, its quality and augmentation value are guaranteed. Simultaneously, the category labels and core features in the prompts restrict the semantic space of the generated data to a credible range, ensuring that the generated synthetic samples not only do not introduce noise but also enhance the target model's discriminative ability in sparse regions, significantly improving the target model's generalization performance.

[0151] In a preferred implementation, in order to make full use of all the obtained labeled data, this embodiment can also push dirty data samples in the sample set to be processed to the manual review interface so that the category labels of the dirty data samples are corrected by the manual reviewer. Thus, after receiving the correction results of the reviewer on the category labels of the dirty data samples, the category labels of the dirty data samples are updated according to the correction results.

[0152] Here, "dirty data sample" refers to a sample whose dirty data label is true, that is, a sample whose dirty data label represents dirty data. Because dirty data samples contain core features, reviewers can quickly determine whether the category label is incorrect and correct it through the core features, avoiding the loss of high-quality difficult sample, and at the same time, it can also evaluate the labeling performance of the labelers.

[0153] Optionally, this embodiment may also store at least one of the following data into the training dataset: samples that are neither dirty nor sparse data in the sample set to be processed, synthetic samples, and dirty data samples after updating the category labels according to the correction results.

[0154] Specifically, samples that are neither dirty nor sparse, as well as dirty data samples whose category labels have been updated based on the correction results, have reliable category labels and high data value, and can be directly used as high-quality training data for storage.

[0155] For synthetic samples, the storage method can be any one of the following or a combination thereof:

[0156] Direct input method: Since the synthetic samples are already subject to the explicit constraints of the category labels and core features in the prompt instructions when they are generated, the synthetic samples directly inherit the category labels and core features of the enhanced seed samples and are directly stored in the training dataset as labeled data for subsequent model training iterations.

[0157] Post-verification data entry method: Synthetic samples are first output to a verification module (this module can be an automated filtering module, such as a rule-based filter or a small discriminant model; it can also be a manual quick verification interface). After passing consistency verification (e.g., confirming that the generated content matches the category label and does not contain obvious noise), they are then stored in the training dataset. Manual quick verification only confirms the usability of the synthetic samples and does not require repeating the core feature delineation operation.

[0158] This embodiment achieves closed-loop iteration and continuous optimization of data quality through closed-loop manual review and multi-strategy data entry. Specifically, dirty data identified by the confidence learning module is not directly discarded, but pushed to reviewers for precise verification and correction. The corrected samples become high-quality training data again, rescuing difficult examples that might have been discarded due to labeling errors and providing more accurate category labels for subsequent model iterations. Simultaneously, the system can assess the professional competence of annotators by statistically analyzing which annotators generate the most dirty data, or identify and correct ambiguities in the labeling rules, thereby driving continuous improvement of the entire labeling process. Furthermore, this embodiment provides a flexible and controllable synthetic sample entry strategy. Direct entry allows for rapid data growth at minimal cost, while post-verification entry further ensures the reliability of synthetic samples through additional filtering or manual verification. Manual verification eliminates the need to repeat core feature definition operations, making it far less costly than labeling from scratch. The resulting training dataset is used for the next round of target model training, significantly improving the model's robustness and generalization ability.

[0159] The above describes a data augmentation method provided by an embodiment of this application. The following describes an apparatus for performing the above data augmentation method.

[0160] Please see Figure 3 , Figure 3 This is a schematic diagram of a data enhancement device provided in an embodiment of this application. Figure 3 As shown, the data enhancement device may include:

[0161] The data acquisition unit 301 is used to acquire a sample set to be processed. Each sample in the sample set to be processed includes: labeled data, category labels corresponding to the labeled data, and core features. The core features are the discriminative basis in the labeled data used to support the labeling decision of the category labels.

[0162] The label determination unit 302 is used to determine dirty data labels and sparse data labels for each sample in the sample set to be processed based on preset sample quality assessment conditions. The dirty data labels are used to characterize whether the corresponding sample is dirty data with incorrect class labels, and the sparse data labels are used to characterize whether the target model to be trained is uncertain about the prediction classification result of the corresponding sample.

[0163] The sample screening unit 303 is used to select samples that are sparse data and not dirty data from the sample set to be processed based on dirty data labels and sparse data labels, and use them as enhancement seed samples.

[0164] The data augmentation unit 304 is used to generate a synthetic sample corresponding to the augmented seed sample based on the category labels and core features contained in the augmented seed sample.

[0165] Each module in the aforementioned data enhancement device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can invoke and execute the operations corresponding to each module.

[0166] This application also provides an electronic device, which may include at least one processor and a memory connected to the processor, wherein:

[0167] Memory is used to store computer programs;

[0168] The processor is used to execute computer programs to enable the electronic device to implement any of the data augmentation methods provided in the embodiments of this application.

[0169] refer to Figure 4 The diagram illustrates a structural schematic suitable for implementing the electronic device in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 4 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0170] like Figure 4 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. When the electronic device is powered on, the RAM 603 also stores various programs and data required for the operation of the electronic device. The processing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0171] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, memory cards, hard drives, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0172] This application also provides a computer program product, including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the data augmentation methods provided in this application. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0173] This application also provides a computer-readable storage medium carrying one or more computer programs. When these programs are executed by an electronic device, they enable the device to implement any of the data augmentation methods provided in this application. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0174] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0175] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0176] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0177] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0178] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A data augmentation method, characterized in that, include: Obtain a sample set to be processed. Each sample in the sample set to be processed includes: labeled data, category labels corresponding to the labeled data, and core features. The core features are the discriminative basis in the labeled data used to support the labeling decision of the category labels. Based on preset sample quality assessment conditions, dirty data labels and sparse data labels are determined for each sample in the sample set to be processed. The dirty data labels are used to characterize whether the corresponding sample is dirty data with incorrect class label, and the sparse data labels are used to characterize whether the target model to be trained is uncertain about the prediction classification result of the corresponding sample. Based on the dirty data label and the sparse data label, samples that are sparse data and not dirty data are selected from the sample set to be processed and used as enhancement seed samples. Based on the category labels and core features contained in the enhanced seed sample, a synthetic sample corresponding to the enhanced seed sample is generated.

2. The data augmentation method according to claim 1, characterized in that, The process of obtaining the sample set to be processed includes: Obtain an initial dataset, which consists of multiple labeled data points; The preset user annotation interface sequentially displays each labeled data in the initial dataset, and performs the following operations for each displayed labeled data: Lock the submit control; In response to the core feature region defined by the user in the labeled data, obtain the core features corresponding to the labeled data; In response to a user's category label classification operation on the labeled data, obtain the category label corresponding to the labeled data; After obtaining the core features and category labels corresponding to the labeled data, unlock the submission control; In response to the user's triggering operation on the submission control, the labeled data, the core features and category labels corresponding to the labeled data are treated as a sample and stored in the sample set to be processed.

3. The data augmentation method according to claim 2, characterized in that, Also includes: Select one initial labeled data from the initial dataset, store it in a pre-built data retention set, and iterate through each labeled data in the initial dataset except for the initial labeled data: Calculate the similarity information between the currently traversed labeled data and the labeled data in the data retention set, wherein the similarity information includes at least one similarity value; If all similarity values ​​in the similarity information are not greater than the preset similarity threshold, then the currently traversed labeled data is stored in the data retention set, and the next labeled data is traversed; otherwise, the next labeled data is traversed directly until the traversal ends. Use the data set obtained at the end of the traversal as the updated initial dataset; The step of sequentially displaying each labeled data item in the initial dataset on a preset user annotation interface includes: The user annotation interface sequentially displays each annotation data in the updated initial dataset.

4. The data augmentation method according to claim 3, characterized in that, Before iterating through each labeled data in the initial dataset except for the initial labeled data, the method further includes: A feature vector is generated for each labeled data in the initial dataset using the first pre-trained model, so as to obtain the feature vector of each labeled data in the initial dataset; The calculation of similarity information between the currently traversed labeled data and the labeled data in the data retention set includes: Construct an approximate nearest neighbor index of the feature vectors of all labeled data in the data retention set; Based on the approximate nearest neighbor index, retrieve the k candidate labeled data that are most similar to the currently traversed labeled data from the data retention set, where k is a preset positive integer; Calculate the similarity between the feature vector of the currently traversed labeled data and the feature vectors of the k candidate labeled data respectively, and obtain k similarity values ​​as the similarity information.

5. The data augmentation method according to claim 1, characterized in that, The method of determining dirty data labels and sparse data labels for each sample in the sample set to be processed, based on preset sample quality assessment conditions, includes: The pre-trained classification model is used to classify and predict each labeled data in the sample set to be processed, and the prediction probability vector corresponding to each labeled data in the sample set to be processed is obtained. The prediction probability vector includes the prediction probability of all categories in the preset category set. The predicted probability vector and category label corresponding to each labeled data in the sample set to be processed are input into the confidence learning module to obtain the label quality score corresponding to each labeled data in the sample set to be processed. Based on the label quality score corresponding to each labeled data in the sample set to be processed, determine the dirty data label of the sample in which each labeled data in the sample set to be processed is located; The maximum predicted probability value of the predicted probability vector corresponding to each labeled data in the sample set to be processed is determined as the category prediction confidence level corresponding to each labeled data in the sample set to be processed. Based on the category prediction confidence of each labeled data in the sample set to be processed, the sparse data label of the sample containing each labeled data in the sample set to be processed is determined.

6. The data augmentation method according to claim 5, characterized in that, The step of determining the dirty data label of the sample containing each labeled data in the sample set to be processed based on the label quality score corresponding to each labeled data in the sample set to be processed includes: For each piece of labeled data in the sample set to be processed, if the label quality score corresponding to the labeled data is less than a preset score threshold, then the dirty data label of the sample containing the labeled data is determined to be true; otherwise, the dirty data label of the sample containing the labeled data is determined to be false. Wherein, the dirty data label being true indicates that the sample is dirty data, and the dirty data label being false indicates that the sample is not dirty data.

7. The data augmentation method according to claim 5, characterized in that, The step of determining the sparse data label of the sample containing each labeled data in the sample set to be processed based on the category prediction confidence score corresponding to each labeled data in the sample set to be processed includes: For each labeled data in the sample set to be processed, if the confidence of the category prediction corresponding to the labeled data is less than the preset confidence threshold, then the sparse data label of the sample containing the labeled data is determined to be true; otherwise, the sparse data label of the sample containing the labeled data is determined to be false. Wherein, the sample corresponding to the sparse data label being a true representation is sparse data, and the sample corresponding to the sparse data label being a false representation is not sparse data.

8. The data augmentation method according to claim 1, characterized in that, The process of generating a synthetic sample corresponding to the enhanced seed sample based on the category labels and core features contained in the enhanced seed sample includes: Based on the category labels and core features contained in the enhanced seed samples, prompt instructions are generated. These prompt instructions are used to instruct the generative large language model to generate semantically similar but different sentence structures and expressions based on the category labels and core features. The prompting instructions are input into the generative large language model to obtain the augmented data output by the model; The augmented data and the category labels and core features contained in the augmented seed sample are combined to form the synthetic sample.

9. The data augmentation method according to claim 1, characterized in that, Also includes: The dirty data samples in the sample set to be processed are pushed to the manual review interface. After receiving the reviewer's correction result on the category label of the dirty data sample, the category label of the dirty data sample is updated according to the correction result. The dirty data sample refers to the sample whose dirty data label is characterized as dirty data. And / or, store at least one of the following data into the training dataset: samples that are neither dirty nor sparse data in the sample set to be processed, the synthetic samples, and dirty data samples whose category labels have been updated according to the correction results.

10. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the data augmentation method as described in any one of claims 1 to 9.

Images